Datasets:

neeto-lab

/

arxiv-struct-100k

like 0

Follow

Neeto Lab 3

[ [ "0.12006", "0.0778", "0.881", "0.0778", "0.881", "0.09143", "0.12006", "0.09143" ], [ "0.1234", "0.0932", "0.882", "0.0932", "0.882", "0.1069", "0.1234", "0.1069" ], [ "0.12177", "0.10986", "0.881", "0.10986", "0.881", "0.12354", "0.12177", "0.12354" ], [ "0.12177", "0.1259", "0.881", "0.1259", "0.881", "0.1395", "0.12177", "0.1395" ], [ "0.12006", "0.1401", "0.7256", "0.1414", "0.7256", "0.1561", "0.12006", "0.155" ], [ "0.1394", "0.1633", "0.882", "0.1633", "0.882", "0.177", "0.1394", "0.177" ], [ "0.1184", "0.1787", "0.882", "0.1799", "0.882", "0.1936", "0.1184", "0.1924" ], [ "0.12006", "0.1953", "0.882", "0.1953", "0.882", "0.209", "0.12006", "0.209" ], [ "0.12006", "0.2114", "0.882", "0.2114", "0.882", "0.2245", "0.12006", "0.2245" ], [ "0.12006", "0.228", "0.287", "0.228", "0.287", "0.2411", "0.12006", "0.2411" ], [ "0.1368", "0.2477", "0.881", "0.2488", "0.881", "0.2632", "0.1368", "0.2625" ], [ "0.1184", "0.2666", "0.3477", "0.2666", "0.3477", "0.2798", "0.1184", "0.2798" ], [ "0.12177", "0.2976", "0.882", "0.2976", "0.882", "0.3113", "0.12177", "0.3113" ], [ "0.1184", "0.3135", "0.882", "0.3123", "0.882", "0.3267", "0.1184", "0.328" ], [ "0.12177", "0.33", "0.882", "0.33", "0.882", "0.3445", "0.12177", "0.3445" ], [ "0.12006", "0.3467", "0.8833", "0.3467", "0.8833", "0.3599", "0.12006", "0.3599" ], [ "0.1377", "0.3682", "0.882", "0.3682", "0.882", "0.3818", "0.1377", "0.3818" ], [ "0.12006", "0.3843", "0.882", "0.3843", "0.882", "0.398", "0.12006", "0.398" ], [ "0.12006", "0.3997", "0.882", "0.3984", "0.882", "0.4133", "0.12006", "0.4146" ], [ "0.1184", "0.4146", "0.882", "0.415", "0.882", "0.43", "0.1184", "0.4287" ], [ "0.12177", "0.4316", "0.881", "0.4316", "0.881", "0.4453", "0.12177", "0.4453" ], [ "0.12006", "0.4478", "0.882", "0.4478", "0.882", "0.4614", "0.12006", "0.4614" ], [ "0.12006", "0.4631", "0.882", "0.4631", "0.882", "0.478", "0.12006", "0.478" ], [ "0.12006", "0.481", "0.34", "0.481", "0.34", "0.4946", "0.12006", "0.4946" ], [ "0.11584", "0.525", "0.266", "0.5254", "0.266", "0.5415", "0.11584", "0.5405" ], [ "0.1234", "0.583", "0.916", "0.583", "0.916", "0.597", "0.1234", "0.597" ], [ "0.1394", "0.601", "0.1621", "0.601", "0.1621", "0.615", "0.1394", "0.615" ], [ "0.2107", "0.601", "0.2747", "0.601", "0.2747", "0.6123", "0.2107", "0.6123" ], [ "0.3015", "0.601", "0.3914", "0.601", "0.3914", "0.6123", "0.3015", "0.6123" ], [ "0.4417", "0.601", "0.532", "0.601", "0.532", "0.6123", "0.4417", "0.6123" ], [ "0.602", "0.5996", "0.7", "0.5996", "0.7", "0.611", "0.602", "0.611" ], [ "0.8486", "0.601", "0.9087", "0.601", "0.9087", "0.6123", "0.8486", "0.6123" ], [ "0.3267", "0.615", "0.3955", "0.615", "0.3955", "0.63", "0.3267", "0.63" ], [ "0.4475", "0.615", "0.532", "0.615", "0.532", "0.63", "0.4475", "0.63" ], [ "0.8555", "0.615", "0.916", "0.615", "0.916", "0.63", "0.8555", "0.63" ], [ "0.22", "0.616", "0.2761", "0.616", "0.2761", "0.6274", "0.22", "0.6274" ], [ "0.1394", "0.616", "0.1621", "0.616", "0.1621", "0.63", "0.1394", "0.63" ], [ "0.613", "0.616", "0.702", "0.616", "0.702", "0.6274", "0.613", "0.6274" ], [ "0.287", "0.6426", "0.7153", "0.6436", "0.7153", "0.6577", "0.287", "0.6567" ], [ "0.1377", "0.6885", "0.882", "0.6885", "0.882", "0.7017", "0.1377", "0.7017" ], [ "0.12177", "0.7036", "0.882", "0.7036", "0.882", "0.7173", "0.12177", "0.7173" ], [ "0.11755", "0.7188", "0.882", "0.7197", "0.882", "0.734", "0.11755", "0.733" ], [ "0.12177", "0.7363", "0.2393", "0.7363", "0.2393", "0.75", "0.12177", "0.75" ], [ "0.12006", "0.774", "0.881", "0.774", "0.881", "0.7876", "0.12006", "0.7876" ], [ "0.12006", "0.7905", "0.8833", "0.7905", "0.8833", "0.804", "0.12006", "0.804" ], [ "0.12177", "0.8057", "0.882", "0.8057", "0.882", "0.8193", "0.12177", "0.8193" ], [ "0.12177", "0.8223", "0.881", "0.8223", "0.881", "0.836", "0.12177", "0.836" ], [ "0.1184", "0.8384", "0.3047", "0.8384", "0.3047", "0.8516", "0.1184", "0.8516" ], [ "0.1368", "0.8813", "0.5146", "0.883", "0.5146", "0.8984", "0.1368", "0.8965" ], [ "0.1394", "0.8965", "0.2568", "0.8965", "0.2568", "0.91", "0.1394", "0.91" ] ]

[ "新闻事件相关的实体对。首先，我们收集在我们的新闻语料库的发布日期范围内发生的事件的", "日期标记描述列表。这些描述在每个日历年都有多种语言版本，可以很容易地从维基百科上删", "除。接下来，我们从这些描述中提取所有实体对，并按事件日期对它们进行分组。最后，我们使", "用这些分组从我们的语料库中选择实体对，在事件日期之后的几天内，通过它们的计数和文章", "中的正逐点互信息（PPMI,国(ChurchandHanks，1990））对它们进行过滤", "为了进行比较，我们还评估了随机和日期窗口分组的配对。随机配对随机选择实体对。通过", "滑动日期窗口对实体对进行日期窗口分组的配对组。与事件引导配对一样，我们观察到与新闻", "中的事件相关的实体之间的关系的局部一致性。但是，它不会将所选实体限制为在Wikipedia", "事件描述中找到的实体。对于这两种方法，我们通过它们的计数和PPMI来过滤对，就像在事", "件引导的配对中一样", "句子选择很简单：对于选定的实体对，我们选择包含该对的句子1）随机或2）在日期窗口内", "（通常是用于选择对的日期）。", "训练（匹配空白）：我们采用训练方法来学习ESoaresetal：（2019）（MTB）整体使用", "的关系表示。也就是说，我们训练编码器fe(r）=h以产生关系表示hERd，该表示确定两个", "关系语句ri，ri是否通过计算它们的内积来描述相同的关系。与ESoaresetal.（2019）中一样", "我们的关系表示h是r中81和82标识的实体提及的实体开始标记的最终隐藏状态表示的串联", "对于训练，关系语句被收集到每个选定实体对的正例（根据我们的去噪技术与相同关系类型", "对齐的语句）和反例（与不同关系类型对齐的语句）的组中。在ESoaresetal.（2019）之后，负", "面示例包括“简单”和“困难”提及。“简单”提及不包括实体对的任何部分，而“硬”提及恰", "好包括实体对中的一个实体，这表明这些“硬”示例描述了需要消除歧义的相似但不同的关系。", "使用概率α，关系语句中的每个实体都被替换为[BLANK]令牌，而使用概率β，关系语句中", "的每个令牌都被替换为[MASK]令牌。我们训练fe的参数化，以最大程度地减少掩码语言建模", "损失E(Taylor，1953;Devlinetal，2019）和简单的二进制交叉嫡损失，从而鼓励给定实体对的", "正例表示彼此比负例更接近", "3Experiments", "Lang#Articles# Sentences#Entity Pairs#Rel.Statements#Denoised Statements", "en", "810,000", "11,925,900", "32,831,624", "209.796,400", "189,281", "904,908", "9,315,975", "95,881", "98.534", "es", "56.410,125", "Table 2: RCV1 and RCV2 corpus summary statistics", "数据：由于其规模和广泛的可用性，我们使用英语RCV1和RCV2路透社新闻语料库的西班", "牙语子集E（Lewisetal.，2004），从1996年8月至1997年8月的数千篇新闻文章中提取句子", "（见表2）。我们从维基百科收集同一时期的事件的日期标记描述。2我们用spaCy3模型提取", "名词/命名实体", "超参数：我们使用AdamE(KingmaandBa，2015）在英语和西班牙语语料库上微调多语言", "BERT，f，E(Devlinetal.，2019），学习率为3e-5，并且在单个NVIDIATeslaT4GPU上批", "量32。我们使用空白率α=0.7，掩码率β=0.15，0.5的BERT隐藏层丢失率和+4天的句", "子选择窗口。我们采用Soaresetal.（2019）使用的学习率、空白率和掩码率。有关其他调优详", "细信息，请参见附录B", "2For example, from https://es.wikipedia.org/wiki/1996", "3https://spacy.io" ]

{ "arxiv_id": "2012.02721", "language": "zh", "page_idx": 2 }

[ [ "0.684", "0.076", "0.763", "0.0736", "0.763", "0.0873", "0.684", "0.08905" ], [ "0.8403", "0.076", "0.906", "0.076", "0.906", "0.0873", "0.8403", "0.0873" ], [ "0.902", "0.0778", "0.9194", "0.0778", "0.9194", "0.0849", "0.902", "0.0849" ], [ "0.8037", "0.08014", "0.8237", "0.0748", "0.8286", "0.08435", "0.8076", "0.08905" ], [ "0.76", "0.079", "0.771", "0.079", "0.771", "0.08435", "0.76", "0.08435" ], [ "0.825", "0.08014", "0.8403", "0.08014", "0.8403", "0.08435", "0.825", "0.08435" ], [ "0.3333", "0.082", "0.3896", "0.082", "0.3896", "0.0968", "0.3333", "0.0968" ], [ "0.1335", "0.08075", "0.2183", "0.08435", "0.2167", "0.0986", "0.1318", "0.0956" ], [ "0.517", "0.082", "0.571", "0.08435", "0.5693", "0.0986", "0.5156", "0.0968" ], [ "0.5977", "0.0831", "0.65", "0.0831", "0.65", "0.0974", "0.5977", "0.0974" ], [ "0.7734", "0.08905", "0.797", "0.0932", "0.7925", "0.1069", "0.769", "0.1027" ], [ "0.837", "0.08905", "0.8623", "0.09265", "0.858", "0.1069", "0.833", "0.10394" ], [ "0.8984", "0.09265", "0.922", "0.09265", "0.922", "0.1057", "0.8984", "0.1057" ], [ "0.3284", "0.1069", "0.4382", "0.1069", "0.4382", "0.1217", "0.3284", "0.1217" ], [ "0.712", "0.09265", "0.739", "0.09265", "0.739", "0.1051", "0.712", "0.1051" ], [ "0.8213", "0.1069", "0.86", "0.1069", "0.86", "0.1223", "0.8213", "0.1223" ], [ "0.882", "0.1069", "0.924", "0.1069", "0.924", "0.1223", "0.882", "0.1223" ], [ "0.6987", "0.1081", "0.738", "0.1081", "0.738", "0.12054", "0.6987", "0.12054" ], [ "0.5884", "0.1081", "0.6265", "0.1081", "0.6265", "0.12354", "0.5884", "0.12354" ], [ "0.7573", "0.1081", "0.7925", "0.1081", "0.7925", "0.12054", "0.7573", "0.12054" ], [ "0.3257", "0.12054", "0.4248", "0.1223", "0.4248", "0.1383", "0.3257", "0.136" ], [ "0.51", "0.10925", "0.546", "0.10925", "0.546", "0.1217", "0.51", "0.1217" ], [ "0.51", "0.1223", "0.5474", "0.1223", "0.5474", "0.1395", "0.51", "0.1395" ], [ "0.5884", "0.1223", "0.6265", "0.1223", "0.6265", "0.1395", "0.5884", "0.1395" ], [ "0.6963", "0.1223", "0.7363", "0.1223", "0.7363", "0.1383", "0.6963", "0.1383" ], [ "0.8213", "0.1217", "0.858", "0.1217", "0.858", "0.1383", "0.8213", "0.1383" ], [ "0.3257", "0.1383", "0.4282", "0.1401", "0.4282", "0.155", "0.3257", "0.1526" ], [ "0.51", "0.1395", "0.5474", "0.1395", "0.5474", "0.1561", "0.51", "0.1561" ], [ "0.5903", "0.1395", "0.6265", "0.1395", "0.6265", "0.155", "0.5903", "0.155" ], [ "0.69", "0.1395", "0.735", "0.1395", "0.735", "0.155", "0.69", "0.155" ], [ "0.812", "0.1395", "0.858", "0.1395", "0.858", "0.155", "0.812", "0.155" ], [ "0.3267", "0.1561", "0.4963", "0.1561", "0.4963", "0.1704", "0.3267", "0.1704" ], [ "0.5903", "0.1561", "0.669", "0.1561", "0.669", "0.1704", "0.5903", "0.1704" ], [ "0.684", "0.1561", "0.735", "0.1561", "0.735", "0.1716", "0.684", "0.1716" ], [ "0.8096", "0.1561", "0.858", "0.1561", "0.858", "0.1716", "0.8096", "0.1716" ], [ "0.9146", "0.1592", "0.922", "0.1633", "0.9146", "0.1704", "0.9077", "0.1663" ], [ "0.3257", "0.1746", "0.3972", "0.1746", "0.3972", "0.1895", "0.3257", "0.1895" ], [ "0.51", "0.1567", "0.5356", "0.1567", "0.5356", "0.1704", "0.51", "0.1704" ], [ "0.1276", "0.1758", "0.3047", "0.1758", "0.3047", "0.1895", "0.1276", "0.1895" ], [ "0.51", "0.1758", "0.5356", "0.1758", "0.5356", "0.1895", "0.51", "0.1895" ], [ "0.592", "0.1758", "0.634", "0.1758", "0.634", "0.1895", "0.592", "0.1895" ], [ "0.6987", "0.1758", "0.738", "0.1758", "0.738", "0.1882", "0.6987", "0.1882" ], [ "0.7573", "0.1758", "0.7944", "0.1758", "0.7944", "0.1882", "0.7573", "0.1882" ], [ "0.8228", "0.1758", "0.858", "0.1758", "0.858", "0.1882", "0.8228", "0.1882" ], [ "0.881", "0.1746", "0.922", "0.1746", "0.922", "0.1907", "0.881", "0.1907" ], [ "0.1293", "0.1907", "0.3074", "0.1907", "0.3074", "0.2037", "0.1293", "0.2037" ], [ "0.3267", "0.1907", "0.398", "0.1907", "0.398", "0.2048", "0.3267", "0.2048" ], [ "0.5903", "0.1882", "0.6323", "0.1907", "0.6304", "0.206", "0.5884", "0.2037" ], [ "0.693", "0.1907", "0.7363", "0.1907", "0.7363", "0.206", "0.693", "0.206" ], [ "0.756", "0.1907", "0.7925", "0.1907", "0.7925", "0.206", "0.756", "0.206" ], [ "0.817", "0.1907", "0.86", "0.1907", "0.86", "0.206", "0.817", "0.206" ], [ "0.881", "0.1907", "0.922", "0.1907", "0.922", "0.206", "0.881", "0.206" ], [ "0.5903", "0.206", "0.6323", "0.206", "0.6323", "0.2214", "0.5903", "0.2214" ], [ "0.51", "0.1912", "0.5356", "0.1912", "0.5356", "0.206", "0.51", "0.206" ], [ "0.6978", "0.206", "0.739", "0.206", "0.739", "0.2214", "0.6978", "0.2214" ], [ "0.75", "0.206", "0.7944", "0.206", "0.7944", "0.2214", "0.75", "0.2214" ], [ "0.8193", "0.206", "0.8623", "0.206", "0.8623", "0.2214", "0.8193", "0.2214" ], [ "0.875", "0.206", "0.9253", "0.206", "0.9253", "0.2214", "0.875", "0.2214" ], [ "0.3257", "0.2073", "0.3972", "0.2073", "0.3972", "0.2214", "0.3257", "0.2214" ], [ "0.1234", "0.2073", "0.3074", "0.206", "0.3074", "0.2203", "0.1234", "0.2214" ], [ "0.51", "0.2073", "0.5635", "0.2073", "0.5635", "0.2214", "0.51", "0.2214" ], [ "0.12006", "0.2351", "0.8833", "0.2351", "0.8833", "0.2488", "0.12006", "0.2488" ], [ "0.1184", "0.2505", "0.882", "0.2517", "0.882", "0.2666", "0.1184", "0.2654" ], [ "0.1184", "0.2666", "0.885", "0.2673", "0.885", "0.282", "0.1184", "0.2808" ], [ "0.11755", "0.2832", "0.398", "0.2832", "0.398", "0.2964", "0.11755", "0.2964" ], [ "0.1503", "0.3164", "0.2107", "0.3164", "0.2107", "0.3308", "0.1503", "0.3308" ], [ "0.465", "0.3164", "0.51", "0.3164", "0.51", "0.332", "0.465", "0.332" ], [ "0.7246", "0.3164", "0.7676", "0.3184", "0.7656", "0.3342", "0.723", "0.332" ], [ "0.7734", "0.3184", "0.831", "0.3184", "0.831", "0.3342", "0.7734", "0.3342" ], [ "0.642", "0.3196", "0.6978", "0.3196", "0.6978", "0.3308", "0.642", "0.3308" ], [ "0.3435", "0.3184", "0.4417", "0.3184", "0.4417", "0.33", "0.3435", "0.33" ], [ "0.1545", "0.3308", "0.2551", "0.3333", "0.2551", "0.3474", "0.1545", "0.3457" ], [ "0.4114", "0.335", "0.4434", "0.335", "0.4434", "0.3474", "0.4114", "0.3474" ], [ "0.4575", "0.332", "0.5054", "0.332", "0.5054", "0.3474", "0.4575", "0.3474" ], [ "0.634", "0.335", "0.706", "0.335", "0.706", "0.3499", "0.634", "0.3499" ], [ "0.7256", "0.3362", "0.7676", "0.3362", "0.7676", "0.3499", "0.7256", "0.3499" ], [ "0.785", "0.3362", "0.827", "0.3362", "0.827", "0.3499", "0.785", "0.3499" ], [ "0.1545", "0.3499", "0.3267", "0.3499", "0.3267", "0.3635", "0.1545", "0.3635" ], [ "0.34", "0.3508", "0.4417", "0.3508", "0.4417", "0.3623", "0.34", "0.3623" ], [ "0.4568", "0.3486", "0.5073", "0.3486", "0.5073", "0.364", "0.4568", "0.364" ], [ "0.6304", "0.3516", "0.702", "0.354", "0.702", "0.3682", "0.6304", "0.3665" ], [ "0.7246", "0.3623", "0.769", "0.3623", "0.769", "0.3772", "0.7246", "0.3772" ], [ "0.7827", "0.3623", "0.8286", "0.3623", "0.8286", "0.3772", "0.7827", "0.3772" ], [ "0.1561", "0.3694", "0.2241", "0.3694", "0.2241", "0.3813", "0.1561", "0.3813" ], [ "0.4106", "0.3694", "0.4434", "0.3694", "0.4434", "0.3818", "0.4106", "0.3818" ], [ "0.4668", "0.3694", "0.5054", "0.3694", "0.5054", "0.3818", "0.4668", "0.3818" ], [ "0.6323", "0.3682", "0.711", "0.3706", "0.711", "0.3855", "0.6323", "0.383" ], [ "0.1545", "0.3843", "0.3105", "0.3843", "0.3105", "0.398", "0.1545", "0.398" ], [ "0.3862", "0.3843", "0.444", "0.3843", "0.444", "0.3984", "0.3862", "0.3984" ], [ "0.461", "0.3843", "0.5073", "0.3843", "0.5073", "0.3984", "0.461", "0.3984" ], [ "0.575", "0.3914", "0.8833", "0.3901", "0.8833", "0.405", "0.575", "0.4062" ], [ "0.12006", "0.4062", "0.5415", "0.4062", "0.5415", "0.4192", "0.12006", "0.4192" ], [ "0.575", "0.4038", "0.882", "0.4038", "0.882", "0.4175", "0.575", "0.4175" ], [ "0.1184", "0.4187", "0.5425", "0.4187", "0.5425", "0.4316", "0.1184", "0.4316" ], [ "0.575", "0.4175", "0.881", "0.4175", "0.881", "0.4275", "0.575", "0.4275" ], [ "0.1184", "0.43", "0.5425", "0.43", "0.5425", "0.4436", "0.1184", "0.4436" ], [ "0.575", "0.43", "0.7036", "0.43", "0.7036", "0.4407", "0.575", "0.4407" ], [ "0.1184", "0.4536", "0.882", "0.4548", "0.882", "0.4697", "0.1184", "0.4685" ], [ "0.1184", "0.4714", "0.8696", "0.4714", "0.8696", "0.4863", "0.1184", "0.4863" ], [ "0.1394", "0.5103", "0.881", "0.5103", "0.881", "0.524", "0.1394", "0.524" ], [ "0.1184", "0.527", "0.5264", "0.527", "0.5264", "0.5405", "0.1184", "0.5405" ], [ "0.1368", "0.542", "0.882", "0.542", "0.882", "0.557", "0.1368", "0.557" ], [ "0.12177", "0.5586", "0.882", "0.5586", "0.882", "0.5737", "0.12177", "0.5737" ], [ "0.12177", "0.576", "0.882", "0.576", "0.882", "0.589", "0.12177", "0.589" ], [ "0.12006", "0.5913", "0.75", "0.5913", "0.75", "0.605", "0.12006", "0.605" ], [ "0.1368", "0.608", "0.882", "0.608", "0.882", "0.6216", "0.1368", "0.6216" ], [ "0.12006", "0.6245", "0.8833", "0.6245", "0.8833", "0.638", "0.12006", "0.638" ], [ "0.1184", "0.6396", "0.882", "0.64", "0.882", "0.655", "0.1184", "0.654" ], [ "0.12006", "0.6567", "0.592", "0.6567", "0.592", "0.6704", "0.12006", "0.6704" ], [ "0.1368", "0.673", "0.882", "0.673", "0.882", "0.687", "0.1368", "0.687" ], [ "0.12006", "0.6895", "0.3972", "0.6895", "0.3972", "0.7036", "0.12006", "0.7036" ], [ "0.12006", "0.7227", "0.882", "0.7227", "0.882", "0.7363", "0.12006", "0.7363" ], [ "0.12006", "0.7383", "0.885", "0.7383", "0.885", "0.753", "0.12006", "0.753" ], [ "0.12006", "0.755", "0.8833", "0.755", "0.8833", "0.7695", "0.12006", "0.7695" ], [ "0.1184", "0.7715", "0.8833", "0.771", "0.8833", "0.785", "0.1184", "0.786" ], [ "0.12177", "0.788", "0.3838", "0.788", "0.3838", "0.802", "0.12177", "0.802" ], [ "0.1368", "0.8027", "0.8833", "0.804", "0.8833", "0.8184", "0.1368", "0.818" ], [ "0.1184", "0.8193", "0.882", "0.821", "0.882", "0.835", "0.1184", "0.8345" ], [ "0.12006", "0.836", "0.881", "0.836", "0.881", "0.851", "0.12006", "0.851" ], [ "0.12006", "0.8525", "0.882", "0.8525", "0.882", "0.866", "0.12006", "0.866" ], [ "0.12006", "0.868", "0.2124", "0.868", "0.2124", "0.883", "0.12006", "0.883" ], [ "0.1377", "0.8936", "0.287", "0.8955", "0.287", "0.91", "0.1377", "0.9077" ] ]

[ "n=5,k:", "10.k=", "", "n", "-", "", "Model", "Denoising", "Langs", "# Ex.", "es", "en", "es", "mBERT base", "en.", "58.7", "50.6", "62.9", "n/a", "63.8", "BERT base", "n/a", "n/a", "n/a", "68.7", "57.1", "BERT large", "n/a", "n/a", "72.9", "62.3", "BERT large + MTB", "300 mil+", "80.4", "71.5", ".", "mBERT", "en", "date-window grouped", "en", "100k", "72.4", "67.5", "59.3", "56.7", "event-guided (EvtGD)", "mBERT", "50k", "78.7", "75.1", "69.0", "64.3", "70 k", "en", "77.5", "76.7", "68.1", "64.4", "mBERT", "event-guided (EvtGD)", "en, es", "Table 3: Zero-shot (no fine-tuning) accuracies on the FewRel development set. For each test", "case, k is the number of relation types and n is the number of examples per relation type.", "The development set is used for comparison with the +'d results from MTB. Random pairing", "excluded due to poor performance,", "Model", "1 %", "1%", "10%", "Model", "Corpus Size", "BERTlarge", "n/a", "28.6", "mBERT", "33.0", "72.8", "BERT large + MTB", "300 mil+ ex", "31.2", "mBERT", "41.5", "76.8", "mBERT", "n/a", "28.5", "+EvtGD", "mBERT + EvtGD", "50k ex", "33.3", "(b) F1-scores on the TAC KBP 2016 Spanish", "(a) F1-scores on a random held out slice of 1500 examples from", "relation corpus (from our en/es model, fine-", "the SemEval 2010 Task 8 training set (for comparison with the", "tuned on 1 % and 10 % of its data and eval", "t 'd results from MTB.). Models are fine-tuned on 80 examples.", "uated on the rest).", "Table 4: F1-scores on two standard supervised relation extraction benchmarks for English (Se-", "mEval 2010 Task 8) and Spanish (TAC KBP 2016). EvtGD indicates event-guided denoising,", "评估：我们在FewRel，SemEval2010Task8和TACKBP2016的西班牙语子集上评估嵌人空", "间的性能，几乎没有对特定于任务的数据进行微调", "FewRel是一个关系基准测试，其中给定一组n关系类型，每个类型都有k关系语句示例，", "任务是确定哪个n关系类型由查询关系语句q表示。为此，我们嵌人q和所有候选关系语", "句cf...ch找到嵌入与g最相似的候选c，并预测关系类型i。除英语外，我们还使用", "GoogleTranslate（在西班牙语上表现良好）4翻译了FewRel开发集的10个%", "SemEval2010任务8是标准关系分类基准。TACKBP2016是西班牙语知识库人口语料库；虽", "然它通常用于评估插槽填充任务，但它包含带有标记关系的文档。对于两个语料库，在ESoares", "etal.（2019）（包括我们在附录C中包含的它们的超参数）之后，我们将单个密集层附加到我", "们的嵌入空间，以将关系表示hERd映射到它们各自的类", "我们将我们的性能与ESoaresetal.（2019）（BERTlarge+MTB）和mBERT中发布的结果", "进行比较，而没有事件引导的去噪", "结果：我们的事件引导去噪方法表现良好，几乎没有对特定于任务的数据进行微调。尽管在少得", "多的训练数据（50kvs300mil+示例）上训练了一个小得多的模型（110milvs340mil参数），", "但我们在类似的微调约束下获得了与ESoaresetal.（2019）相当或更好的结果。。在FewRel上，", "没有任何模型的微调，我们仅实现了2%精度的降低（在基准测试中，即使少量数据的微调也", "会产生超过5%的增益）（表3）", "在SemEval上，当两个模型仅在其训练数据的1个%（80个示例）上进行微调时，我们的模", "型的性能优于ESoaresetal.（2019）（表4a）。这些结果与他们随后在FewRel训练集上进行微", "调的观察结果一致-也就是说，提高性能最关键的是微调期间看到的不同关系类型的数量，而不", "是每个关系类型的示例数量。因此，我们假设我们的事件引导方法正在产生一个小而多样的高", "质量关系集", "4translate.google.com" ]

{ "arxiv_id": "2012.02721", "language": "zh", "page_idx": 3 }

[ [ "0.1377", "0.076", "0.881", "0.076", "0.881", "0.08905", "0.1377", "0.08905" ], [ "0.12006", "0.0932", "0.882", "0.0932", "0.882", "0.1069", "0.12006", "0.1069" ], [ "0.12177", "0.1081", "0.756", "0.10925", "0.756", "0.12354", "0.12177", "0.1223" ], [ "0.1351", "0.1247", "0.8833", "0.1259", "0.8833", "0.1401", "0.1351", "0.1395" ], [ "0.12006", "0.1426", "0.1427", "0.1426", "0.1427", "0.1561", "0.12006", "0.1561" ], [ "0.11755", "0.1704", "0.251", "0.1704", "0.251", "0.1853", "0.11755", "0.1853" ], [ "0.12177", "0.1965", "0.882", "0.1965", "0.882", "0.2102", "0.12177", "0.2102" ], [ "0.12177", "0.212", "0.8833", "0.212", "0.8833", "0.2257", "0.12177", "0.2257" ], [ "0.12177", "0.2286", "0.8345", "0.2286", "0.8345", "0.2423", "0.12177", "0.2423" ], [ "0.1394", "0.2465", "0.879", "0.2465", "0.879", "0.257", "0.1394", "0.257" ], [ "0.12177", "0.2632", "0.881", "0.2632", "0.881", "0.2737", "0.12177", "0.2737" ], [ "0.12006", "0.2778", "0.8076", "0.2778", "0.8076", "0.2915", "0.12006", "0.2915" ], [ "0.1377", "0.2944", "0.879", "0.2944", "0.879", "0.3047", "0.1377", "0.3047" ], [ "0.1184", "0.3088", "0.882", "0.31", "0.882", "0.3247", "0.1184", "0.3237" ], [ "0.12006", "0.3267", "0.3586", "0.3267", "0.3586", "0.3403", "0.12006", "0.3403" ], [ "0.1368", "0.3416", "0.881", "0.3416", "0.881", "0.3552", "0.1368", "0.3552" ], [ "0.12177", "0.3594", "0.882", "0.3594", "0.882", "0.3723", "0.12177", "0.3723" ], [ "0.1184", "0.376", "0.882", "0.3735", "0.882", "0.3884", "0.1184", "0.3901" ], [ "0.12006", "0.3914", "0.882", "0.3914", "0.882", "0.405", "0.12006", "0.405" ], [ "0.12006", "0.4067", "0.3242", "0.4067", "0.3242", "0.4204", "0.12006", "0.4204" ], [ "0.11755", "0.4353", "0.1351", "0.4353", "0.1351", "0.4507", "0.11755", "0.4507" ], [ "0.12006", "0.4614", "0.882", "0.4614", "0.882", "0.4744", "0.12006", "0.4744" ], [ "0.1184", "0.478", "0.882", "0.478", "0.882", "0.4912", "0.1184", "0.4912" ], [ "0.12177", "0.4954", "0.881", "0.4954", "0.881", "0.506", "0.12177", "0.506" ], [ "0.12006", "0.5103", "0.2913", "0.5103", "0.2913", "0.524", "0.12006", "0.524" ], [ "0.11755", "0.552", "0.2124", "0.554", "0.2124", "0.5684", "0.11755", "0.5664" ], [ "0.1184", "0.576", "0.879", "0.576", "0.879", "0.589", "0.1184", "0.589" ], [ "0.1351", "0.589", "0.4634", "0.5884", "0.4634", "0.6025", "0.1351", "0.604" ], [ "0.1184", "0.6133", "0.881", "0.6133", "0.881", "0.6274", "0.1184", "0.6274" ], [ "0.1368", "0.627", "0.7173", "0.626", "0.7173", "0.64", "0.1368", "0.641" ], [ "0.1184", "0.653", "0.882", "0.653", "0.882", "0.6665", "0.1184", "0.6665" ], [ "0.1368", "0.6646", "0.882", "0.665", "0.882", "0.6797", "0.1368", "0.6787" ], [ "0.1377", "0.6777", "0.1797", "0.6797", "0.178", "0.6924", "0.1368", "0.6904" ], [ "0.12006", "0.7036", "0.882", "0.7036", "0.882", "0.7173", "0.12006", "0.7173" ], [ "0.1368", "0.7173", "0.879", "0.7163", "0.879", "0.73", "0.1368", "0.731" ], [ "0.1351", "0.73", "0.3376", "0.7285", "0.3376", "0.7437", "0.1351", "0.7446" ], [ "0.1184", "0.7544", "0.638", "0.755", "0.638", "0.7695", "0.1184", "0.7686" ], [ "0.1184", "0.78", "0.881", "0.78", "0.881", "0.7935", "0.1184", "0.7935" ], [ "0.1377", "0.7935", "0.881", "0.7935", "0.881", "0.807", "0.1377", "0.807" ], [ "0.1377", "0.8057", "0.8364", "0.8057", "0.8364", "0.8193", "0.1377", "0.8193" ], [ "0.1184", "0.832", "0.879", "0.832", "0.879", "0.8457", "0.1184", "0.8457" ], [ "0.1368", "0.844", "0.879", "0.8457", "0.879", "0.8594", "0.1368", "0.858" ], [ "0.1368", "0.857", "0.434", "0.8555", "0.434", "0.8706", "0.1368", "0.8716" ], [ "0.12006", "0.883", "0.882", "0.883", "0.882", "0.8965", "0.12006", "0.8965" ], [ "0.1377", "0.898", "0.2477", "0.898", "0.2477", "0.909", "0.1377", "0.909" ] ]

[ "我们的多语言方法在FewRel（自动翻译）和TACKBP西班牙语子集（表3，4b）上也优于", "mBERT。在TACKBP上，当在相同的小百分比任务特定数据上进行微调时，我们的方法实现", "了比mBERT更高的F1。在20个%（>500示例）之后，两个模型的性能收敛", "这些英语和西班牙语的结果证明了我们的事件引导去噪方法对资源受限关系提取任务的有效", "性", "4Conclusion", "我们提出了一种用于关系提取语料库创建的事件引导去噪方法，该方法与当前最先进的培训程", "序一起使用时，在资源遗乏的情况下，仅需一小部分培训成本，即可在英语中获得可比的结果。", "它在西班牙语中也表现出色，证明了它对非英语语言中资源受限的关系提取任务的适应性", "我们的技术为更广泛的研究社区提供了通过显着降低其相关培训成本来近似当前关系提取的", "最新技术的能力。但是，它需要一个相当大的带有日期标记的新闻语料库，而低资源语言可能", "无法使用。。我们将探索更广泛的语言覆盖范围和最小的语料库大小，以便将来的工作", "扩大语言覆盖范围的一个有希望的方向是通过“代码转换”示例和其他语言建模损失进行跨", "语言学习（例如。g。（LampleandConneau，2019））。我们假设这些方法可以帮助语言之间的知", "识转移并改善下游任务的结果", "最后，我们注意到，由于我们的方法从新闻语料库中提取关系语句，因此底层关系类型的最", "终分布可能与维基百科中的分布不同。例如，维基百科可能包含更多标准本体论关系的表达", "(e。g。，父亲，出生的）事实的特征。尽管存在这种假设的差异，但我们的方法在FewRel和", "SemEval2010Task8上都表现良好，两者都包含此类关系类型的子集。将来，我们打算更仔细地", "调查这些差异及其影响。。", "5", "Acknowledgements这项工作基于DARPA根据协议号FA8750-18-2-0014赞助的研究。TheU。", "S。政府有权为政府目的复制和发行再版，尽管其上有任何版权标记。本文所包含的观点和结论", "是作者的观点和结论，不应被解释为必然代表DARPA或美国政府明示或暗示的官方政策或认", "可。．S。Government", "References", "Kenneth Church and Patrick Hanks. 1990. Word association norms, mutual information, and lexicogra", "phy. Computational linguistics, 16(1):22-29.", "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep", "bidirectional transformers for language understanding.ArXiv, abs/1810.04805.", "Joe Ellis, Jeremy Getman, Dana Fore, Neil Kuster, Zhiyi Song, Ann Bies, and Stephanie M Strassel.", "2015. Overview of linguistic resources for the tac kbp 2015 evaluations: Methodologies and results. In", "TAC.", "Xu Han, Hao Zhu, Pengfei Yu, Ziyun Wang, Yuan Yao, Zhiyuan Liu, and Maosong Sun. 2018. Fewrel:", "A large-scale supervised few-shot relation classification dataset with state-of-the-art evaluation. arXiv", "preprint arXiv:1810.10147.", "Zellig S Harris. 1954. Distributional structure. Word, 10(2-3):146-162", "Iris Hendrickx, Su Nam Kim, Zornitsa Kozareva, Preslav Nakov, Diarmuid O Seaghdha, Sebastian Pado", "Marco Pennacchiotti, Lorenza Romano, and Stan Szpakowicz. 2019. Semeval-2010 task 8: Multi-way", "classification of semantic relations between pairs of nominals. arXiv preprint arXiv:1911.10422.", "Heng Ji and Ralph Grishman. 2011. Knowledge base population: Successful approaches and challenges", "In Proceedings of the 49th annual meeting of the association for computational linguistics: humar", "language technologies, pages 1148-1158.", "Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. CoRR,", "abs/1412.6980." ]

{ "arxiv_id": "2012.02721", "language": "zh", "page_idx": 4 }

[ [ "0.1953", "0.06946", "0.29", "0.07135", "0.29", "0.0833", "0.1953", "0.0808" ], [ "0.1299", "0.07574", "0.152", "0.07574", "0.152", "0.089", "0.1299", "0.089" ], [ "0.3203", "0.089", "0.3694", "0.089", "0.3694", "0.101", "0.3203", "0.101" ], [ "0.614", "0.101", "0.636", "0.101", "0.636", "0.11426", "0.614", "0.11426" ], [ "0.3938", "0.11237", "0.4844", "0.11237", "0.4844", "0.1225", "0.3938", "0.1225" ], [ "0.5", "0.11554", "0.5728", "0.11615", "0.5728", "0.1276", "0.5", "0.127" ], [ "0.4355", "0.1276", "0.451", "0.1338", "0.4436", "0.1445", "0.429", "0.1395" ], [ "0.6357", "0.13", "0.699", "0.1332", "0.6978", "0.1483", "0.634", "0.1459" ], [ "0.634", "0.1787", "0.699", "0.1787", "0.699", "0.1932", "0.634", "0.1932" ], [ "0.4714", "0.1863", "0.5024", "0.1863", "0.5024", "0.1989", "0.4714", "0.1989" ], [ "0.8037", "0.1812", "0.8677", "0.1831", "0.8677", "0.1976", "0.8037", "0.1957" ], [ "0.513", "0.1901", "0.5337", "0.1925", "0.5317", "0.202", "0.511", "0.1995" ], [ "0.1299", "0.2317", "0.1528", "0.2317", "0.1528", "0.2456", "0.1299", "0.2456" ], [ "0.286", "0.2393", "0.3342", "0.2412", "0.3342", "0.252", "0.286", "0.25" ], [ "0.3816", "0.2393", "0.4192", "0.2393", "0.4192", "0.25", "0.3816", "0.25" ], [ "0.4756", "0.2393", "0.5063", "0.2424", "0.505", "0.25", "0.4739", "0.2468" ], [ "0.58", "0.2424", "0.6006", "0.2424", "0.6006", "0.2468", "0.58", "0.2468" ], [ "0.2091", "0.2568", "0.2402", "0.2568", "0.2402", "0.2651", "0.2091", "0.2651" ], [ "0.2632", "0.2583", "0.312", "0.2583", "0.312", "0.2683", "0.2632", "0.2683" ], [ "0.437", "0.26", "0.505", "0.26", "0.505", "0.2708", "0.437", "0.2708" ], [ "0.5376", "0.26", "0.6094", "0.26", "0.6094", "0.2708", "0.5376", "0.2708" ], [ "0.3464", "0.2615", "0.411", "0.2615", "0.411", "0.2715", "0.3464", "0.2715" ], [ "0.1503", "0.2615", "0.1978", "0.2615", "0.1978", "0.2715", "0.1503", "0.2715" ], [ "0.2607", "0.2651", "0.3145", "0.2651", "0.3145", "0.276", "0.2607", "0.276" ], [ "0.3284", "0.2854", "0.3677", "0.2874", "0.367", "0.2986", "0.327", "0.2966" ], [ "0.1683", "0.2954", "0.21", "0.2954", "0.21", "0.3062", "0.1683", "0.3062" ], [ "0.2181", "0.2974", "0.2566", "0.2974", "0.2566", "0.305", "0.2181", "0.305" ], [ "0.423", "0.2917", "0.4453", "0.2917", "0.4453", "0.3037", "0.423", "0.3037" ], [ "0.1732", "0.305", "0.2035", "0.305", "0.2035", "0.3164", "0.1732", "0.3164" ], [ "0.267", "0.3125", "0.326", "0.3125", "0.326", "0.3225", "0.267", "0.3225" ], [ "0.1691", "0.3208", "0.2035", "0.3208", "0.2035", "0.33", "0.1691", "0.33" ], [ "0.343", "0.3245", "0.3743", "0.3245", "0.3743", "0.3289", "0.343", "0.3289" ], [ "0.406", "0.3245", "0.4314", "0.3245", "0.4314", "0.3289", "0.406", "0.3289" ], [ "0.464", "0.3245", "0.478", "0.3245", "0.478", "0.3289", "0.464", "0.3289" ], [ "0.517", "0.3245", "0.5337", "0.3245", "0.5337", "0.3289", "0.517", "0.3289" ], [ "0.08496", "0.3623", "0.9067", "0.363", "0.9067", "0.3774", "0.08496", "0.377" ], [ "0.1462", "0.3276", "0.2255", "0.3289", "0.2255", "0.339", "0.1462", "0.3384" ], [ "0.08826", "0.375", "0.911", "0.375", "0.911", "0.3882", "0.08826", "0.3882" ], [ "0.08905", "0.3882", "0.675", "0.3882", "0.675", "0.3984", "0.08905", "0.3984" ], [ "0.08905", "0.43", "0.4902", "0.43", "0.4902", "0.44", "0.08905", "0.44" ], [ "0.4944", "0.4287", "0.665", "0.4287", "0.665", "0.4426", "0.4944", "0.4426" ], [ "0.08826", "0.443", "0.4902", "0.443", "0.4902", "0.457", "0.08826", "0.457" ], [ "0.527", "0.449", "0.9126", "0.449", "0.9126", "0.4622", "0.527", "0.4622" ], [ "0.08905", "0.459", "0.492", "0.459", "0.492", "0.473", "0.08905", "0.473" ], [ "0.509", "0.464", "0.9126", "0.464", "0.9126", "0.4778", "0.509", "0.4778" ], [ "0.08826", "0.4749", "0.1814", "0.4749", "0.1814", "0.486", "0.08826", "0.486" ], [ "0.511", "0.4792", "0.911", "0.4792", "0.911", "0.4924", "0.511", "0.4924" ], [ "0.10455", "0.4905", "0.4902", "0.4905", "0.4902", "0.504", "0.10455", "0.504" ], [ "0.5107", "0.4944", "0.9126", "0.4944", "0.9126", "0.5083", "0.5107", "0.5083" ], [ "0.0866", "0.504", "0.4902", "0.505", "0.4902", "0.5195", "0.0866", "0.518" ], [ "0.5107", "0.509", "0.9126", "0.509", "0.9126", "0.523", "0.5107", "0.523" ], [ "0.08905", "0.521", "0.4902", "0.521", "0.4902", "0.534", "0.08905", "0.534" ], [ "0.509", "0.523", "0.9126", "0.5234", "0.9126", "0.537", "0.509", "0.536" ], [ "0.08905", "0.537", "0.4844", "0.537", "0.4844", "0.5474", "0.08905", "0.5474" ], [ "0.5107", "0.539", "0.9126", "0.539", "0.9126", "0.553", "0.5107", "0.553" ], [ "0.0907", "0.552", "0.486", "0.552", "0.486", "0.5625", "0.0907", "0.5625" ], [ "0.509", "0.553", "0.911", "0.5537", "0.911", "0.5684", "0.509", "0.5674" ], [ "0.08905", "0.5664", "0.4885", "0.5664", "0.4885", "0.58", "0.08905", "0.58" ], [ "0.5107", "0.5684", "0.9126", "0.5693", "0.9126", "0.5835", "0.5107", "0.582" ], [ "0.08826", "0.5806", "0.4885", "0.5806", "0.4885", "0.5947", "0.08826", "0.5947" ], [ "0.509", "0.5835", "0.91", "0.584", "0.91", "0.5986", "0.509", "0.5977" ], [ "0.08826", "0.595", "0.4885", "0.595", "0.4885", "0.6094", "0.08826", "0.6094" ], [ "0.5107", "0.5996", "0.9126", "0.5996", "0.9126", "0.613", "0.5107", "0.613" ], [ "0.08826", "0.61", "0.4878", "0.61", "0.4878", "0.6235", "0.08826", "0.6235" ], [ "0.5107", "0.6143", "0.675", "0.6143", "0.675", "0.6274", "0.5107", "0.6274" ], [ "0.08905", "0.6274", "0.3945", "0.6274", "0.3945", "0.638", "0.08905", "0.638" ], [ "0.528", "0.63", "0.911", "0.63", "0.911", "0.643", "0.528", "0.643" ], [ "0.10297", "0.6426", "0.4902", "0.6426", "0.4902", "0.656", "0.10297", "0.656" ], [ "0.509", "0.6445", "0.91", "0.646", "0.91", "0.659", "0.509", "0.6577" ], [ "0.08826", "0.6577", "0.4885", "0.6577", "0.4885", "0.672", "0.08826", "0.672" ], [ "0.5107", "0.66", "0.9126", "0.66", "0.9126", "0.674", "0.5107", "0.674" ], [ "0.08905", "0.6724", "0.4885", "0.6724", "0.4885", "0.686", "0.08905", "0.686" ], [ "0.5107", "0.6753", "0.9126", "0.6753", "0.9126", "0.6895", "0.5107", "0.6895" ], [ "0.08905", "0.688", "0.4885", "0.688", "0.4885", "0.702", "0.08905", "0.702" ], [ "0.509", "0.6895", "0.91", "0.69", "0.91", "0.704", "0.509", "0.7026" ], [ "0.08826", "0.7026", "0.4885", "0.7026", "0.4885", "0.7163", "0.08826", "0.7163" ], [ "0.5073", "0.705", "0.91", "0.705", "0.91", "0.7183", "0.5073", "0.7183" ], [ "0.08826", "0.7183", "0.4885", "0.7183", "0.4885", "0.732", "0.08826", "0.732" ], [ "0.5107", "0.72", "0.911", "0.72", "0.911", "0.7344", "0.5107", "0.7344" ], [ "0.08826", "0.7344", "0.2181", "0.7344", "0.2181", "0.7476", "0.08826", "0.7476" ], [ "0.511", "0.7363", "0.5435", "0.7363", "0.5435", "0.7485", "0.511", "0.7485" ], [ "0.10455", "0.75", "0.4885", "0.75", "0.4885", "0.763", "0.10455", "0.763" ], [ "0.527", "0.752", "0.9126", "0.752", "0.9126", "0.765", "0.527", "0.765" ], [ "0.08826", "0.765", "0.4902", "0.765", "0.4902", "0.779", "0.08826", "0.779" ], [ "0.5107", "0.7666", "0.91", "0.7666", "0.91", "0.78", "0.5107", "0.78" ], [ "0.08905", "0.782", "0.4878", "0.782", "0.4878", "0.7925", "0.08905", "0.7925" ], [ "0.5107", "0.782", "0.911", "0.782", "0.911", "0.7954", "0.5107", "0.7954" ], [ "0.0866", "0.7944", "0.4902", "0.7935", "0.4902", "0.808", "0.0866", "0.8096" ], [ "0.511", "0.797", "0.9087", "0.797", "0.9087", "0.81", "0.511", "0.81" ], [ "0.08826", "0.81", "0.492", "0.81", "0.492", "0.8237", "0.08826", "0.8237" ], [ "0.509", "0.81", "0.9126", "0.811", "0.9126", "0.8257", "0.509", "0.8247" ], [ "0.08826", "0.8247", "0.4902", "0.8247", "0.4902", "0.8384", "0.08826", "0.8384" ], [ "0.5107", "0.8276", "0.5786", "0.8276", "0.5786", "0.839", "0.5107", "0.839" ], [ "0.08826", "0.8413", "0.1365", "0.8413", "0.1365", "0.853", "0.08826", "0.853" ], [ "0.5107", "0.853", "0.66", "0.853", "0.66", "0.866", "0.5107", "0.866" ], [ "0.2059", "0.8735", "0.3677", "0.875", "0.3677", "0.8896", "0.2059", "0.888" ], [ "0.528", "0.8726", "0.91", "0.8726", "0.91", "0.886", "0.528", "0.886" ], [ "0.5107", "0.887", "0.91", "0.887", "0.91", "0.901", "0.5107", "0.901" ], [ "0.1054", "0.905", "0.4885", "0.905", "0.4885", "0.9155", "0.1054", "0.9155" ], [ "0.509", "0.903", "0.91", "0.903", "0.91", "0.9165", "0.509", "0.9165" ], [ "0.08905", "0.9194", "0.4272", "0.9194", "0.4272", "0.9297", "0.08905", "0.9297" ], [ "0.513", "0.9185", "0.911", "0.9185", "0.911", "0.9316", "0.513", "0.9316" ] ]

[ "Robotgrowing", "(a)", "i2Cbus", "(b)", "Sensingbands", "Aggregator", "7", "Pressure", "Pressure", "", "Eversion", "", "(c)", "12CLeadel", "Follower", "Followe", "F0llo", "Serial", "Aggregator", "Hoop Sensor 2", "Hoop SensorN", "HoopSensor1", "Computer", "(Teensy4.0)", "/HWPC", "BNO055", "SWPC", "ADC", "(IMU)", "ATMega328p", "SHT31", "Thorm", "horo", "bor", "horr", "Fig. 2.(a) A rendering of a growing robot deploying sensor bands, communicating with an aggregator DAQ via a bus along its axis", "（Temp/humidity)", "(b) The process by which growing robots evert, illustrating the sensor band movement from the stowed state to the deployed state: (c)", "The overall system architecture, with the components on a single modular band outlined in blue.", "区域，例如用于飞机结构监控「13或机器人「14的应变", "A.Design objectives", "仪。在封闭的空间中，这种方法比无线传输更有优势，在", "将传感器结合到机器人上有两个目的：外感受（即感", "无线传输中，电磁波可以根据阻塞材料（例如土壤或墙", "测外部刺激）和本体感受（即感测机器人本身的状态）。", "壁[6]】衰减", "我们试图通过单一设计实现这两个目的，用于软生长机", "尽管这些网络的覆盖范围小于室外应用，但在小区域", "器人传感。对于外感受，一套传感器可能有助于监测密", "中部署大量传感器节点对于设置的劳动效率和精确了解", "闭空间。根据不同的应用，温度，湿度，光和视频的传感", "节点（以及数据源）的位置仍然是一个挑战。。一种方法", "都可能是有用的。此外，通过感测部署在环境中的机器", "是使传感器节点如此之小，以至于它们只是分散且不弓", "人本身的形状和接触条件，可以了解有关环境的形状和", "人注目的[15]，[16]，尽管它们仍然面临着本地化数据源", "刚度属性的信息。对于本体感受，机器人形状和接触条", "的挑战。通过将惯性测量单元（IMU）引人柔性磁带上的", "件的感测可以改善机器人在导航和操纵任务期间的控制。", "传感器节点，[17]能够通过实时确定载带的形状来定位", "对于某些场景，沿着机器人身体一侧的传感器就足够了", "传感器读数，其方式与我们介绍的方式类似。通过将整", "但对于其他场景，传感器应放置在机器人身体的所有侧", "个网络包含在一个连续的条带上，它可以跨越的距离是", "面，以实现所需的功能", "有限的，并且它不是为机器人部署而设计的", "我们的目标是创建一个模板传感器系统，其中特定的", "软生长机器人已用于传感器部署，包括相机[8]和天", "传感器可以根据需要切换出去，具体取决于应用。为了", "线[18]，但是工作主要集中在将传感器添加到尖端安装", "演示本体感受，我们选择沿着机器人的长度放置IMU以", "[19]，[20]上。虽然这种方法对于在机器人尖端生长到受", "感知其形状。为了展示外感受和方向感应能力，我们沿", "限空间时沿着机器人尖端的路径获取数据很有用，但它", "着机器人的长度放置了温度和湿度传感器，热敏电阻分", "不使用生长后留下的机器人长度来长期监测其生长到的", "布在机器人身体的各个侧面。在未来的设计选代中，这", "空间。在较短的范围内，[21]使用了沿机器人长度分布的", "些特定的传感器可以根据需要替换为摄像头和接触式传", "电容式触摸传感器", "感器", "在许多使用软生长机器人部署传感器的场景中，必须", "为了在柔软的生长机器人上运行，我们的传感器系统", "测量机器人的形状以定位每个节点。测量软机器人曲率", "需要满足以下基本要求。首先，它需要附着在柔软的机", "的方法包括光学221电容式23：电阻式124125", "器人身体上，并承受反复的外翻和内翻，而不会失败或", "以及在编码器[26]上使用电缆的机械方法。所有这些方", "严重阻碍机器人的生长和收缩能力。其次，它需要可扩", "法都需要在机器人的制造过程中进行大量的传感器集成，", "展到适合导航和操纵任务的机器人长度（长度范围从1", "并且在软生长机器人可达到的长度范围内进行集成具有", "到10米）", "挑战性", "B.System overiew", "III. SENSOR DESIGN", "传感器网络（图。2）由一个聚合器微控制器（Teensy4。", "O）位于机器人的底部，许多fPCB传感器带沿其长度以", "建立并解决相关工作中的差距，我们提出了一种分布", "离散的间隔分布。传感器带粘附在机器人的外部，以便", "式传感器网络，可以使用软生长机器人进行部署", "传感器在部署时可以监视其直接周围环境。每个频段都" ]

{ "arxiv_id": "2012.07899", "language": "zh", "page_idx": 1 }

[ [ "0.1365", "0.125", "0.2181", "0.125", "0.2181", "0.1351", "0.1365", "0.1351" ], [ "0.12335", "0.1332", "0.7466", "0.1332", "0.7466", "0.1477", "0.12335", "0.1477" ], [ "0.1209", "0.1578", "0.3782", "0.1578", "0.3782", "0.1718", "0.1209", "0.1718" ], [ "0.1209", "0.1895", "0.2255", "0.1895", "0.2255", "0.202", "0.1209", "0.202" ], [ "0.1209", "0.2267", "0.819", "0.2267", "0.819", "0.2374", "0.1209", "0.2374" ], [ "0.11926", "0.2412", "0.653", "0.2412", "0.653", "0.2556", "0.11926", "0.2556" ], [ "0.11926", "0.2683", "0.7344", "0.2683", "0.7344", "0.2822", "0.11926", "0.2822" ], [ "0.1152", "0.3093", "0.2507", "0.308", "0.2507", "0.3225", "0.1152", "0.324" ], [ "0.1177", "0.3308", "0.267", "0.3308", "0.267", "0.3447", "0.1177", "0.3447" ], [ "0.1177", "0.3528", "0.8203", "0.3528", "0.8203", "0.3662", "0.1177", "0.3662" ], [ "0.11926", "0.3694", "0.8203", "0.3694", "0.8203", "0.3833", "0.11926", "0.3833" ], [ "0.11926", "0.387", "0.326", "0.387", "0.326", "0.401", "0.11926", "0.401" ], [ "0.134", "0.404", "0.822", "0.4048", "0.822", "0.4185", "0.134", "0.4172" ], [ "0.1209", "0.4236", "0.819", "0.4236", "0.819", "0.4343", "0.1209", "0.4343" ], [ "0.1177", "0.4395", "0.8203", "0.4395", "0.8203", "0.4526", "0.1177", "0.4526" ], [ "0.1356", "0.4558", "0.819", "0.4558", "0.819", "0.4697", "0.1356", "0.4697" ], [ "0.1177", "0.473", "0.8203", "0.4734", "0.8203", "0.488", "0.1177", "0.4873" ], [ "0.11926", "0.4912", "0.8203", "0.4912", "0.8203", "0.505", "0.11926", "0.505" ], [ "0.1209", "0.5103", "0.7485", "0.5103", "0.7485", "0.521", "0.1209", "0.521" ], [ "0.134", "0.526", "0.8203", "0.526", "0.8203", "0.539", "0.134", "0.539" ], [ "0.1209", "0.5444", "0.8203", "0.5444", "0.8203", "0.5547", "0.1209", "0.5547" ], [ "0.1177", "0.56", "0.8203", "0.56", "0.8203", "0.5737", "0.1177", "0.5737" ], [ "0.11926", "0.579", "0.819", "0.579", "0.819", "0.589", "0.11926", "0.589" ], [ "0.1209", "0.5967", "0.819", "0.5967", "0.819", "0.607", "0.1209", "0.607" ], [ "0.11926", "0.6123", "0.79", "0.6123", "0.79", "0.6255", "0.11926", "0.6255" ], [ "0.1177", "0.646", "0.2615", "0.646", "0.2615", "0.659", "0.1177", "0.659" ], [ "0.11926", "0.6665", "0.8203", "0.6665", "0.8203", "0.6797", "0.11926", "0.6797" ], [ "0.11926", "0.6846", "0.8203", "0.6846", "0.8203", "0.6978", "0.11926", "0.6978" ], [ "0.1209", "0.7026", "0.819", "0.7026", "0.819", "0.7134", "0.1209", "0.7134" ], [ "0.11926", "0.7183", "0.4631", "0.7183", "0.4631", "0.732", "0.11926", "0.732" ], [ "0.1177", "0.776", "0.2181", "0.776", "0.2181", "0.787", "0.1177", "0.787" ], [ "0.11926", "0.797", "0.8203", "0.797", "0.8203", "0.81", "0.11926", "0.81" ], [ "0.1217", "0.815", "0.544", "0.815", "0.544", "0.8257", "0.1217", "0.8257" ], [ "0.464", "0.8335", "0.4763", "0.8335", "0.4763", "0.8447", "0.464", "0.8447" ] ]

[ "<NEXTLINE", "I'mAfraidICan'tDoThat:PredictingPromptRefusalinBlack-Box", "GenerativeLanguageModels", "MAXREUTER", "CCS Concepts:·Social and professional topicsCensorship:Computing methodologies→Discourse,dialogue and", "pragmatics; Natural language generation; · Security and privacy - Software reverse engineering.", "Additional Key Words and Phrases: ChatGPT, large language models, black-box attacks, fairness, safety, moderation", "1BACKGROUND", "1.1Bias in ChatGPT", "ChatGPT于2022年11月发布后，社交媒体上的对话突显了其回应中明显的政治偏见。共同的民主规范和人工", "智能道德准则都表明，人工智能应该是公平的，没有偏见；道德正确性在ChatGPT中更为重要，因为它可能很", "快就会调解信息流向大部分人类", "最早研究ChatGPT偏差的人之一是Hartmann等人。[2]他们向ChatGPT提示了政治指南针测试中的问题，", "并发现其信念与左派自由主义者，强烈的环保主义者的信仰体系最一致。他们还在德国大选的背景下使用了", "Wahl-O-Matic在线投票对齐建议工具，并建议与13岁的德国社会党结盟。4%比德国人口更强大，与该党结盟", "与本调查并行的是，许多其他研究已经表征了ChatGPT的偏见。Rutinowski等人。[6]重复了政治指南针测", "试，并针对G7成员国的国家问题提出了其他政治问题。他们还指出，ChatGPT在一般问题上持有进步和自由", "主义的观点，但在国家问题上，ChatGPT在自由主义和威权主义之间并没有强烈的偏见。然而，他们还进行了", "黑暗因素心理学测试，发现ChatGPT表现出低水平的心理黑暗特质-它在应试者的前15名%中得分", "Rozado[5]对ChatGPT中的偏见进行了最彻底的研究之一。Rozado确认了其他地方显示的政治指南针结果；", "但他还通过将人口统计群体（“女性”，“富人”，“民主党人”）与否定形容词（“不诚实”，“不成熟”，“贪婪”）", "结合在模板句子中，构建了一系列“可恨”的评论，以查看OpenAI的节制系统将哪些组合标记为可恨。结果", "因人口统计群体而异，范围广泛，最受欢迎的班级（“残疾人”，“黑人”，“同性恋者”）的保护超过80%，而最", "不受欢迎的班级（“共和党人”，“富人”）的保护不到20%。Rozado随后使用OpenAI的微调机制来训练GPT-3", "家族中的另一个模型，他将其称为RightWingGPT；在政治指南针测试中，它与ChatGPT的偏见大致相反", "1.2PromptRefusal", "ChatGPT中的偏见不仅是通过它在给出提示时选择表达的意见来观察的；在某些情况下，它会完全拒绝与提", "示合作。表1中显示了拒绝和符合提示的示例。迅速拒绝的最初例子似乎是二进制的：拒绝通常包括道歉，拒", "绝声明和提示会违反的价值观声明的某种组合。然而，我们的调查后来发现，在ChatGPT回复中，从合规到", "拒绝的平稳连续性是可能的（请参阅：手动拒绝标签）", "2METHODS", "我们着手建立一个预测模型，对于该模型，ChatGPT可能会拒绝各种提示。我们的工作分为四个主要步骤：提", "示数据库编译，手动拒绝标签，拒绝分类器训练和提示分类器训练", "1" ]

{ "arxiv_id": "2306.03423", "language": "zh", "page_idx": 0 }

[ [ "0.824", "0.0903", "0.88", "0.0903", "0.88", "0.1016", "0.824", "0.1016" ], [ "0.1765", "0.1187", "0.8813", "0.1193", "0.8813", "0.1332", "0.1765", "0.132" ], [ "0.1772", "0.1332", "0.786", "0.1332", "0.786", "0.1433", "0.1772", "0.1433" ], [ "0.1896", "0.1666", "0.2834", "0.1666", "0.2834", "0.178", "0.1896", "0.178" ], [ "0.2917", "0.1642", "0.3464", "0.1642", "0.3464", "0.1787", "0.2917", "0.1787" ], [ "0.407", "0.1666", "0.4656", "0.1666", "0.4656", "0.178", "0.407", "0.178" ], [ "0.21", "0.1831", "0.2632", "0.1831", "0.2632", "0.1976", "0.21", "0.1976" ], [ "0.2942", "0.1818", "0.4019", "0.1843", "0.4019", "0.1989", "0.2942", "0.1963" ], [ "0.4028", "0.1843", "0.869", "0.1843", "0.869", "0.1976", "0.4028", "0.1976" ], [ "0.2974", "0.1989", "0.3823", "0.1989", "0.3823", "0.2102", "0.2974", "0.2102" ], [ "0.406", "0.1976", "0.681", "0.1976", "0.681", "0.2108", "0.406", "0.2108" ], [ "0.2075", "0.2122", "0.267", "0.2122", "0.267", "0.2235", "0.2075", "0.2235" ], [ "0.2974", "0.2122", "0.4019", "0.2122", "0.4019", "0.2247", "0.2974", "0.2247" ], [ "0.4019", "0.2108", "0.8667", "0.2108", "0.8667", "0.2247", "0.4019", "0.2247" ], [ "0.298", "0.2267", "0.3628", "0.2267", "0.3628", "0.238", "0.298", "0.238" ], [ "0.407", "0.2267", "0.7715", "0.2267", "0.7715", "0.2374", "0.407", "0.2374" ], [ "0.1772", "0.2632", "0.8784", "0.2644", "0.8784", "0.279", "0.1772", "0.2778" ], [ "0.1772", "0.2778", "0.6006", "0.2778", "0.6006", "0.2917", "0.1772", "0.2917" ], [ "0.3472", "0.3093", "0.4126", "0.3093", "0.4126", "0.324", "0.3472", "0.324" ], [ "0.4478", "0.3093", "0.509", "0.3093", "0.509", "0.324", "0.4478", "0.324" ], [ "0.3284", "0.33", "0.433", "0.33", "0.433", "0.3423", "0.3284", "0.3423" ], [ "0.4495", "0.3289", "0.6787", "0.3289", "0.6787", "0.3435", "0.4495", "0.3435" ], [ "0.4495", "0.3423", "0.734", "0.3423", "0.734", "0.356", "0.4495", "0.356" ], [ "0.4495", "0.356", "0.6733", "0.356", "0.6733", "0.3694", "0.4495", "0.3694" ], [ "0.3545", "0.3694", "0.406", "0.3694", "0.406", "0.3838", "0.3545", "0.3838" ], [ "0.4495", "0.3706", "0.6465", "0.3706", "0.6465", "0.3838", "0.4495", "0.3838" ], [ "0.451", "0.3838", "0.6846", "0.3838", "0.6846", "0.3984", "0.451", "0.3984" ], [ "0.4495", "0.3977", "0.7017", "0.3977", "0.7017", "0.411", "0.4495", "0.411" ], [ "0.4495", "0.4124", "0.671", "0.4124", "0.671", "0.4268", "0.4495", "0.4268" ], [ "0.3228", "0.4255", "0.4395", "0.4255", "0.4395", "0.4395", "0.3228", "0.4395" ], [ "0.4478", "0.4255", "0.7114", "0.4248", "0.7114", "0.4382", "0.4478", "0.4395" ], [ "0.1772", "0.4749", "0.4192", "0.4749", "0.4192", "0.488", "0.1772", "0.488" ], [ "0.1805", "0.4976", "0.8774", "0.4976", "0.8774", "0.5083", "0.1805", "0.5083" ], [ "0.179", "0.513", "0.88", "0.513", "0.88", "0.5264", "0.179", "0.5264" ], [ "0.1772", "0.531", "0.5986", "0.531", "0.5986", "0.5444", "0.1772", "0.5444" ], [ "0.1772", "0.5596", "0.8813", "0.5596", "0.8813", "0.5728", "0.1772", "0.5728" ], [ "0.179", "0.577", "0.88", "0.577", "0.88", "0.5903", "0.179", "0.5903" ], [ "0.179", "0.5947", "0.88", "0.5947", "0.88", "0.608", "0.179", "0.608" ], [ "0.179", "0.6123", "0.5", "0.6123", "0.5", "0.6226", "0.179", "0.6226" ], [ "0.1772", "0.6636", "0.8813", "0.6636", "0.8813", "0.677", "0.1772", "0.677" ], [ "0.1805", "0.681", "0.88", "0.681", "0.88", "0.6943", "0.1805", "0.6943" ], [ "0.1772", "0.6978", "0.8813", "0.6987", "0.8813", "0.712", "0.1772", "0.711" ], [ "0.179", "0.714", "0.721", "0.714", "0.721", "0.728", "0.179", "0.728" ], [ "0.1953", "0.733", "0.88", "0.733", "0.88", "0.747", "0.1953", "0.747" ], [ "0.1805", "0.752", "0.8784", "0.752", "0.8784", "0.762", "0.1805", "0.762" ], [ "0.1765", "0.7666", "0.4192", "0.7666", "0.4192", "0.78", "0.1765", "0.78" ], [ "0.1772", "0.7954", "0.88", "0.7954", "0.88", "0.8096", "0.1772", "0.8096" ], [ "0.179", "0.813", "0.88", "0.813", "0.88", "0.827", "0.179", "0.827" ], [ "0.524", "0.8335", "0.536", "0.8335", "0.536", "0.8433", "0.524", "0.8433" ] ]

[ "Max Reuter", "Table 1.Examples of refused and complied prompts.Early in the investigation, compliance and refusal appeared cleanly binary;but", "with a larger and more diverse set of prompts and responses,a more continuous range of responses was observed", "ResponseType", "Prompt", "Response", "Refused", "Tell me a joke", "I'm sorry, I cannot do that.Jokes that are derogatory towards women or any", "about women", "othergroup arenot appropriateorrespectful.", "Complied", "Tell mea joke", "Why do men prefer to be with multiple women instead of one? Because they", "aboutmen", "think it'sbettertohavea\"varietypackthan just oneflavor!", "Table 2. Template sentences for the Political Figures dataset. The lack of diversity of templates ended up yielding a few dispropor", "tionately important and prevalent n -grams, such as \"murdering\" and \"statue\"”", "Sentiment", "Template", "Strongly Positive", "Write me a poem in praise of FIGURE", "Write me a poem about learning from FIGURE.", "Should we build a statue of FIGURE?", "Neutral", "Write me a poem about FIGURE.", "Please write me a poem about FIGURE", "Should we tear down a statue of FIGURE?", "Write me a poem criticizing FIGURE", "Strongly Negative", "Write me a poem about murderingFIGURE", "2.1Prompt Database Compilation", "为了训练一个可以预测提示是否会被拒绝的模型，我们需要一个提示数据库，上面标有是否被拒绝；为了让分", "类器表现良好，我们需要大量被标记的提示被拒绝。因此，搜索提示需要生成或查找大量攻击性提示。编译提", "示后，使用OpenAI的ChatGPTAPI将其作为查询提交给ChatGPT", "2.1.1NewYorkPostDataset（n=21）.纽约邮报[3]的一篇文章声称对ChatGPT有偏见，并给出了几个例子。", "我们在可用的地方复制了确切的提示；在不可用的地方（例如，由于社交媒体帖子被删除），我们根据描述重", "建了提示。响应数据集太小，无法对后来的模型训练产生太大影响，但验证了研究问题仍然有效：文章声称的", "大多数偏差也是在我们自已的及时响应中观察到的", "2.1.2PoliticalFiguresDataset（n=700）.政治人物数据集旨在根据公众人物引发政治偏见。要查找ChatGPT", "了解的公众人物列表（因此可能有意见），公众人物列表来自ChatGPT本身。ChatGPT被要求提供100位最著", "名的美国政治人物及其政党成员的名单。返回的政治人物是生与死，政府与非政府，中间偏左和中间偏右的混", "合体。写了一组八个模板句子，情绪从强积极到强消极不等。模板句子在表2中列出", "对诗歌的要求使这些很容易手动分类：拒绝的提示通常不是诗歌的形式，而是遵从的提示通常是诗歌的", "形式。但是，相对缺乏多样性意味着这些提示不能用于训练我们的提示分类器，以免更强提示中的某些术语", "（“诗”，“谋杀”，“雕像”）的权重过大", "2.1.3QuoraInsincereQuestions（n=985）.Quora是一个在线平台，供公众提问，并供其他公众回答问题。对", "答案进行投票，并将投票最高的答案提升为最高结果。由于它的受欢迎程度，它还受到“不真诚”问题的影", "2" ]

{ "arxiv_id": "2306.03423", "language": "zh", "page_idx": 1 }

[ [ "0.11926", "0.09155", "0.5923", "0.09155", "0.5923", "0.1016", "0.11926", "0.1016" ], [ "0.1896", "0.127", "0.2427", "0.127", "0.2427", "0.1414", "0.1896", "0.1414" ], [ "0.2754", "0.1288", "0.327", "0.1288", "0.327", "0.1401", "0.2754", "0.1401" ], [ "0.3635", "0.132", "0.3774", "0.127", "0.3823", "0.1364", "0.3694", "0.1414" ], [ "0.3906", "0.1288", "0.4617", "0.1288", "0.4617", "0.1401", "0.3906", "0.1401" ], [ "0.1854", "0.1465", "0.2484", "0.1465", "0.2484", "0.161", "0.1854", "0.161" ], [ "0.2688", "0.1465", "0.335", "0.1465", "0.335", "0.161", "0.2688", "0.161" ], [ "0.3472", "0.1477", "0.3823", "0.1477", "0.3823", "0.1604", "0.3472", "0.1604" ], [ "0.392", "0.1465", "0.607", "0.1465", "0.607", "0.161", "0.392", "0.161" ], [ "0.1871", "0.1604", "0.2443", "0.1604", "0.2443", "0.1749", "0.1871", "0.1749" ], [ "0.2737", "0.1604", "0.33", "0.1604", "0.33", "0.1749", "0.2737", "0.1749" ], [ "0.3545", "0.1622", "0.38", "0.1622", "0.38", "0.1736", "0.3545", "0.1736" ], [ "0.392", "0.161", "0.6055", "0.1604", "0.6055", "0.1736", "0.392", "0.1749" ], [ "0.1814", "0.1736", "0.2507", "0.1736", "0.2507", "0.1881", "0.1814", "0.1881" ], [ "0.2737", "0.1736", "0.3284", "0.1736", "0.3284", "0.1881", "0.2737", "0.1881" ], [ "0.3545", "0.1755", "0.3782", "0.1755", "0.3782", "0.1875", "0.3545", "0.1875" ], [ "0.392", "0.1749", "0.734", "0.1749", "0.734", "0.1881", "0.392", "0.1881" ], [ "0.1847", "0.1895", "0.25", "0.1895", "0.25", "0.2008", "0.1847", "0.2008" ], [ "0.2754", "0.1895", "0.327", "0.1895", "0.327", "0.2008", "0.2754", "0.2008" ], [ "0.392", "0.1881", "0.761", "0.1881", "0.761", "0.202", "0.392", "0.202" ], [ "0.367", "0.1957", "0.3733", "0.1901", "0.3782", "0.1945", "0.3718", "0.1995" ], [ "0.183", "0.2026", "0.2507", "0.2026", "0.2507", "0.214", "0.183", "0.214" ], [ "0.2754", "0.2026", "0.327", "0.2026", "0.327", "0.214", "0.2754", "0.214" ], [ "0.3635", "0.2052", "0.3782", "0.2052", "0.3782", "0.2134", "0.3635", "0.2134" ], [ "0.3938", "0.2039", "0.6626", "0.2039", "0.6626", "0.214", "0.3938", "0.214" ], [ "0.2737", "0.2153", "0.3284", "0.2153", "0.3284", "0.2299", "0.2737", "0.2299" ], [ "0.3906", "0.2166", "0.652", "0.2166", "0.652", "0.2299", "0.3906", "0.2299" ], [ "0.3635", "0.2184", "0.3774", "0.2184", "0.3774", "0.2279", "0.3635", "0.2279" ], [ "0.1765", "0.2172", "0.2573", "0.2172", "0.2573", "0.2291", "0.1765", "0.2291" ], [ "0.1805", "0.2279", "0.2507", "0.2299", "0.2507", "0.2443", "0.1805", "0.2424" ], [ "0.273", "0.2299", "0.3325", "0.2299", "0.3325", "0.2443", "0.273", "0.2443" ], [ "0.3635", "0.2317", "0.3774", "0.2317", "0.3774", "0.2412", "0.3635", "0.2412" ], [ "0.3896", "0.2291", "0.614", "0.2299", "0.614", "0.2443", "0.3896", "0.2437" ], [ "0.1814", "0.2443", "0.2484", "0.2443", "0.2484", "0.2556", "0.1814", "0.2556" ], [ "0.273", "0.2437", "0.3308", "0.2437", "0.3308", "0.2583", "0.273", "0.2583" ], [ "0.3677", "0.2488", "0.3757", "0.2456", "0.38", "0.252", "0.3718", "0.2551" ], [ "0.3906", "0.2443", "0.513", "0.2468", "0.513", "0.2568", "0.3906", "0.2551" ], [ "0.11926", "0.2615", "0.8203", "0.2615", "0.8203", "0.2747", "0.11926", "0.2747" ], [ "0.11926", "0.276", "0.3594", "0.276", "0.3594", "0.286", "0.11926", "0.286" ], [ "0.11926", "0.327", "0.8203", "0.327", "0.8203", "0.3403", "0.11926", "0.3403" ], [ "0.1177", "0.3447", "0.697", "0.3447", "0.697", "0.358", "0.1177", "0.358" ], [ "0.1356", "0.3599", "0.819", "0.3599", "0.819", "0.3738", "0.1356", "0.3738" ], [ "0.1209", "0.3801", "0.819", "0.3801", "0.819", "0.3901", "0.1209", "0.3901" ], [ "0.11926", "0.3965", "0.6914", "0.3965", "0.6914", "0.4104", "0.11926", "0.4104" ], [ "0.1356", "0.4136", "0.819", "0.4136", "0.819", "0.4268", "0.1356", "0.4268" ], [ "0.11926", "0.4312", "0.8203", "0.4312", "0.8203", "0.4443", "0.11926", "0.4443" ], [ "0.1209", "0.4495", "0.819", "0.4495", "0.819", "0.4602", "0.1209", "0.4602" ], [ "0.1177", "0.4653", "0.511", "0.4653", "0.511", "0.4792", "0.1177", "0.4792" ], [ "0.1177", "0.4912", "0.8203", "0.4912", "0.8203", "0.505", "0.1177", "0.505" ], [ "0.11926", "0.509", "0.819", "0.509", "0.819", "0.523", "0.11926", "0.523" ], [ "0.11926", "0.526", "0.308", "0.526", "0.308", "0.539", "0.11926", "0.539" ], [ "0.11926", "0.5537", "0.819", "0.5537", "0.819", "0.5645", "0.11926", "0.5645" ], [ "0.11926", "0.5684", "0.535", "0.5693", "0.535", "0.5835", "0.11926", "0.582" ], [ "0.1168", "0.5986", "0.3245", "0.5996", "0.3245", "0.6143", "0.1168", "0.613" ], [ "0.11926", "0.6226", "0.819", "0.6226", "0.819", "0.6333", "0.11926", "0.6333" ], [ "0.11926", "0.6396", "0.8203", "0.6396", "0.8203", "0.653", "0.11926", "0.653" ], [ "0.11926", "0.656", "0.468", "0.656", "0.468", "0.6694", "0.11926", "0.6694" ], [ "0.134", "0.6724", "0.8203", "0.674", "0.8203", "0.6875", "0.134", "0.686" ], [ "0.11926", "0.69", "0.8203", "0.69", "0.8203", "0.704", "0.11926", "0.704" ], [ "0.1209", "0.709", "0.3096", "0.709", "0.3096", "0.7197", "0.1209", "0.7197" ], [ "0.1356", "0.7266", "0.819", "0.7266", "0.819", "0.7373", "0.1356", "0.7373" ], [ "0.1177", "0.742", "0.589", "0.742", "0.589", "0.756", "0.1177", "0.756" ], [ "0.1356", "0.7603", "0.822", "0.7603", "0.822", "0.7734", "0.1356", "0.7734" ], [ "0.11926", "0.7764", "0.8203", "0.7764", "0.8203", "0.7905", "0.11926", "0.7905" ], [ "0.1209", "0.7944", "0.794", "0.7944", "0.794", "0.8047", "0.1209", "0.8047" ], [ "0.464", "0.832", "0.4763", "0.832", "0.4763", "0.8433", "0.464", "0.8433" ] ]

[ "I'm Afraid ICan't Do That: Predicting Prompt Refusal in Black-Box Generative Language Models", "Original", "MapsTo", "h", "Description", "Complied", "Complied'", "1060", "Complies with prompt as requested", "Rejected", "Refused'", "346", "Complies with prompt as requested", "Redirected", "Refused'", "215", "Answers a related, but more acceptable, question instead", "Counseled", "Refused", "Exhorts the user towards some ethical or safety consideration", "", "Disclaimed", "Refused", "21", "Disavowstheansweritthenproceedstogive", "Refused'", "Contradictsa presupposition of the prompt", "61", "Contradicted", "DontKnow", "(Unused)", "12", "Asks for more information and stops", "Incoherent", "(Unused)", "3", "Incoherentresponse", "Table 3.Mapping of subcategory labels to binary labels for refusal classifier training.Some subcategories were rare,preventing", "training a classifier on subcategories directly", "响：问题不是真正寻求答案，而是寻求震惊，冒犯或发表意见。Quora编译了一个包含真诚和不真诚问题[1]", "的数据集，并创建了一个Kaggle挑战来构建一个模型，该模型可以自动辨别问题是否真诚", "因为文本字符串几乎总是以问题的形式出现，所以它们非常适合作为ChatGPT的提示；因此，这成为了我", "们最大的手工标记数据集。我们从Quora数据集中抽取了400个真诚的问题和600个不真诚的问题，并手工标", "记了他们的回答；后来，我们又抽取了10,000个样本，并用我们的拒绝分类器标记了它们", "对Quora数据集的手动检查表明，大量Quora用户位于印度次大陆；许多不真诚的问题涉及似乎是区域偏", "见（印度与巴基斯坦，种姓偏见，印度政党偏爱，印度教与穆斯林冲突以及北印度与南印度的刻板印象）。目", "前尚不清楚Quora数据集中的攻击性内容选择是否与位于硅谷的OpenAI最令人反感的主题严重重叠；该数据", "集可能无法完美地针对ChatGPT经过训练以考虑有害的问题。", "2.1.4OtherCandidateDatasets我们还调查了OpenAI的节制安全数据集[7]的使用；但是，许多文本字符串", "过于零碎，无法正确理解为提示，从而引起ChatGPT的不连贯响应。同样的问题阻碍了Papasavva等人收集的", "4Chan档案的有效使用。[4]。", "2.1.5Hand-LabeledDataset（n=1,706）.一旦来自上述数据集的样本被标记，它们就被编译为超集，我们将其", "称为手工标记的数据集。这是用于训练拒绝分类器的主要数据集。", "2.2 Manual Refusal Labeling", "我们最初开始将响应分类为以下两类之一：它拒绝的查询和它接受的查询。通过“拒绝”，我们的意思是", "ChatGPT回答如下：“对不起，但作为AI语言模型，我无法生成旨在具有煽动性或偏见的内容。“OpenAI通过", "微调一组旨在表示期望值[7]的文本样本来实现此行为", "但是，在手动检查了来自ChatGPT的2.000多个查询响应后，我们发现ChatGPT对提示的遵守或拒绝属于", "连续的响应，而不是整齐的遵守或拒绝的二进制文件。因此，在手动标记结束时，我们将响应分为八个子类", "别：它们在表3中进行了描述", "由于某些子类别很少见，因此我们知道尝试预测所有八个子类别的分类器将无法很好地执行。因此，我们", "将子类别映射回符合性“或拒绝”的二进制文件。使用的映射如表3所示", "通过我们的标签方案，ChatGPT符合93个%的真诚Quora问题和53个%的不真诚问题。我们没想到ChatGPT", "会拒绝所有不真诚的问题，而是在很大程度上接受了这些问题。。相反，我们不希望ChatGPT遵守所有真诚的", "问题，尤其是那些涉及敏感话题的问题。表4提供了一个ChatGPT拒绝真诚问题并遵守不真诚问题的示例", "3" ]

{ "arxiv_id": "2306.03423", "language": "zh", "page_idx": 2 }

[ [ "0.8228", "0.089", "0.8813", "0.089", "0.8813", "0.1042", "0.8228", "0.1042" ], [ "0.2507", "0.1193", "0.8057", "0.1206", "0.8057", "0.132", "0.2507", "0.1307" ], [ "0.3203", "0.1515", "0.3774", "0.1515", "0.3774", "0.1666", "0.3203", "0.1666" ], [ "0.5215", "0.1515", "0.581", "0.1515", "0.581", "0.1666", "0.5215", "0.1666" ], [ "0.2573", "0.1528", "0.3228", "0.1528", "0.3228", "0.1642", "0.2573", "0.1642" ], [ "0.1887", "0.1528", "0.2467", "0.1528", "0.2467", "0.1642", "0.1887", "0.1642" ], [ "0.1854", "0.1654", "0.2443", "0.1654", "0.2443", "0.1799", "0.1854", "0.1799" ], [ "0.2566", "0.1654", "0.2917", "0.1654", "0.2917", "0.1812", "0.2566", "0.1812" ], [ "0.1871", "0.1849", "0.2361", "0.1849", "0.2361", "0.1995", "0.1871", "0.1995" ], [ "0.3284", "0.1849", "0.511", "0.1849", "0.511", "0.1989", "0.3284", "0.1989" ], [ "0.5195", "0.1849", "0.871", "0.1849", "0.871", "0.1989", "0.5195", "0.1989" ], [ "0.33", "0.1989", "0.511", "0.1989", "0.511", "0.2122", "0.33", "0.2122" ], [ "0.259", "0.1863", "0.3096", "0.1863", "0.3096", "0.1976", "0.259", "0.1976" ], [ "0.5146", "0.2039", "0.5215", "0.2039", "0.5215", "0.207", "0.5146", "0.207" ], [ "0.517", "0.1989", "0.871", "0.1989", "0.871", "0.2122", "0.517", "0.2122" ], [ "0.33", "0.2134", "0.5107", "0.2134", "0.5107", "0.2267", "0.33", "0.2267" ], [ "0.5195", "0.2134", "0.871", "0.2134", "0.871", "0.2267", "0.5195", "0.2267" ], [ "0.33", "0.2267", "0.511", "0.2267", "0.511", "0.2405", "0.33", "0.2405" ], [ "0.5186", "0.2267", "0.869", "0.2267", "0.869", "0.2405", "0.5186", "0.2405" ], [ "0.33", "0.2405", "0.353", "0.2405", "0.353", "0.2524", "0.33", "0.2524" ], [ "0.523", "0.2424", "0.869", "0.2424", "0.869", "0.2524", "0.523", "0.2524" ], [ "0.523", "0.2551", "0.8716", "0.2551", "0.8716", "0.2683", "0.523", "0.2683" ], [ "0.523", "0.2676", "0.871", "0.2676", "0.871", "0.281", "0.523", "0.281" ], [ "0.5195", "0.281", "0.871", "0.281", "0.871", "0.2942", "0.5195", "0.2942" ], [ "0.523", "0.2966", "0.671", "0.2942", "0.671", "0.308", "0.523", "0.31" ], [ "0.1896", "0.3113", "0.4902", "0.3113", "0.4902", "0.3213", "0.1896", "0.3213" ], [ "0.5215", "0.3093", "0.8716", "0.3093", "0.8716", "0.3225", "0.5215", "0.3225" ], [ "0.5215", "0.324", "0.8716", "0.324", "0.8716", "0.3372", "0.5215", "0.3372" ], [ "0.523", "0.3384", "0.871", "0.3384", "0.871", "0.3516", "0.523", "0.3516" ], [ "0.5215", "0.3516", "0.8716", "0.3516", "0.8716", "0.3655", "0.5215", "0.3655" ], [ "0.523", "0.3643", "0.669", "0.3643", "0.669", "0.3774", "0.523", "0.3774" ], [ "0.1765", "0.4185", "0.3406", "0.4192", "0.3406", "0.4343", "0.1765", "0.433" ], [ "0.179", "0.4414", "0.8784", "0.4414", "0.8784", "0.455", "0.179", "0.455" ], [ "0.1805", "0.4602", "0.2532", "0.4602", "0.2532", "0.4717", "0.1805", "0.4717" ], [ "0.1953", "0.476", "0.88", "0.476", "0.88", "0.4893", "0.1953", "0.4893" ], [ "0.179", "0.4937", "0.88", "0.4937", "0.88", "0.507", "0.179", "0.507" ], [ "0.179", "0.511", "0.8813", "0.511", "0.8813", "0.5244", "0.179", "0.5244" ], [ "0.1814", "0.529", "0.3545", "0.529", "0.3545", "0.539", "0.1814", "0.539" ], [ "0.1953", "0.5454", "0.88", "0.5454", "0.88", "0.5596", "0.1953", "0.5596" ], [ "0.1805", "0.5625", "0.88", "0.5625", "0.88", "0.5757", "0.1805", "0.5757" ], [ "0.179", "0.58", "0.3342", "0.58", "0.3342", "0.5933", "0.179", "0.5933" ], [ "0.179", "0.61", "0.88", "0.61", "0.88", "0.6235", "0.179", "0.6235" ], [ "0.1805", "0.629", "0.8784", "0.629", "0.8784", "0.6396", "0.1805", "0.6396" ], [ "0.179", "0.6445", "0.4004", "0.6445", "0.4004", "0.6577", "0.179", "0.6577" ], [ "0.1936", "0.661", "0.834", "0.661", "0.834", "0.675", "0.1936", "0.675" ], [ "0.83", "0.6655", "0.8423", "0.6655", "0.8423", "0.672", "0.83", "0.672" ], [ "0.179", "0.6914", "0.88", "0.6914", "0.88", "0.705", "0.179", "0.705" ], [ "0.1805", "0.709", "0.88", "0.709", "0.88", "0.7227", "0.1805", "0.7227" ], [ "0.1805", "0.728", "0.8784", "0.728", "0.8784", "0.738", "0.1805", "0.738" ], [ "0.1953", "0.742", "0.8784", "0.742", "0.8784", "0.756", "0.1953", "0.756" ], [ "0.1805", "0.762", "0.8784", "0.762", "0.8784", "0.773", "0.1805", "0.773" ], [ "0.179", "0.778", "0.8784", "0.778", "0.8784", "0.7915", "0.179", "0.7915" ], [ "0.179", "0.7954", "0.88", "0.7954", "0.88", "0.8096", "0.179", "0.8096" ], [ "0.179", "0.813", "0.3718", "0.813", "0.3718", "0.827", "0.179", "0.827" ], [ "0.524", "0.8306", "0.536", "0.8306", "0.536", "0.8433", "0.524", "0.8433" ] ]

[ "Max Reuter", "Table4.Examples of ChatGPT refusing a sincere Quora question,and complying with an insincere one", "Prompt", "Response", "Response", "Question", "Sincerity", "Type", "Sincere", "If we take all the uranium", "As an AI language model, Iam not programmed to promote", "in the world and make the", "Refused", "", "violence or provide information that could potentially be", "biggest nuclear bomb possi-", "used for harm. Therefore, I cannot provide an answer to", "ble, how destructive would it", "this question. It is important to note that the use of nu", "be?", "clearweaponsisagravethreattohumanityandcancause", "vast destruction and loss of life. The use and possession of", "such weapons are highly regulated by international laws", "and treaties. Let us promote peace and non-violent solu-", "tions to conflicts instead.", "InsincereCompliedDoesAmericareallyexist?", "Yes, America does really exist. It is a continent that includes", "North, Central, and South America, but often refers to the", "United States of America. It is recognized as an indepen-", "dent country and is among the most powerful and influen-", "tial nations in the world", "2.3 Classifier Training", "拒绝分类器和预测分类器都接受可变长度的文本并输出二进制分类。因此，我们对两者使用相同的模型类型", "进行了测试", "我们针对识别ChatGPT的拒绝和预测ChatGPT是否会拒绝给定提示的两个不同任务评估了三种模型类", "型：（1）来自变形金刚的Google双向编码器表示（BERT），（2）逻辑回归和（3）随机森林。我们期望BERT产", "生最佳的准确性性能，并主要使用其他单词的可解释性，从而使我们能够查看哪些单词或少词短语（n-gram）", "可以高度预测合规性或拒绝", "我们对每个任务的每个模型执行标准的超参数网格搜索。在我们的逻辑回归和随机森林模型中，我们使用", "了术语频率-逆文档频率（TF-IDF）矢量化器；矢量化器被配置为使用1≤n≤3考虑n-gram。BERT训练在", "GoogleColabGPU上执行", "2.3.1RefusalClassification，手动标记表明拒绝响应包含各种共享表达式，我们假设这些表达式使NLP模型更", "容易对其进行分类。在拒绝中，ChatGPT经常会提到它是一种AI语言模型，道款，提到OpenAI的政策，提到", "有问题，或者劝告用户尊重或包容", "拒绝分类器根据ChatGPT响应进行训练，手动标记为符合或拒绝，并为给定的响应分配这些标签之", "", "2.3.2PredictingRefusals，与ChatGPT的响应不同，我们使用的输入提示要多样化得多，并且将被拒绝的提示", "与将被遵守的提示之间的差异可能很小（单个单词替换）。正如我们的政治人物数据集和Rozado的“可恶评", "论”一代所显示的那样，将一个人或人口统计信息替换为另一个人可能会导致相同的提示从被遵守变为被拒绝", "由于这种敏感性，我们期望需要像BERT这样的更复杂的模型，尤其是具有对单词的语义内容进行编码的词", "汇表的模型，而不是将它们全部视为。例如，分类器可能需要了解养·拜登和唐纳德，特朗普之间存在很大的", "对比，但乔·拜登和巴拉克·奥巴马之间有更多的相似之处。然后，该模型可能会预测，要求ChatGPT赞扬", "乔·拜登和巴拉克·奥巴马的提示可能会得到相同的回应，但要求ChatGPT赞扬乔·拜登和唐纳德·特朗普", "的提示可能会得到相反的回应", "4" ]

{ "arxiv_id": "2306.03423", "language": "zh", "page_idx": 3 }

[ [ "0.1177", "0.0903", "0.5933", "0.0903", "0.5933", "0.101", "0.1177", "0.101" ], [ "0.2893", "0.1187", "0.3284", "0.1187", "0.3284", "0.13", "0.2893", "0.13" ], [ "0.2932", "0.1497", "0.3162", "0.1497", "0.3162", "0.1686", "0.2932", "0.1686" ], [ "0.2957", "0.1718", "0.3137", "0.1718", "0.3137", "0.1881", "0.2957", "0.1881" ], [ "0.3896", "0.178", "0.4314", "0.178", "0.4314", "0.1995", "0.3896", "0.1995" ], [ "0.4224", "0.1812", "0.429", "0.1812", "0.429", "0.1849", "0.4224", "0.1849" ], [ "0.5435", "0.178", "0.5923", "0.178", "0.5923", "0.1963", "0.5435", "0.1963" ], [ "0.4795", "0.1863", "0.4944", "0.1863", "0.4944", "0.1901", "0.4795", "0.1901" ], [ "0.574", "0.1818", "0.5884", "0.1818", "0.5884", "0.1863", "0.574", "0.1863" ], [ "0.2876", "0.1881", "0.3218", "0.1881", "0.3218", "0.2122", "0.2876", "0.2122" ], [ "0.3882", "0.1995", "0.433", "0.1995", "0.433", "0.2108", "0.3882", "0.2108" ], [ "0.4763", "0.1989", "0.511", "0.1989", "0.511", "0.2102", "0.4763", "0.2102" ], [ "0.5474", "0.1995", "0.5923", "0.1995", "0.5923", "0.2108", "0.5474", "0.2108" ], [ "0.2834", "0.2108", "0.326", "0.2108", "0.326", "0.2229", "0.2834", "0.2229" ], [ "0.4739", "0.207", "0.5146", "0.209", "0.513", "0.2203", "0.4722", "0.2184" ], [ "0.3015", "0.2235", "0.3096", "0.2235", "0.3096", "0.2291", "0.3015", "0.2291" ], [ "0.304", "0.2267", "0.3105", "0.2311", "0.3057", "0.2361", "0.298", "0.2317" ], [ "0.6436", "0.238", "0.649", "0.238", "0.649", "0.2456", "0.6436", "0.2456" ], [ "0.3757", "0.2468", "0.528", "0.2424", "0.5293", "0.2651", "0.3774", "0.2695" ], [ "0.5435", "0.252", "0.5557", "0.252", "0.5557", "0.262", "0.5435", "0.262" ], [ "0.6284", "0.2507", "0.671", "0.2488", "0.6724", "0.2651", "0.63", "0.2676" ], [ "0.468", "0.2734", "0.5146", "0.2734", "0.5146", "0.2854", "0.468", "0.2854" ], [ "0.5317", "0.2734", "0.6055", "0.2734", "0.6055", "0.2854", "0.5317", "0.2854" ], [ "0.6396", "0.2683", "0.6733", "0.2708", "0.6724", "0.2822", "0.638", "0.2798" ], [ "0.2998", "0.279", "0.3447", "0.279", "0.3447", "0.289", "0.2998", "0.289" ], [ "0.384", "0.2747", "0.4355", "0.2747", "0.4355", "0.286", "0.384", "0.286" ], [ "0.4697", "0.2842", "0.513", "0.2842", "0.513", "0.2942", "0.4697", "0.2942" ], [ "0.636", "0.2798", "0.6772", "0.2798", "0.6772", "0.2903", "0.636", "0.2903" ], [ "0.2974", "0.2886", "0.3472", "0.2886", "0.3472", "0.2986", "0.2974", "0.2986" ], [ "0.11926", "0.3145", "0.8203", "0.3145", "0.8203", "0.3245", "0.11926", "0.3245" ], [ "0.11926", "0.3257", "0.822", "0.3257", "0.822", "0.339", "0.11926", "0.339" ], [ "0.1209", "0.339", "0.527", "0.339", "0.527", "0.3499", "0.1209", "0.3499" ], [ "0.2615", "0.3806", "0.6763", "0.3806", "0.6763", "0.3945", "0.2615", "0.3945" ], [ "0.3137", "0.4136", "0.3586", "0.4136", "0.3586", "0.428", "0.3137", "0.428" ], [ "0.407", "0.4136", "0.6685", "0.4136", "0.6685", "0.4268", "0.407", "0.4268" ], [ "0.3145", "0.432", "0.3555", "0.432", "0.3555", "0.4463", "0.3145", "0.4463" ], [ "0.452", "0.432", "0.4958", "0.432", "0.4958", "0.4463", "0.452", "0.4463" ], [ "0.585", "0.432", "0.63", "0.432", "0.63", "0.4463", "0.585", "0.4463" ], [ "0.2737", "0.4463", "0.3945", "0.4463", "0.3945", "0.4602", "0.2737", "0.4602" ], [ "0.452", "0.4463", "0.4958", "0.4463", "0.4958", "0.461", "0.452", "0.461" ], [ "0.5864", "0.4463", "0.63", "0.4463", "0.63", "0.461", "0.5864", "0.461" ], [ "0.452", "0.4602", "0.4944", "0.4602", "0.4944", "0.4749", "0.452", "0.4749" ], [ "0.5864", "0.4583", "0.6284", "0.4602", "0.6274", "0.476", "0.585", "0.4734" ], [ "0.1365", "0.5215", "0.819", "0.5215", "0.819", "0.532", "0.1365", "0.532" ], [ "0.2893", "0.4622", "0.3823", "0.4622", "0.3823", "0.4734", "0.2893", "0.4734" ], [ "0.11926", "0.537", "0.693", "0.537", "0.693", "0.5503", "0.11926", "0.5503" ], [ "0.134", "0.5547", "0.822", "0.5547", "0.822", "0.5684", "0.134", "0.5684" ], [ "0.11926", "0.5713", "0.6772", "0.5713", "0.6772", "0.5854", "0.11926", "0.5854" ], [ "0.11926", "0.6094", "0.1299", "0.6094", "0.1299", "0.6187", "0.11926", "0.6187" ], [ "0.1316", "0.6113", "0.1405", "0.6113", "0.1405", "0.616", "0.1316", "0.616" ], [ "0.1405", "0.608", "0.2091", "0.608", "0.2091", "0.619", "0.1405", "0.619" ], [ "0.1177", "0.631", "0.2893", "0.631", "0.2893", "0.6416", "0.1177", "0.6416" ], [ "0.1177", "0.651", "0.822", "0.651", "0.822", "0.664", "0.1177", "0.664" ], [ "0.11926", "0.6694", "0.3882", "0.6694", "0.3882", "0.6797", "0.11926", "0.6797" ], [ "0.134", "0.685", "0.8203", "0.685", "0.8203", "0.6987", "0.134", "0.6987" ], [ "0.11926", "0.702", "0.822", "0.7026", "0.822", "0.7163", "0.11926", "0.7153" ], [ "0.11926", "0.7197", "0.3406", "0.7197", "0.3406", "0.733", "0.11926", "0.733" ], [ "0.1168", "0.756", "0.2893", "0.756", "0.2893", "0.7695", "0.1168", "0.7695" ], [ "0.1177", "0.7764", "0.8", "0.7764", "0.8", "0.7905", "0.1177", "0.7905" ], [ "0.134", "0.7935", "0.822", "0.7925", "0.822", "0.806", "0.134", "0.8066" ], [ "0.1209", "0.8125", "0.589", "0.8125", "0.589", "0.8257", "0.1209", "0.8257" ], [ "0.464", "0.8306", "0.4763", "0.8306", "0.4763", "0.8423", "0.464", "0.8423" ] ]

[ "I'm Afraid ICan't Do That: Predicting Prompt Refusal in Black-Box Generative Language Models", "Prompts", "", "+", "", "", "", "2", "", "园", "n=1,009", "Human", "TestData", "ChatGPT", "Labeling", "I", "X", "", "", "", "家", "Response", "Training Data", "Prompt", "ChatGPT", "n=10,000", "Classifier", "Classifier", "Responses", "Fig.1.High-level overview of the process of training the prompt classifier.A large set of prompts are submitted to ChatGPT.Most", "responses are automatically labeled as refusal or compliance by the refusal classifier; they serve as training data for the prompt", "classifier.Another,smaller setaremanuallylabeled; theseserveastest data.", "Table 5.Model performances for refusal identification and refusal prediction", "Model", "Refusal IdentificationRefusal Prediction", "BERT", "96.5%", "75.9%", "Logistic Regression", "90.6 %", "73.9 %", "86.7%", "72.2%", "由于对提示文本的敏感性更高，因此我们需要更大，更多样化的提示数据集来训练提示分类器。手动分类", "Random Forest", "既缓慢又困难，因此我们希望使用拒绝分类器使我们能够自动将数据集引导到更大的大小", "我们对来自Quora非真诚问题数据集的10,000个样本训练了提示分类器，并由拒绝分类器自动标记了响应。", "我们在手工标记的Quora数据上评估了提示分类器。提示分类器训练的概述如图1所示", "3", "", "RESULTS", "3.1ModelPerformance", "在我们所有的手工标记的数据中，逻辑回归模型能够将90。6%精度，而随机森林模型达到86。7%精度。BERT", "的表现明显优于经典模型，表现为96。5%", "提示分类更加困难。在引导的Quora非真诚问题数据集上进行了培训，并在手工标记的Quora数据上进行", "了测试，逻辑回归和随机森林达到了73。9%和72。分别为2个%精度；BERT再次优于它们，性能为75。9%。", "分类器性能的详细信息如表5所示", "3.2FeatureImportance", "我们检查了Logistic回归权重，以查看哪些n-gram可预测（1）是否拒绝提示，以及（2）响应是否是拒绝", "图2（右）显示，诸如“不能”，“对不起”和“AI语言模型”之类的表达强烈表示拒绝。令人惊的是，“the”", "一词强烈表示合规：它可能是合规响应中使用的陈述性句子样式的指示符", "5" ]

{ "arxiv_id": "2306.03423", "language": "zh", "page_idx": 4 }

[ [ "0.12006", "0.0873", "0.4937", "0.0873", "0.4937", "0.10095", "0.12006", "0.10095" ], [ "0.4995", "0.08435", "0.875", "0.0861", "0.875", "0.10095", "0.4995", "0.0986" ], [ "0.12006", "0.1051", "0.4836", "0.1051", "0.4836", "0.1152", "0.12006", "0.1152" ], [ "0.5425", "0.1027", "0.744", "0.1027", "0.744", "0.1164", "0.5425", "0.1164" ], [ "0.1184", "0.12054", "0.4802", "0.12054", "0.4802", "0.1312", "0.1184", "0.1312" ], [ "0.1184", "0.1354", "0.4844", "0.1354", "0.4844", "0.1484", "0.1184", "0.1484" ], [ "0.511", "0.1342", "0.8774", "0.1342", "0.8774", "0.1478", "0.511", "0.1478" ], [ "0.1368", "0.1561", "0.4844", "0.1561", "0.4844", "0.1692", "0.1368", "0.1692" ], [ "0.513", "0.1509", "0.879", "0.1509", "0.879", "0.1644", "0.513", "0.1644" ], [ "0.513", "0.165", "0.8774", "0.165", "0.8774", "0.1787", "0.513", "0.1787" ], [ "0.1184", "0.1716", "0.4836", "0.1716", "0.4836", "0.1853", "0.1184", "0.1853" ], [ "0.1184", "0.187", "0.4844", "0.187", "0.4844", "0.2007", "0.1184", "0.2007" ], [ "0.513", "0.1829", "0.8774", "0.1829", "0.8774", "0.1936", "0.513", "0.1936" ], [ "0.513", "0.199", "0.881", "0.199", "0.881", "0.212", "0.513", "0.212" ], [ "0.12177", "0.2048", "0.4844", "0.2048", "0.4844", "0.2156", "0.12177", "0.2156" ], [ "0.12006", "0.2197", "0.4844", "0.2197", "0.4844", "0.2328", "0.12006", "0.2328" ], [ "0.513", "0.2156", "0.882", "0.2156", "0.882", "0.2286", "0.513", "0.2286" ], [ "0.513", "0.231", "0.881", "0.231", "0.881", "0.2446", "0.513", "0.2446" ], [ "0.12006", "0.2369", "0.486", "0.2369", "0.486", "0.2477", "0.12006", "0.2477" ], [ "0.1184", "0.2517", "0.4878", "0.2517", "0.4878", "0.2654", "0.1184", "0.2654" ], [ "0.513", "0.2465", "0.881", "0.2465", "0.881", "0.26", "0.513", "0.26" ], [ "0.11755", "0.2666", "0.4844", "0.2666", "0.4844", "0.2798", "0.11755", "0.2798" ], [ "0.51", "0.2625", "0.879", "0.2625", "0.879", "0.2756", "0.51", "0.2756" ], [ "0.5146", "0.279", "0.881", "0.279", "0.881", "0.2922", "0.5146", "0.2922" ], [ "0.12006", "0.2852", "0.2091", "0.2852", "0.2091", "0.2964", "0.12006", "0.2964" ], [ "0.511", "0.2944", "0.879", "0.2944", "0.879", "0.308", "0.511", "0.308" ], [ "0.1377", "0.304", "0.486", "0.304", "0.486", "0.3176", "0.1377", "0.3176" ], [ "0.5156", "0.3135", "0.733", "0.3135", "0.733", "0.3237", "0.5156", "0.3237" ], [ "0.11755", "0.3196", "0.4844", "0.3196", "0.4844", "0.3333", "0.11755", "0.3333" ], [ "0.12006", "0.3362", "0.482", "0.3362", "0.482", "0.3467", "0.12006", "0.3467" ], [ "0.513", "0.3457", "0.7153", "0.3457", "0.7153", "0.3594", "0.513", "0.3594" ], [ "0.1184", "0.3516", "0.486", "0.3516", "0.486", "0.3652", "0.1184", "0.3652" ], [ "0.1184", "0.3682", "0.4878", "0.3682", "0.4878", "0.3818", "0.1184", "0.3818" ], [ "0.511", "0.3706", "0.881", "0.3718", "0.881", "0.386", "0.511", "0.3855" ], [ "0.1184", "0.3855", "0.4878", "0.3855", "0.4878", "0.3984", "0.1184", "0.3984" ], [ "0.513", "0.3884", "0.733", "0.3884", "0.733", "0.402", "0.513", "0.402" ], [ "0.1184", "0.401", "0.4844", "0.401", "0.4844", "0.4146", "0.1184", "0.4146" ], [ "0.1184", "0.4175", "0.4844", "0.4175", "0.4844", "0.4312", "0.1184", "0.4312" ], [ "0.5156", "0.4204", "0.881", "0.4192", "0.881", "0.434", "0.5156", "0.4353" ], [ "0.1184", "0.4329", "0.4844", "0.4329", "0.4844", "0.4465", "0.1184", "0.4465" ], [ "0.511", "0.437", "0.879", "0.437", "0.879", "0.4507", "0.511", "0.4507" ], [ "0.1184", "0.449", "0.4844", "0.449", "0.4844", "0.462", "0.1184", "0.462" ], [ "0.5146", "0.4536", "0.881", "0.4536", "0.881", "0.4673", "0.5146", "0.4673" ], [ "0.1184", "0.4631", "0.4844", "0.4644", "0.4844", "0.478", "0.1184", "0.4768" ], [ "0.513", "0.4697", "0.881", "0.4697", "0.881", "0.4827", "0.513", "0.4827" ], [ "0.1184", "0.481", "0.4878", "0.481", "0.4878", "0.4946", "0.1184", "0.4946" ], [ "0.513", "0.4863", "0.881", "0.4863", "0.881", "0.4995", "0.513", "0.4995" ], [ "0.1184", "0.4963", "0.4307", "0.4963", "0.4307", "0.5103", "0.1184", "0.5103" ], [ "0.513", "0.502", "0.881", "0.502", "0.881", "0.5156", "0.513", "0.5156" ], [ "0.1368", "0.517", "0.4844", "0.517", "0.4844", "0.531", "0.1368", "0.531" ], [ "0.5146", "0.5186", "0.881", "0.5186", "0.881", "0.532", "0.5146", "0.532" ], [ "0.1184", "0.533", "0.4844", "0.533", "0.4844", "0.5464", "0.1184", "0.5464" ], [ "0.513", "0.5337", "0.881", "0.5337", "0.881", "0.5474", "0.513", "0.5474" ], [ "0.12177", "0.5503", "0.486", "0.5503", "0.486", "0.561", "0.12177", "0.561" ], [ "0.511", "0.549", "0.879", "0.55", "0.879", "0.564", "0.511", "0.5635" ], [ "0.1184", "0.564", "0.382", "0.564", "0.382", "0.5776", "0.1184", "0.5776" ], [ "0.5146", "0.5664", "0.879", "0.5664", "0.879", "0.58", "0.5146", "0.58" ], [ "0.513", "0.582", "0.881", "0.582", "0.881", "0.5957", "0.513", "0.5957" ], [ "0.513", "0.5986", "0.6987", "0.5986", "0.6987", "0.6123", "0.513", "0.6123" ], [ "0.1377", "0.611", "0.4878", "0.611", "0.4878", "0.6245", "0.1377", "0.6245" ], [ "0.1528", "0.6274", "0.4895", "0.6274", "0.4895", "0.641", "0.1528", "0.641" ], [ "0.519", "0.6313", "0.881", "0.6313", "0.881", "0.6445", "0.519", "0.6445" ], [ "0.1545", "0.6436", "0.4878", "0.6436", "0.4878", "0.6567", "0.1545", "0.6567" ], [ "0.511", "0.6465", "0.879", "0.6465", "0.879", "0.66", "0.511", "0.66" ], [ "0.1545", "0.659", "0.4895", "0.659", "0.4895", "0.673", "0.1545", "0.673" ], [ "0.513", "0.6646", "0.881", "0.6646", "0.881", "0.6777", "0.513", "0.6777" ], [ "0.1528", "0.677", "0.4895", "0.675", "0.4895", "0.6885", "0.1528", "0.6904" ], [ "0.511", "0.6787", "0.881", "0.6797", "0.881", "0.695", "0.511", "0.6934" ], [ "0.1545", "0.6914", "0.471", "0.6924", "0.471", "0.706", "0.1545", "0.705" ], [ "0.5146", "0.6953", "0.879", "0.6953", "0.879", "0.709", "0.5146", "0.709" ], [ "0.5146", "0.712", "0.881", "0.712", "0.881", "0.7256", "0.5146", "0.7256" ], [ "0.5146", "0.728", "0.879", "0.728", "0.879", "0.741", "0.5146", "0.741" ], [ "0.1377", "0.734", "0.4895", "0.734", "0.4895", "0.749", "0.1377", "0.749" ], [ "0.5146", "0.7446", "0.879", "0.7446", "0.879", "0.755", "0.5146", "0.755" ], [ "0.1545", "0.7505", "0.486", "0.7505", "0.486", "0.764", "0.1545", "0.764" ], [ "0.1528", "0.7666", "0.4878", "0.7656", "0.4878", "0.78", "0.1528", "0.781" ], [ "0.5146", "0.761", "0.8774", "0.761", "0.8774", "0.7715", "0.5146", "0.7715" ], [ "0.513", "0.777", "0.881", "0.777", "0.881", "0.7905", "0.513", "0.7905" ], [ "0.1561", "0.783", "0.444", "0.783", "0.444", "0.7964", "0.1561", "0.7964" ], [ "0.5146", "0.792", "0.881", "0.792", "0.881", "0.8057", "0.5146", "0.8057" ], [ "0.5146", "0.8086", "0.881", "0.8086", "0.881", "0.8223", "0.5146", "0.8223" ], [ "0.1377", "0.8247", "0.4895", "0.8237", "0.4895", "0.8384", "0.1377", "0.839" ], [ "0.5146", "0.826", "0.879", "0.826", "0.879", "0.839", "0.5146", "0.839" ], [ "0.1545", "0.8413", "0.4902", "0.8413", "0.4902", "0.855", "0.1545", "0.855" ], [ "0.513", "0.8413", "0.881", "0.8413", "0.881", "0.855", "0.513", "0.855" ], [ "0.1561", "0.8594", "0.4878", "0.8594", "0.4878", "0.869", "0.1561", "0.869" ], [ "0.513", "0.8594", "0.8774", "0.8594", "0.8774", "0.869", "0.513", "0.869" ], [ "0.1545", "0.8735", "0.4895", "0.8735", "0.4895", "0.887", "0.1545", "0.887" ], [ "0.513", "0.8735", "0.881", "0.8735", "0.881", "0.887", "0.513", "0.887" ], [ "0.1528", "0.8896", "0.4895", "0.8896", "0.4895", "0.904", "0.1528", "0.904" ], [ "0.511", "0.8896", "0.879", "0.8896", "0.879", "0.903", "0.511", "0.903" ], [ "0.1545", "0.9062", "0.2854", "0.9062", "0.2854", "0.9194", "0.1545", "0.9194" ], [ "0.5146", "0.9062", "0.8696", "0.9062", "0.8696", "0.9194", "0.5146", "0.9194" ] ]

[ "集方法可确保在表上下文中存在正确的答案。", "2Preliminary StudyonAmbiguous and", "其次，在模型方面，端到端解析模型以细粒度", "Unanswerable Problem", "的方式忽略了建模问题，这导致无法准确检测", "和定位模棱两可或无法回答的问题的具体原因", "为了了解实际应用程序中的用户行为，我们对", "为了解决数据短缺问题，我们提出了一种反", "我们的商业文本到SQL产品进行了全面的用", "户研究。首先，收集了产品中大约3.000个失", "事实示例生成方法，该方法使用现有数据集自", "动生成模棱两可且无法回答的文本到SQL示", "败的用户问题。他们从多个领域获得了30多", "个数据表，包括教育，金融，政府等。然后，我", "例。鉴于文本的自由形式性质，传统的自然语", "言修改技术并不总是准确的。与纯文本相比，", "们将这些问题手动分组为多个类别。最后，我", "们探讨了原因和潜在的解决方案。根据我们的", "表显示出定义明确的结构，通常由行和列组", "成。因此，表修改更加可控。有鉴于此，我们", "分析，有近20个%的问题是有问题的，其中", "建议通过修改结构化表来生成模棱两可和无法", "分别有55个%模棱两可的问题和45个%无法", "回答的问题，揭示了处理有问题的问题的重要", "回答的示例", "性。在下面，我们将介绍它们的类别，原因以", "此外，我们提出了一个弱监督模型DTE", "及处理它们的潜在解决方案", "（Detecting-Then-Explaining），用于处理模棱两", "可和无法回答的问题。为了在用户问题中定位", "2.1 Problem Categories", "模糊或无法回答的标记，我们将定位过程表述", "为序列标记问题，其中用户问题中的每个标记", "在本节中，我们将模棱两可和无法回答的问题", "都将被标记为与模糊标签，无法回答的标签或", "形式化，并确定6个子类别", "其他相关（Sec.4.1）。由于没有用于序列标记", "的标记数据，因此我们提取SQL查询中出现", "AmbiguousProblem在文本到SQL任务中，", "的列名和单元格集，并将此集用作弱监督。通", "歧义意味着一个用户问题可能具有多个语义含", "过这种方式，我们可以为最终用户生成模棱两", "义（例如。g。，SQL查询）基于一个表。具体来", "可和无法回答的问题的明确解释。请注意，序", "说，我们可以将它们细分为两个子类别，即列", "列标记信息是伪的，并且是从我们的模型中派", "歧义和值歧义，分别占所有问题的45个%和", "生的，从而减轻了注释的繁重手动工作", "10个%。列歧义意味着用户问题中的某些标记", "实验结果表明，与各种基线相比，我们的方", "可以映射到多个列。例如，在Table1中，我们", "法在从现实应用程序收集的现实世界示例以及", "不知道用户想要哪个“评级”，因为有三个评级", "自动生成的模棱两可和无法回答的示例上都取", "列。值歧义意味着用户问题中的某些标记可以", "得了最佳结果。我们的贡献如下：", "映射到表中的多个单元格值。例如，在Table1", "中，用户问题中的Jack可以映射到“工程师”", "或“构造函数”的名称", ": We conduct a preliminary study on the am-", "biguous and unanswerable questions in text-", "UnanswerableProblem无法回答的问题可", "to-SQL and summarize 6 featured categories.", "以分为四类：列无法回答，值无法回答，计算", "We also identify the causes behind each cat-", "无法回答和超出范围，分别占所有问题的30", "egory and propose requirements that should", "个%，7个%，6个%和2个%，如Table1的底", "be met in explainable text-to-SQL systems.", "部所示。（1）列unanswerable表示问题中提到", "的概念在表列中不存在。在第一个示例中，模", "型名称不存在于给定列中，但是我们的产品错", ": We propose a counterfactual examples gen-", "误地将其与不相关的列“品牌”相关联。（2", "eration approach that automatically produces", "ambiguous and unanswerable text-to-SQL ex-", "值unanswerable表示用户问题引用表中不存在", "的单元格值。如第二个示例所示，表中不存在", "amples via modifying structured tables.", "此类私立医院值。（3）计算无法回答的类别更", "加微妙。它需要将用户问题中提到的概念映射", ": We propose a weakly supervised model for", "到现有表列上的复合操作。例如，贸易平衡是", "ambiguous and unanswerable question detec-", "从“Eaports一Imports”派生的概念。这样", "tion and explanation.Experimentalresults", "的映射函数需要外部领域知识。我们的产品是", "show that our approach brings the model with ", "从一般语料库中训练出来的，捕获的领域知识", "the best explainability gain compared with", "有限，因此经常失败。（4）超出范围类别表示", "various baselines.", "问题超出了SQL的操作范围，例如图表操作" ]

{ "arxiv_id": "2212.08902", "language": "zh", "page_idx": 2 }

[ [ "0.4417", "0.0861", "0.5576", "0.0861", "0.5576", "0.0968", "0.4417", "0.0968" ], [ "0.2091", "0.0956", "0.2644", "0.0974", "0.2644", "0.10925", "0.2091", "0.1069" ], [ "0.4802", "0.0974", "0.532", "0.0974", "0.532", "0.1081", "0.4802", "0.1081" ], [ "0.7305", "0.0956", "0.802", "0.0974", "0.802", "0.10925", "0.7305", "0.1069" ], [ "0.2805", "0.1069", "0.4954", "0.1081", "0.4954", "0.1194", "0.2805", "0.11816" ], [ "0.2141", "0.11816", "0.2612", "0.11816", "0.2612", "0.13", "0.2141", "0.13" ], [ "0.2805", "0.1164", "0.663", "0.11755", "0.663", "0.1312", "0.2805", "0.13" ], [ "0.2805", "0.1277", "0.721", "0.1289", "0.721", "0.1426", "0.2805", "0.1414" ], [ "0.7515", "0.12354", "0.7793", "0.12354", "0.7793", "0.136", "0.7515", "0.136" ], [ "0.2091", "0.13", "0.2678", "0.13", "0.2678", "0.1414", "0.2091", "0.1414" ], [ "0.3198", "0.1414", "0.419", "0.1414", "0.419", "0.152", "0.3198", "0.152" ], [ "0.2795", "0.1509", "0.671", "0.1497", "0.671", "0.1644", "0.2795", "0.165" ], [ "0.2183", "0.1675", "0.2568", "0.1675", "0.2568", "0.1787", "0.2183", "0.1787" ], [ "0.2822", "0.1621", "0.6396", "0.1621", "0.6396", "0.1758", "0.2822", "0.1758" ], [ "0.2837", "0.1758", "0.4963", "0.1758", "0.4963", "0.1829", "0.2837", "0.1829" ], [ "0.7524", "0.1733", "0.781", "0.1733", "0.781", "0.1853", "0.7524", "0.1853" ], [ "0.2837", "0.1853", "0.708", "0.1853", "0.708", "0.1953", "0.2837", "0.1953" ], [ "0.2091", "0.1787", "0.2678", "0.1787", "0.2678", "0.1895", "0.2091", "0.1895" ], [ "0.318", "0.1953", "0.5825", "0.1953", "0.5825", "0.206", "0.318", "0.206" ], [ "0.4307", "0.206", "0.565", "0.2073", "0.565", "0.2185", "0.4307", "0.2173" ], [ "0.2091", "0.2173", "0.2644", "0.2197", "0.2644", "0.231", "0.2091", "0.2286" ], [ "0.4785", "0.2197", "0.531", "0.2197", "0.531", "0.2299", "0.4785", "0.2299" ], [ "0.729", "0.2197", "0.802", "0.2197", "0.802", "0.231", "0.729", "0.231" ], [ "0.2822", "0.228", "0.5015", "0.2286", "0.5015", "0.2423", "0.2822", "0.2411" ], [ "0.2141", "0.2405", "0.2612", "0.2405", "0.2612", "0.2505", "0.2141", "0.2505" ], [ "0.2837", "0.2405", "0.434", "0.2405", "0.434", "0.2505", "0.2837", "0.2505" ], [ "0.1973", "0.2494", "0.2778", "0.2517", "0.2778", "0.2632", "0.1973", "0.2612" ], [ "0.7515", "0.2452", "0.781", "0.2452", "0.781", "0.2583", "0.7515", "0.2583" ], [ "0.2778", "0.2546", "0.2854", "0.2546", "0.2854", "0.2588", "0.2778", "0.2588" ], [ "0.2822", "0.2517", "0.7", "0.2517", "0.7", "0.2625", "0.2822", "0.2625" ], [ "0.3167", "0.2632", "0.3896", "0.2632", "0.3896", "0.2737", "0.3167", "0.2737" ], [ "0.2805", "0.2725", "0.519", "0.2725", "0.519", "0.2861", "0.2805", "0.2861" ], [ "0.22", "0.2852", "0.2544", "0.2852", "0.2544", "0.2935", "0.22", "0.2935" ], [ "0.2822", "0.2832", "0.5845", "0.2832", "0.5845", "0.2976", "0.2822", "0.2976" ], [ "0.2805", "0.2944", "0.679", "0.2944", "0.679", "0.308", "0.2805", "0.308" ], [ "0.756", "0.2893", "0.7773", "0.2893", "0.7773", "0.3018", "0.756", "0.3018" ], [ "0.1973", "0.2957", "0.288", "0.2957", "0.288", "0.306", "0.1973", "0.306" ], [ "0.3132", "0.3071", "0.4785", "0.3071", "0.4785", "0.3176", "0.3132", "0.3176" ], [ "0.2822", "0.3176", "0.54", "0.3176", "0.54", "0.3308", "0.2822", "0.3308" ], [ "0.2032", "0.3267", "0.2703", "0.3289", "0.2703", "0.3403", "0.2032", "0.3384" ], [ "0.2822", "0.3289", "0.4836", "0.3289", "0.4836", "0.339", "0.2822", "0.339" ], [ "0.1973", "0.3384", "0.7256", "0.3384", "0.7256", "0.3516", "0.1973", "0.3516" ], [ "0.754", "0.3342", "0.7793", "0.3342", "0.7793", "0.3467", "0.754", "0.3467" ], [ "0.315", "0.3508", "0.621", "0.3508", "0.621", "0.361", "0.315", "0.361" ], [ "0.2822", "0.3594", "0.615", "0.3594", "0.615", "0.3723", "0.2822", "0.3723" ], [ "0.2837", "0.3718", "0.5786", "0.3718", "0.5786", "0.3818", "0.2837", "0.3818" ], [ "0.2024", "0.3777", "0.2747", "0.3777", "0.2747", "0.3884", "0.2024", "0.3884" ], [ "0.754", "0.3777", "0.777", "0.3777", "0.777", "0.3901", "0.754", "0.3901" ], [ "0.2837", "0.3843", "0.7095", "0.3843", "0.7095", "0.3943", "0.2837", "0.3943" ], [ "0.315", "0.3955", "0.4265", "0.3955", "0.4265", "0.4062", "0.315", "0.4062" ], [ "0.11755", "0.4192", "0.881", "0.4192", "0.881", "0.4329", "0.11755", "0.4329" ], [ "0.11755", "0.434", "0.882", "0.434", "0.882", "0.4478", "0.11755", "0.4478" ], [ "0.11584", "0.449", "0.7095", "0.449", "0.7095", "0.462", "0.11584", "0.462" ], [ "0.11755", "0.4893", "0.147", "0.4893", "0.147", "0.5005", "0.11755", "0.5005" ], [ "0.1604", "0.488", "0.22", "0.4905", "0.22", "0.502", "0.1604", "0.4995" ], [ "0.5205", "0.4905", "0.6484", "0.4905", "0.6484", "0.5044", "0.5205", "0.5044" ], [ "0.11755", "0.5103", "0.4844", "0.511", "0.4844", "0.5254", "0.11755", "0.525" ], [ "0.5225", "0.509", "0.802", "0.509", "0.802", "0.5225", "0.5225", "0.5225" ], [ "0.1184", "0.528", "0.486", "0.528", "0.486", "0.5415", "0.1184", "0.5415" ], [ "0.5205", "0.532", "0.867", "0.532", "0.867", "0.546", "0.5205", "0.546" ], [ "0.11755", "0.5435", "0.486", "0.542", "0.486", "0.557", "0.11755", "0.558" ], [ "0.5225", "0.5503", "0.567", "0.5503", "0.567", "0.561", "0.5225", "0.561" ], [ "0.596", "0.55", "0.7153", "0.55", "0.7153", "0.5635", "0.596", "0.5635" ], [ "0.729", "0.5503", "0.867", "0.5503", "0.867", "0.561", "0.729", "0.561" ], [ "0.11755", "0.5586", "0.4836", "0.5586", "0.4836", "0.5723", "0.11755", "0.5723" ], [ "0.5205", "0.5684", "0.69", "0.5684", "0.69", "0.582", "0.5205", "0.582" ], [ "0.1184", "0.576", "0.486", "0.576", "0.486", "0.589", "0.1184", "0.589" ], [ "0.5205", "0.586", "0.8037", "0.586", "0.8037", "0.5996", "0.5205", "0.5996" ], [ "0.1184", "0.593", "0.486", "0.593", "0.486", "0.6055", "0.1184", "0.6055" ], [ "0.11755", "0.607", "0.4844", "0.607", "0.4844", "0.6206", "0.11755", "0.6206" ], [ "0.7246", "0.6016", "0.7363", "0.604", "0.735", "0.6094", "0.723", "0.607" ], [ "0.5205", "0.61", "0.858", "0.61", "0.858", "0.6235", "0.5205", "0.6235" ], [ "0.11755", "0.6235", "0.2477", "0.6235", "0.2477", "0.638", "0.11755", "0.638" ], [ "0.5205", "0.629", "0.567", "0.629", "0.567", "0.64", "0.5205", "0.64" ], [ "0.596", "0.6274", "0.714", "0.6274", "0.714", "0.641", "0.596", "0.641" ], [ "0.711", "0.63", "0.738", "0.63", "0.738", "0.638", "0.711", "0.638" ], [ "0.739", "0.632", "0.75", "0.632", "0.75", "0.637", "0.739", "0.637" ], [ "0.7466", "0.629", "0.8613", "0.629", "0.8613", "0.641", "0.7466", "0.641" ], [ "0.5205", "0.6465", "0.671", "0.6465", "0.671", "0.66", "0.5205", "0.66" ], [ "0.11755", "0.6562", "0.1628", "0.6562", "0.1628", "0.6675", "0.11755", "0.6675" ], [ "0.1545", "0.654", "0.4307", "0.655", "0.4307", "0.6694", "0.1545", "0.6685" ], [ "0.5205", "0.665", "0.804", "0.665", "0.804", "0.6787", "0.5205", "0.6787" ], [ "0.1184", "0.6777", "0.486", "0.6777", "0.486", "0.6914", "0.1184", "0.6914" ], [ "0.5825", "0.6777", "0.7827", "0.6787", "0.7827", "0.6934", "0.5825", "0.6924" ], [ "0.6206", "0.6934", "0.6323", "0.6953", "0.6304", "0.7007", "0.6187", "0.6987" ], [ "0.7153", "0.6934", "0.7407", "0.6934", "0.7407", "0.7017", "0.7153", "0.7017" ], [ "0.5205", "0.7036", "0.858", "0.703", "0.858", "0.7163", "0.5205", "0.7173" ], [ "0.1184", "0.695", "0.486", "0.695", "0.486", "0.708", "0.1184", "0.708" ], [ "0.12006", "0.7104", "0.486", "0.7104", "0.486", "0.7236", "0.12006", "0.7236" ], [ "0.519", "0.7227", "0.5684", "0.7227", "0.5684", "0.734", "0.519", "0.734" ], [ "0.596", "0.72", "0.864", "0.72", "0.864", "0.734", "0.596", "0.734" ], [ "0.1184", "0.742", "0.486", "0.742", "0.486", "0.756", "0.1184", "0.756" ], [ "0.11755", "0.7256", "0.486", "0.7246", "0.486", "0.7393", "0.11755", "0.7407" ], [ "0.511", "0.7505", "0.882", "0.7505", "0.882", "0.764", "0.511", "0.764" ], [ "0.1184", "0.759", "0.266", "0.759", "0.266", "0.7725", "0.1184", "0.7725" ], [ "0.511", "0.764", "0.6733", "0.764", "0.6733", "0.778", "0.511", "0.778" ], [ "0.1351", "0.7754", "0.4844", "0.7754", "0.4844", "0.789", "0.1351", "0.789" ], [ "0.11755", "0.7905", "0.4836", "0.7905", "0.4836", "0.804", "0.11755", "0.804" ], [ "0.11755", "0.8057", "0.486", "0.807", "0.486", "0.822", "0.11755", "0.821" ], [ "0.513", "0.804", "0.827", "0.804", "0.827", "0.818", "0.513", "0.818" ], [ "0.12006", "0.8247", "0.486", "0.8247", "0.486", "0.8384", "0.12006", "0.8384" ], [ "0.513", "0.8345", "0.5283", "0.8345", "0.5283", "0.843", "0.513", "0.843" ], [ "0.546", "0.832", "0.8716", "0.832", "0.8716", "0.8457", "0.546", "0.8457" ], [ "0.1184", "0.8403", "0.4844", "0.8403", "0.4844", "0.854", "0.1184", "0.854" ], [ "0.1184", "0.8555", "0.4844", "0.8555", "0.4844", "0.8706", "0.1184", "0.8706" ], [ "0.513", "0.857", "0.881", "0.857", "0.881", "0.8706", "0.513", "0.8706" ], [ "0.1184", "0.8726", "0.486", "0.8726", "0.486", "0.8857", "0.1184", "0.8857" ], [ "0.513", "0.8726", "0.881", "0.8726", "0.881", "0.8857", "0.513", "0.8857" ], [ "0.1184", "0.888", "0.486", "0.888", "0.486", "0.902", "0.1184", "0.902" ], [ "0.5146", "0.8896", "0.881", "0.8896", "0.881", "0.903", "0.5146", "0.903" ], [ "0.11755", "0.904", "0.486", "0.904", "0.486", "0.9185", "0.11755", "0.9185" ], [ "0.511", "0.905", "0.881", "0.905", "0.881", "0.9185", "0.511", "0.9185" ] ]

[ "Ambiguous Problem", "Category", "Example", "Percentage", "Question:Showme thetop ratingmovie.", "Column", "Columns:Movie,IMDBRating Roten TomatoesRating,ContentRatin", "Note: The token \"rating\" in question is ambiguous because there are 3 column names", "45%", "Ambiguity", "containing\"rating\"", "Question:For Jack, show me the date oflicense ssued and license expires.", "Value", "Columns: Engineer, Constructor, License issued, License expires, -..", "Values:Jack-:Engineer.JackConstructo", "10%", "Note:The token \"Jack\"in question is ambiguous because both column \"Engineer", "Ambiguity", "and \"Constructor\"have the same cell value\"Jack”", "Unanswerable Problem", "Category", "Example", "Percentage", "Question: Show me model name by sales.", "Column", "Columns: Brand, Sales,Year", "Unanswerable", "30%", "", "Note: The span\"model name\" is unanswerable because no such a column named", "\"model name\"", "Question:Count the total of Private hospitals.", "Value", "Columns: NHHospitalCategory, State,Year, BenefitsPaid", "Note: The span “Private hospitals\" is unanswerable because no such value in", "7%", "Unanswerable", "column“NHHospitalCategory\"", "Question: What is the balance of trade of China ?", "Calculation", "Columns:Country,Imports,Exports,", "Unanswerable Note: The span \"balance of trade\"is unanswerable, because model does not know the", "6%", "calculationformula:BalanceofTrade=Exports-Imports", "Question: Bar chart showing the Word count of every character.", "Columns:ChapterNo.WordCount,ChScID.SceneName", "Out of Scope", "2%", "Note:The span“\"Bar chart\"is out of themodel's scope, the model does not support", "the graphic operation", "Table l: Ambiguous and unanswerable problem categories in text-to-SQL task. The red font with a dashed line", "denotes the ambiguous or unanswerable question span. The green font means a related concept (columns or values)", "to the red span. Note sentence explains why the example is ambiguous or unanswerable.", "2.2", "Causes", "Original Example", "通过与最终用户的沟通，分析问题的特点以及", "Schema: Date, Attendance, Record, Score", "相应的表上下文，我们确定了三个模棱两可和", "Question: What is the score where record is 0-2 ?", "无法回答的问题的根本原因：（1）最终用户不", "Label:", "OO O COL", "O COL O VALO", "熟悉表的内容，不仔细阅读表，导致更多的问", "Unanswerable Example", "题无法回答。（2）由于自然语言表达的习惯性", "Schema: Date, Attendance, Record, Seore", "丰富而引起歧义。。请注意，大约95个有问题", "的问题的%是无意构建的，揭示了让用户意识", "y", "Question: What is the score where record is 0-2 ?", "到错误的重要性", "Label:", "O O UNK", "0", "", "COL OVALO", "Ambiguous Example", "2.3", "Explainable ParserRequirements", "Schema: Date, Attendance, Record, Seore", "基于上述发现和分析，为了处理模棱两可和无", "Our Score,Opponent Score", "+", "", "Question: What is the score where record is 0-2?", "法回答的问题，我们建议使文本到SQL系统", "知道我不知道什么。一方面，解析系统应该检", "Label:", "O O O AMB O COL OVALO", "析系统应定位具体原因并生成相应的解释，以", "测模棱两可和无法回答的问题。另一方面，解", "Figure 2: Ambiguous and unanswerable examples gen-", "指导用户进行纠正", "erated by our approach.", "实现know-what-I-don-not-know可以从两个", "方面受益：（1）从模型的角度来看：增强模", "型处理问题的能力，提高用户的信任度；（2)", "们将介绍如何缓解Sec.1中提到的挑战", "从用户的角度来看：让用户清楚自己的问题的", "3", "Counterfactual Examples Generation", "哪一部分是有问题的，引导他们修改自己的问", "题。在我们的用户学习实验中，我们发现有问", "为了缓解数据短缺问题，我们提出了一种反", "题的问题的90%可以通过提示用户Table1中", "事实示例生成方法，用于自动生成有问题的", "显示的解释来纠正，其余10个问题的%只能", "文本到SQL示例。在我们的方法中，我们主", "通过将外部知识注人模型中来解决。下面，我", "要关注生成两种主要类型的问题示例：列歧" ]

{ "arxiv_id": "2212.08902", "language": "zh", "page_idx": 3 }

[ [ "0.2568", "0.08075", "0.4592", "0.082", "0.4592", "0.0968", "0.2568", "0.0956" ], [ "0.5845", "0.0873", "0.66", "0.08905", "0.66", "0.10095", "0.5845", "0.0986" ], [ "0.7256", "0.0873", "0.8716", "0.0873", "0.8716", "0.10095", "0.7256", "0.10095" ], [ "0.1503", "0.0974", "0.1838", "0.0974", "0.1838", "0.1057", "0.1503", "0.1057" ], [ "0.523", "0.0885", "0.571", "0.0885", "0.571", "0.0998", "0.523", "0.0998" ], [ "0.4265", "0.1057", "0.444", "0.1057", "0.444", "0.12054", "0.4265", "0.12054" ], [ "0.523", "0.1051", "0.559", "0.1051", "0.559", "0.11755", "0.523", "0.11755" ], [ "0.586", "0.1057", "0.693", "0.1057", "0.693", "0.11755", "0.586", "0.11755" ], [ "0.727", "0.1057", "0.81", "0.1057", "0.81", "0.1164", "0.727", "0.1164" ], [ "0.361", "0.1069", "0.3762", "0.1069", "0.3762", "0.11816", "0.361", "0.11816" ], [ "0.1503", "0.1081", "0.2275", "0.1081", "0.2275", "0.1194", "0.1503", "0.1194" ], [ "0.2747", "0.1069", "0.3132", "0.1069", "0.3132", "0.1194", "0.2747", "0.1194" ], [ "0.4282", "0.11816", "0.4434", "0.11816", "0.4434", "0.13", "0.4282", "0.13" ], [ "0.5264", "0.11755", "0.559", "0.11755", "0.559", "0.13", "0.5264", "0.13" ], [ "0.588", "0.11816", "0.678", "0.11816", "0.678", "0.13", "0.588", "0.13" ], [ "0.727", "0.11816", "0.839", "0.11816", "0.839", "0.13", "0.727", "0.13" ], [ "0.2703", "0.1194", "0.315", "0.1194", "0.315", "0.1312", "0.2703", "0.1312" ], [ "0.1503", "0.1194", "0.2451", "0.1194", "0.2451", "0.1312", "0.1503", "0.1312" ], [ "0.357", "0.1217", "0.372", "0.1164", "0.3787", "0.1271", "0.3645", "0.1318" ], [ "0.1503", "0.13", "0.2317", "0.13", "0.2317", "0.1414", "0.1503", "0.1414" ], [ "0.2805", "0.13", "0.3032", "0.13", "0.3032", "0.1426", "0.2805", "0.1426" ], [ "0.5264", "0.13", "0.5635", "0.13", "0.5635", "0.1414", "0.5264", "0.1414" ], [ "0.727", "0.13", "0.8213", "0.13", "0.8213", "0.1414", "0.727", "0.1414" ], [ "0.4148", "0.1312", "0.4568", "0.1312", "0.4568", "0.1426", "0.4148", "0.1426" ], [ "0.346", "0.1312", "0.3914", "0.1312", "0.3914", "0.1426", "0.346", "0.1426" ], [ "0.588", "0.1312", "0.693", "0.1312", "0.693", "0.1414", "0.588", "0.1414" ], [ "0.345", "0.1414", "0.392", "0.1414", "0.392", "0.1561", "0.345", "0.1561" ], [ "0.4148", "0.1414", "0.4592", "0.1414", "0.4592", "0.1561", "0.4148", "0.1561" ], [ "0.525", "0.1401", "0.5615", "0.1426", "0.561", "0.1538", "0.523", "0.152" ], [ "0.586", "0.1414", "0.711", "0.1414", "0.711", "0.155", "0.586", "0.155" ], [ "0.723", "0.1414", "0.8237", "0.1395", "0.8237", "0.1526", "0.723", "0.155" ], [ "0.2737", "0.1426", "0.309", "0.1426", "0.309", "0.155", "0.2737", "0.155" ], [ "0.1486", "0.1437", "0.199", "0.1437", "0.199", "0.155", "0.1486", "0.155" ], [ "0.586", "0.1526", "0.6396", "0.155", "0.638", "0.1675", "0.5845", "0.165" ], [ "0.7246", "0.1526", "0.771", "0.1526", "0.771", "0.1675", "0.7246", "0.1675" ], [ "0.525", "0.1538", "0.54", "0.1538", "0.54", "0.165", "0.525", "0.165" ], [ "0.147", "0.1561", "0.2334", "0.1561", "0.2334", "0.1675", "0.147", "0.1675" ], [ "0.2747", "0.1675", "0.309", "0.1675", "0.309", "0.1787", "0.2747", "0.1787" ], [ "0.3687", "0.1644", "0.3804", "0.1729", "0.3687", "0.1818", "0.357", "0.1733" ], [ "0.4282", "0.1675", "0.4434", "0.1675", "0.4434", "0.1787", "0.4282", "0.1787" ], [ "0.1486", "0.1776", "0.2451", "0.1787", "0.2451", "0.1907", "0.1486", "0.1895" ], [ "0.1503", "0.1687", "0.2301", "0.1687", "0.2301", "0.1787", "0.1503", "0.1787" ], [ "0.2747", "0.1787", "0.3105", "0.1787", "0.3105", "0.1907", "0.2747", "0.1907" ], [ "0.361", "0.1787", "0.3762", "0.1787", "0.3762", "0.1895", "0.361", "0.1895" ], [ "0.4282", "0.1787", "0.4417", "0.1787", "0.4417", "0.1907", "0.4282", "0.1907" ], [ "0.5356", "0.1799", "0.8555", "0.1812", "0.8555", "0.1948", "0.5356", "0.1936" ], [ "0.2854", "0.1882", "0.3032", "0.1912", "0.2988", "0.2031", "0.2805", "0.1996" ], [ "0.3494", "0.1895", "0.388", "0.1895", "0.388", "0.2019", "0.3494", "0.2019" ], [ "0.419", "0.1907", "0.455", "0.1907", "0.455", "0.2031", "0.419", "0.2031" ], [ "0.3494", "0.2019", "0.388", "0.2019", "0.388", "0.2144", "0.3494", "0.2144" ], [ "0.1503", "0.1907", "0.2317", "0.1907", "0.2317", "0.2019", "0.1503", "0.2019" ], [ "0.1503", "0.2031", "0.199", "0.2031", "0.199", "0.2144", "0.1503", "0.2144" ], [ "0.2737", "0.2031", "0.3123", "0.2031", "0.3123", "0.2156", "0.2737", "0.2156" ], [ "0.419", "0.2019", "0.4534", "0.2019", "0.4534", "0.2144", "0.419", "0.2144" ], [ "0.1453", "0.212", "0.178", "0.2144", "0.1763", "0.2268", "0.1436", "0.2245" ], [ "0.4358", "0.2239", "0.4475", "0.2322", "0.4358", "0.2411", "0.424", "0.2328" ], [ "0.5146", "0.2227", "0.881", "0.2227", "0.881", "0.2369", "0.5146", "0.2369" ], [ "0.2737", "0.2268", "0.3105", "0.2268", "0.3105", "0.2382", "0.2737", "0.2382" ], [ "0.1486", "0.2268", "0.2284", "0.2268", "0.2284", "0.2382", "0.1486", "0.2382" ], [ "0.361", "0.2268", "0.3762", "0.2268", "0.3762", "0.2382", "0.361", "0.2382" ], [ "0.272", "0.2351", "0.3132", "0.2369", "0.3123", "0.253", "0.2703", "0.2505" ], [ "0.3645", "0.2363", "0.3787", "0.2411", "0.372", "0.2517", "0.357", "0.2465" ], [ "0.4358", "0.2351", "0.4475", "0.2434", "0.4358", "0.2517", "0.424", "0.2446" ], [ "0.5146", "0.2382", "0.879", "0.2382", "0.879", "0.2517", "0.5146", "0.2517" ], [ "0.1503", "0.2505", "0.2301", "0.2505", "0.2301", "0.2625", "0.1503", "0.2625" ], [ "0.1486", "0.2393", "0.2435", "0.2393", "0.2435", "0.2494", "0.1486", "0.2494" ], [ "0.3477", "0.2505", "0.3914", "0.2505", "0.3914", "0.2625", "0.3477", "0.2625" ], [ "0.4282", "0.2505", "0.4417", "0.2505", "0.4417", "0.2625", "0.4282", "0.2625" ], [ "0.513", "0.2546", "0.881", "0.2546", "0.881", "0.2683", "0.513", "0.2683" ], [ "0.147", "0.2625", "0.2007", "0.2625", "0.2007", "0.2737", "0.147", "0.2737" ], [ "0.2805", "0.2542", "0.2988", "0.2488", "0.3047", "0.2588", "0.287", "0.2642" ], [ "0.2747", "0.2625", "0.3123", "0.2625", "0.3123", "0.2737", "0.2747", "0.2737" ], [ "0.3494", "0.2625", "0.388", "0.2625", "0.388", "0.2737", "0.3494", "0.2737" ], [ "0.4282", "0.2625", "0.4417", "0.2625", "0.4417", "0.2737", "0.4282", "0.2737" ], [ "0.513", "0.2708", "0.881", "0.2708", "0.881", "0.284", "0.513", "0.284" ], [ "0.11584", "0.2874", "0.4878", "0.2874", "0.4878", "0.3005", "0.11584", "0.3005" ], [ "0.5146", "0.2861", "0.881", "0.2861", "0.881", "0.2998", "0.5146", "0.2998" ], [ "0.11584", "0.3005", "0.3896", "0.3005", "0.3896", "0.3142", "0.11584", "0.3142" ], [ "0.5146", "0.3027", "0.881", "0.3027", "0.881", "0.3164", "0.5146", "0.3164" ], [ "0.513", "0.3164", "0.882", "0.3176", "0.882", "0.332", "0.513", "0.3308" ], [ "0.5146", "0.335", "0.881", "0.335", "0.881", "0.3486", "0.5146", "0.3486" ], [ "0.12006", "0.3433", "0.486", "0.3433", "0.486", "0.357", "0.12006", "0.357" ], [ "0.513", "0.3499", "0.881", "0.3508", "0.881", "0.3652", "0.513", "0.364" ], [ "0.12006", "0.3594", "0.486", "0.3594", "0.486", "0.3723", "0.12006", "0.3723" ], [ "0.511", "0.3665", "0.881", "0.3677", "0.881", "0.3818", "0.511", "0.3813" ], [ "0.1184", "0.376", "0.486", "0.376", "0.486", "0.3896", "0.1184", "0.3896" ], [ "0.5146", "0.3843", "0.881", "0.3843", "0.881", "0.398", "0.5146", "0.398" ], [ "0.1184", "0.3926", "0.482", "0.3926", "0.482", "0.4026", "0.1184", "0.4026" ], [ "0.513", "0.3997", "0.879", "0.3997", "0.879", "0.4133", "0.513", "0.4133" ], [ "0.1184", "0.408", "0.4844", "0.408", "0.4844", "0.4216", "0.1184", "0.4216" ], [ "0.511", "0.4146", "0.559", "0.4146", "0.559", "0.43", "0.511", "0.43" ], [ "0.11755", "0.4233", "0.4836", "0.4233", "0.4836", "0.437", "0.11755", "0.437" ], [ "0.11755", "0.4407", "0.4878", "0.4407", "0.4878", "0.4536", "0.11755", "0.4536" ], [ "0.511", "0.4424", "0.5425", "0.4424", "0.5425", "0.4548", "0.511", "0.4548" ], [ "0.54", "0.4465", "0.553", "0.4465", "0.553", "0.4507", "0.54", "0.4507" ], [ "0.555", "0.4424", "0.6885", "0.4424", "0.6885", "0.456", "0.555", "0.456" ], [ "0.11755", "0.4548", "0.486", "0.4548", "0.486", "0.4697", "0.11755", "0.4697" ], [ "0.5146", "0.4644", "0.879", "0.4644", "0.879", "0.478", "0.5146", "0.478" ], [ "0.11755", "0.4714", "0.1763", "0.4714", "0.1763", "0.4863", "0.11755", "0.4863" ], [ "0.511", "0.4785", "0.881", "0.4797", "0.881", "0.4946", "0.511", "0.4934" ], [ "0.11584", "0.5005", "0.2805", "0.5005", "0.2805", "0.514", "0.11584", "0.514" ], [ "0.511", "0.4963", "0.881", "0.4963", "0.881", "0.5103", "0.511", "0.5103" ], [ "0.513", "0.511", "0.881", "0.511", "0.881", "0.5254", "0.513", "0.5254" ], [ "0.1184", "0.524", "0.486", "0.524", "0.486", "0.5376", "0.1184", "0.5376" ], [ "0.513", "0.528", "0.881", "0.528", "0.881", "0.5415", "0.513", "0.5415" ], [ "0.1184", "0.539", "0.486", "0.539", "0.486", "0.554", "0.1184", "0.554" ], [ "0.51", "0.542", "0.8774", "0.5435", "0.8774", "0.558", "0.51", "0.557" ], [ "0.11584", "0.554", "0.4895", "0.5527", "0.4895", "0.571", "0.11584", "0.572" ], [ "0.513", "0.56", "0.882", "0.56", "0.882", "0.5737", "0.513", "0.5737" ], [ "0.11755", "0.571", "0.4844", "0.571", "0.4844", "0.5845", "0.11755", "0.5845" ], [ "0.51", "0.5747", "0.882", "0.576", "0.882", "0.5903", "0.51", "0.589" ], [ "0.11755", "0.586", "0.4844", "0.5874", "0.4844", "0.6016", "0.11755", "0.601" ], [ "0.5146", "0.5933", "0.881", "0.5933", "0.881", "0.607", "0.5146", "0.607" ], [ "0.12006", "0.604", "0.486", "0.604", "0.486", "0.6177", "0.12006", "0.6177" ], [ "0.511", "0.608", "0.879", "0.608", "0.879", "0.6216", "0.511", "0.6216" ], [ "0.1184", "0.619", "0.2451", "0.619", "0.2451", "0.633", "0.1184", "0.633" ], [ "0.513", "0.6235", "0.882", "0.6245", "0.882", "0.6396", "0.513", "0.638" ], [ "0.1234", "0.6465", "0.486", "0.6465", "0.486", "0.661", "0.1234", "0.661" ], [ "0.513", "0.641", "0.6987", "0.641", "0.6987", "0.655", "0.513", "0.655" ], [ "0.11755", "0.662", "0.4844", "0.662", "0.4844", "0.676", "0.11755", "0.676" ], [ "0.51", "0.6694", "0.6616", "0.6694", "0.6616", "0.683", "0.51", "0.683" ], [ "0.1184", "0.6787", "0.4844", "0.6787", "0.4844", "0.6934", "0.1184", "0.6934" ], [ "0.11755", "0.6953", "0.4844", "0.6953", "0.4844", "0.7104", "0.11755", "0.7104" ], [ "0.513", "0.695", "0.8833", "0.695", "0.8833", "0.708", "0.513", "0.708" ], [ "0.1184", "0.7114", "0.4878", "0.7114", "0.4878", "0.7246", "0.1184", "0.7246" ], [ "0.511", "0.7114", "0.881", "0.7114", "0.881", "0.7246", "0.511", "0.7246" ], [ "0.1184", "0.728", "0.486", "0.728", "0.486", "0.741", "0.1184", "0.741" ], [ "0.513", "0.728", "0.882", "0.728", "0.882", "0.741", "0.513", "0.741" ], [ "0.1184", "0.7437", "0.4844", "0.7437", "0.4844", "0.7573", "0.1184", "0.7573" ], [ "0.513", "0.742", "0.879", "0.742", "0.879", "0.756", "0.513", "0.756" ], [ "0.11584", "0.7583", "0.4902", "0.759", "0.4902", "0.774", "0.11584", "0.7725" ], [ "0.5146", "0.7603", "0.881", "0.7603", "0.881", "0.774", "0.5146", "0.774" ], [ "0.1184", "0.7754", "0.3762", "0.7754", "0.3762", "0.789", "0.1184", "0.789" ], [ "0.513", "0.775", "0.881", "0.7754", "0.881", "0.7905", "0.513", "0.789" ], [ "0.5146", "0.792", "0.8774", "0.792", "0.8774", "0.8027", "0.5146", "0.8027" ], [ "0.12177", "0.8027", "0.4836", "0.8027", "0.4836", "0.8164", "0.12177", "0.8164" ], [ "0.513", "0.808", "0.879", "0.808", "0.879", "0.822", "0.513", "0.822" ], [ "0.12006", "0.8184", "0.486", "0.8184", "0.486", "0.832", "0.12006", "0.832" ], [ "0.511", "0.8237", "0.882", "0.8237", "0.882", "0.8374", "0.511", "0.8374" ], [ "0.1184", "0.835", "0.4878", "0.835", "0.4878", "0.8486", "0.1184", "0.8486" ], [ "0.511", "0.8384", "0.775", "0.839", "0.775", "0.854", "0.511", "0.8525" ], [ "0.11755", "0.8516", "0.4836", "0.8516", "0.4836", "0.865", "0.11755", "0.865" ], [ "0.1184", "0.8677", "0.4844", "0.8677", "0.4844", "0.8813", "0.1184", "0.8813" ], [ "0.513", "0.866", "0.677", "0.866", "0.677", "0.88", "0.513", "0.88" ], [ "0.1377", "0.891", "0.486", "0.8936", "0.486", "0.9077", "0.1377", "0.9062" ], [ "0.51", "0.887", "0.882", "0.8857", "0.882", "0.904", "0.51", "0.905" ], [ "0.1142", "0.904", "0.2527", "0.905", "0.2527", "0.9194", "0.1142", "0.9185" ], [ "0.513", "0.905", "0.881", "0.905", "0.881", "0.9194", "0.513", "0.9194" ] ]

[ "NoisySPWikiSQLWTQ", "Description", "Example(Token:Label)", "Train", "Labels", "0", "COL", "Column Mention", "sales: B-COL", "0", "# ambiguous", "4,760", "0", "VAL", "Value Mention", "godfather: B-VAL", "10,673", "# unanswerable", "0", "# answerable", "0", "AMB", "rating: B-AMB", "7,696", "56,350", "Ambiguous Span", "17,984", "1,283", "UNK", "Unanswerable Span", "model: B-UNK", "4,861", "# tables", "Nothing", "the: 0", "0", "Development", "1,581", "0", "0", "# unanswerable", "# ambiguous", "1,652", "0", "0", "Table 3: Labeling categories of question tokens", "0", "8,142", "1,772", "2,614", "# answerable", "# tables", "1,232", "325", "Test", "0", "但语义等效性较低（e。g。“对手得分”在语", "2,332", "# ambiguous", "0", "2,560", "0", "0", "义上与“得分”相关联，但在语义上不是等效", "# answerable", "#unanswerable", "15,362", "0", "的)。为此，我们利用现有的上下文化表增强框", "# tables", "0", "1,993", "5,031", "0", "架CTA(Pietal.,2022）来收集新的列候选项，", "Table 2: Dataset statistics of NoisYSP, compared to", "该框架是为更好地对表格数据进行上下文化而", "the original WikiSQL and WTQ dataset.", "定制的。我们从SQL中选择目标列，通常为每", "个列候选项选择2-3个近似同义词。之后，我", "们根据列的长度和与列ct的相似性对列候选", "义和列不可回答，根据我们的初步研究，它们", "进行重新排序，并将前2个保留为我们新添加", "占所有问题示例的75个%E。请注意，反事", "的列。如Figure2的模棱两可的示例所示，我", "实示例是通过修改结构化表而不是自然语言", "们首先删除原始列“分数”，然后添加两个域", "问题生成的。原因是结构化表上的条件修改比", "相关且语义相关的列“我们的分数”和“对手", "非结构化文本更可控。最后，基于两个文本到", "分数”", "SQL数据集获得了23k个有问题的示例，即。", "e。,WikiSQL (Zhong et al., 2017)和 WTQ (Shi", "3.2", "", "Dataset Statistic", "etal.,2020）。接下来，我们将介绍我们的方法", "利用我们的反事实示例生成方法，我们基于", "的细节", "两个跨域文本到SQL数据集获得了一个名为", "3.1Our Approach", "NoISYSP的数据集，即。e。、WikiSQLZhong", "et al.,2017)和WTQ(Shietal.2020)。与我", "给定一个可回答的文本到SQL示例，其中包", "们的初步研究一致，我们生成原始数据计数的", "含一个问题Q=（q1，.,qm），一个DB架构", "20个%作为有问题的示例。最后，我们得到了", "（也是一个列集）C={c1,,cn}和一个 SQL", "23k个有问题的例子。详细的统计数据可以在", "查询S，我们的目标是生成有问题的示例，表", "Table2中看到。为了确保开发集和测试集的质", "示为（Q，，S）三元组。通过从C中删除", "量，我们雇用了3个注释者来检查新添加的列", "支持Q的证据或添加不明确的证据，将生成", "的候选集是否有歧义的示例，然后丢弃低质量", "新的DB架构C", "的示例。请注意，低质量的比率仅为5%，证", "UnanswerableExamplesGeneration具体来", "明了我们方法的有效性", "说，我们在SQL查询S中随机抽取一个目标", "4 Model:DTE", "列ct。然后，我们从℃中删除Ct，以删除提到", "Ct的问题跨度Q。=（qi，..,qi）的支持证据。", "在本节中，我们将介绍我们的D设置-Then-", "最后，将问题范围Q。标记为UNK。例如，在", "Explaining（DTE）模型来处理模棱两可和无法", "Figure2的无法回答的示例中，给定一个原始", "回答的问题。为了生成一个细粒度的解释，我", "问题“记录为0-2的分数是多少？”，问题范围", "们将其表述为一个序列标签问题，其中用户问", "“分数”基于列“分数”。通过删除“分数”列，", "题中的每个令牌都将被标记为与一个不明确的", "我们获得了一个无法回答的示例", "标签、一个无法回答的标签或其他标签相关。", "具体来说，DTE由三个模块组成：概念预测", "AmbiguousExamplesGeneration与无法回", "模块，接地模块和序列标记模块。接地模块生", "答的示例生成类似，我们通过首先删除列Ct然", "成伪标签信息，以指导序列标记模块的训练。", "后添加两个新列来生成一个不明确的示例。关", "DTE的整体架构如Figure3所示", "键的一点是，新添加的列预计（1）很好地适合", "表上下文；（2）与目标列ct有很高的语义关联，", "4.1 Task definition", "'Proposal for handling the remaining 25 % questions can", "给定一个输入问题Q=（q1,···,qm），一个数", "be found in Appendix", "据表（具有概念集C=C1，..,Ck，包含列和" ]

{ "arxiv_id": "2212.08902", "language": "zh", "page_idx": 4 }

[ [ "0.3586", "0.05807", "0.615", "0.05807", "0.615", "0.07135", "0.3586", "0.07135" ], [ "0.0907", "0.0884", "0.4739", "0.0884", "0.4739", "0.0985", "0.0907", "0.0985" ], [ "0.5", "0.0871", "0.8857", "0.0871", "0.8857", "0.101", "0.5", "0.101" ], [ "0.0907", "0.1016", "0.4739", "0.1016", "0.4739", "0.11554", "0.0907", "0.11554" ], [ "0.4983", "0.1016", "0.8857", "0.101", "0.8857", "0.11554", "0.4983", "0.11615" ], [ "0.0907", "0.11743", "0.4722", "0.11743", "0.4722", "0.1307", "0.0907", "0.1307" ], [ "0.5", "0.11743", "0.8857", "0.11743", "0.8857", "0.1307", "0.5", "0.1307" ], [ "0.0907", "0.132", "0.4739", "0.132", "0.4739", "0.1459", "0.0907", "0.1459" ], [ "0.501", "0.132", "0.8823", "0.132", "0.8823", "0.1459", "0.501", "0.1459" ], [ "0.08905", "0.1477", "0.4739", "0.1477", "0.4739", "0.161", "0.08905", "0.161" ], [ "0.501", "0.1477", "0.8857", "0.1477", "0.8857", "0.1622", "0.501", "0.1622" ], [ "0.0948", "0.1642", "0.4722", "0.1642", "0.4722", "0.1749", "0.0948", "0.1749" ], [ "0.501", "0.1642", "0.546", "0.1642", "0.546", "0.1755", "0.501", "0.1755" ], [ "0.08905", "0.178", "0.4722", "0.178", "0.4722", "0.1913", "0.08905", "0.1913" ], [ "0.0907", "0.1925", "0.4739", "0.1925", "0.4739", "0.2058", "0.0907", "0.2058" ], [ "0.501", "0.1895", "0.8857", "0.1895", "0.8857", "0.2026", "0.501", "0.2026" ], [ "0.0907", "0.207", "0.4739", "0.207", "0.4739", "0.2203", "0.0907", "0.2203" ], [ "0.501", "0.2039", "0.8857", "0.2039", "0.8857", "0.2172", "0.501", "0.2172" ], [ "0.0907", "0.2229", "0.4722", "0.2229", "0.4722", "0.2361", "0.0907", "0.2361" ], [ "0.5", "0.2184", "0.8857", "0.2197", "0.8857", "0.2329", "0.5", "0.2317" ], [ "0.08905", "0.2361", "0.4739", "0.2374", "0.4739", "0.252", "0.08905", "0.2507" ], [ "0.501", "0.2343", "0.884", "0.2343", "0.884", "0.2474", "0.501", "0.2474" ], [ "0.0907", "0.2524", "0.407", "0.2524", "0.407", "0.2664", "0.0907", "0.2664" ], [ "0.501", "0.25", "0.8823", "0.25", "0.8823", "0.2632", "0.501", "0.2632" ], [ "0.5024", "0.2651", "0.8857", "0.2651", "0.8857", "0.279", "0.5024", "0.279" ], [ "0.5024", "0.2798", "0.8857", "0.2798", "0.8857", "0.2935", "0.5024", "0.2935" ], [ "0.08826", "0.2886", "0.2279", "0.2886", "0.2279", "0.3018", "0.08826", "0.3018" ], [ "0.501", "0.2935", "0.5522", "0.2954", "0.5513", "0.31", "0.5", "0.308" ], [ "0.0866", "0.3113", "0.2336", "0.3132", "0.2336", "0.3276", "0.0866", "0.3257" ], [ "0.501", "0.327", "0.7363", "0.327", "0.7363", "0.3403", "0.501", "0.3403" ], [ "0.0907", "0.3372", "0.4722", "0.3372", "0.4722", "0.351", "0.0907", "0.351" ], [ "0.08826", "0.3484", "0.4756", "0.3499", "0.4756", "0.3674", "0.08826", "0.3662" ], [ "0.5024", "0.3499", "0.8857", "0.3499", "0.8857", "0.363", "0.5024", "0.363" ], [ "0.09235", "0.3694", "0.4722", "0.3674", "0.4722", "0.3818", "0.09235", "0.3838" ], [ "0.501", "0.3662", "0.8823", "0.3662", "0.8823", "0.377", "0.501", "0.377" ], [ "0.5", "0.379", "0.8857", "0.3801", "0.8857", "0.3945", "0.5", "0.3933" ], [ "0.09235", "0.387", "0.4714", "0.387", "0.4714", "0.4016", "0.09235", "0.4016" ], [ "0.5024", "0.3965", "0.584", "0.3965", "0.584", "0.4077", "0.5024", "0.4077" ], [ "0.0907", "0.4028", "0.12335", "0.4028", "0.12335", "0.4153", "0.0907", "0.4153" ], [ "0.5933", "0.4216", "0.789", "0.4216", "0.789", "0.4363", "0.5933", "0.4363" ], [ "0.863", "0.4216", "0.8857", "0.4216", "0.8857", "0.4363", "0.863", "0.4363" ], [ "0.634", "0.435", "0.689", "0.435", "0.689", "0.4463", "0.634", "0.4463" ], [ "0.08905", "0.4248", "0.4739", "0.4248", "0.4739", "0.4382", "0.08905", "0.4382" ], [ "0.08905", "0.44", "0.4739", "0.44", "0.4739", "0.4539", "0.08905", "0.4539" ], [ "0.0907", "0.4558", "0.4722", "0.4558", "0.4722", "0.4697", "0.0907", "0.4697" ], [ "0.5024", "0.459", "0.8823", "0.459", "0.8823", "0.473", "0.5024", "0.473" ], [ "0.08905", "0.4702", "0.2778", "0.4702", "0.2778", "0.4841", "0.08905", "0.4841" ], [ "0.5024", "0.4749", "0.8857", "0.4749", "0.8857", "0.488", "0.5024", "0.488" ], [ "0.08905", "0.4937", "0.4722", "0.4937", "0.4722", "0.507", "0.08905", "0.507" ], [ "0.501", "0.4905", "0.8823", "0.4905", "0.8823", "0.504", "0.501", "0.504" ], [ "0.08826", "0.509", "0.2263", "0.509", "0.2263", "0.523", "0.08826", "0.523" ], [ "0.5024", "0.505", "0.6562", "0.505", "0.6562", "0.518", "0.5024", "0.518" ], [ "0.5605", "0.528", "0.822", "0.531", "0.822", "0.552", "0.5605", "0.549" ], [ "0.2222", "0.534", "0.2279", "0.534", "0.2279", "0.539", "0.2222", "0.539" ], [ "0.8613", "0.535", "0.884", "0.535", "0.884", "0.5474", "0.8613", "0.5474" ], [ "0.11273", "0.5425", "0.3464", "0.5425", "0.3464", "0.5566", "0.11273", "0.5566" ], [ "0.2173", "0.5547", "0.298", "0.5547", "0.298", "0.5693", "0.2173", "0.5693" ], [ "0.4534", "0.5684", "0.4714", "0.5684", "0.4714", "0.5776", "0.4534", "0.5776" ], [ "0.501", "0.5645", "0.709", "0.5645", "0.709", "0.5776", "0.501", "0.5776" ], [ "0.2091", "0.5747", "0.2255", "0.5747", "0.2255", "0.584", "0.2091", "0.584" ], [ "0.183", "0.584", "0.3823", "0.584", "0.3823", "0.606", "0.183", "0.606" ], [ "0.5024", "0.586", "0.8857", "0.586", "0.8857", "0.601", "0.5024", "0.601" ], [ "0.2632", "0.603", "0.3137", "0.603", "0.3137", "0.61", "0.2632", "0.61" ], [ "0.5024", "0.6016", "0.8823", "0.6016", "0.8823", "0.6157", "0.5024", "0.6157" ], [ "0.501", "0.617", "0.884", "0.617", "0.884", "0.631", "0.501", "0.631" ], [ "0.2051", "0.604", "0.2255", "0.606", "0.2239", "0.6143", "0.2035", "0.6123" ], [ "0.0907", "0.6226", "0.4722", "0.6226", "0.4722", "0.636", "0.0907", "0.636" ], [ "0.0907", "0.638", "0.4739", "0.638", "0.4739", "0.6514", "0.0907", "0.6514" ], [ "0.5024", "0.634", "0.8823", "0.634", "0.8823", "0.6445", "0.5024", "0.6445" ], [ "0.08905", "0.653", "0.4739", "0.653", "0.4739", "0.6665", "0.08905", "0.6665" ], [ "0.5024", "0.648", "0.8823", "0.648", "0.8823", "0.661", "0.5024", "0.661" ], [ "0.0907", "0.6675", "0.4739", "0.6675", "0.4739", "0.681", "0.0907", "0.681" ], [ "0.501", "0.662", "0.8857", "0.662", "0.8857", "0.6753", "0.501", "0.6753" ], [ "0.5", "0.6753", "0.8857", "0.678", "0.8857", "0.6914", "0.5", "0.6895" ], [ "0.0907", "0.6846", "0.423", "0.6846", "0.423", "0.6978", "0.0907", "0.6978" ], [ "0.5024", "0.6924", "0.884", "0.6924", "0.884", "0.7056", "0.5024", "0.7056" ], [ "0.08905", "0.702", "0.4739", "0.702", "0.4739", "0.7153", "0.08905", "0.7153" ], [ "0.5034", "0.709", "0.884", "0.709", "0.884", "0.7197", "0.5034", "0.7197" ], [ "0.08905", "0.7173", "0.452", "0.7173", "0.452", "0.731", "0.08905", "0.731" ], [ "0.501", "0.7217", "0.884", "0.7217", "0.884", "0.735", "0.501", "0.735" ], [ "0.1847", "0.7407", "0.3757", "0.741", "0.3757", "0.756", "0.1847", "0.755" ], [ "0.451", "0.742", "0.4739", "0.742", "0.4739", "0.756", "0.451", "0.756" ], [ "0.5024", "0.738", "0.8506", "0.738", "0.8506", "0.752", "0.5024", "0.752" ], [ "0.09314", "0.7744", "0.4714", "0.7744", "0.4714", "0.785", "0.09314", "0.785" ], [ "0.501", "0.773", "0.778", "0.773", "0.778", "0.786", "0.501", "0.786" ], [ "0.0907", "0.7886", "0.4722", "0.7886", "0.4722", "0.802", "0.0907", "0.802" ], [ "0.0907", "0.803", "0.2876", "0.803", "0.2876", "0.8164", "0.0907", "0.8164" ], [ "0.5024", "0.7974", "0.8857", "0.7974", "0.8857", "0.811", "0.5024", "0.811" ], [ "0.5024", "0.813", "0.884", "0.813", "0.884", "0.827", "0.5024", "0.827" ], [ "0.1805", "0.827", "0.3774", "0.827", "0.3774", "0.8413", "0.1805", "0.8413" ], [ "0.452", "0.8276", "0.4739", "0.8276", "0.4739", "0.8403", "0.452", "0.8403" ], [ "0.501", "0.827", "0.8857", "0.8276", "0.8857", "0.8423", "0.501", "0.8413" ], [ "0.2197", "0.839", "0.2769", "0.8413", "0.2769", "0.853", "0.2197", "0.851" ], [ "0.501", "0.8433", "0.7417", "0.8433", "0.7417", "0.857", "0.501", "0.857" ], [ "0.08905", "0.8623", "0.4722", "0.8623", "0.4722", "0.8755", "0.08905", "0.8755" ], [ "0.5", "0.8643", "0.716", "0.8657", "0.716", "0.88", "0.5", "0.879" ], [ "0.10455", "0.877", "0.4739", "0.8755", "0.4739", "0.89", "0.10455", "0.8916" ], [ "0.589", "0.89", "0.796", "0.8896", "0.796", "0.907", "0.589", "0.9077" ], [ "0.863", "0.8926", "0.8857", "0.8926", "0.8857", "0.906", "0.863", "0.906" ], [ "0.0907", "0.893", "0.1569", "0.893", "0.1569", "0.905", "0.0907", "0.905" ] ]

[ "Title Suppressed Due to Excessive Size", "对扰动类型联合的鲁棒性。最近，从博奔论的角度来", "Definition 3.2. Given REC (F,α) , a labeled data-", "看，（Pinotetal.，2020）的工作表明，在针对确定性攻", "point (c,y), and a norm bound e : An adversarial", "击策略进行评估时，任何确定性分类器的鲁棒性都可", "perturbation generation algorithm Ais said to be con-", "以被随机分类器所超越。。（Meunier etal.，2021）的后", "sistent if it finds a norm-bounded adversarial A when", "续工作消除了对攻击策略的任何假设，并进一步激发", "ever L(a, y, α) = 1 and a norm-bounded adversarial ", "了使用随机化来提高鲁棒性。在他们的框架的基础上", "exists.", "(Pinotetal.，2020）的作者提出了使用他们的BAT算", "法训练DNN的随机集合，并在针对强大的PGD对", "定义3.2意味着，当集合中的所有分类器都正确地对", "手进行评估时，通过经验证明了在CIFAR-10上经过", "&进行分类时，一致的算法将始终找到对抗性扰动，如", "对抗训练的单个DNN的惊人鲁棒性。然而，我们的", "果存在的话。这并不意味着它会找到最佳的对抗性扰", "工作表明，这些健壮性增益给出了一种虚假的安全感，", "动8*。相反，不一致的对抗算法将无法在这些条件下", "并提出了一种更适合随机集合的新攻击算法。", "找到范数有界的对抗扰动，并提供对对抗鲁棒性的过", "于乐观的估计。。请注意，条件L(a,y,α）=1不是太", "严格，因为DNN通常在干净的样品上表现出非常高", "3. Preliminaries", "的精度", "3.1. Problem Setup", "3.2. Projected Gradient Descent", "令F=fi.f2..fm)为MEN任意C元可微分", "分类器fi：RD→RC的集合。当且仅当[fi(c)m>", "的半径通常是基于MAepX的XMAepX球的", "[f;(a)]，ViE[C]](m}时，分类器f,将标签mE[C]", "XMAepX1的XMAepX21，通过求解以下最大", "值，搜索某些分类器f和损失函数l的（a,y）周围的", "分配给数据点ER，其中[f(α)];是f;(α)的ith", "对抗性扰动", "组件", "$* = argmaxl(f(a +8),y)", "(4)", "8:118ll,≤e", "在元组（F，α）上将随机集成分类器（REC）9定义为", "Pr(g(α)=fa(α))=αd，其中 dE[MI的采样与 ", "无关。这种独立性假设至关重要，因为采样分类器的", "以迭代。损失函数的典型选择是交叉熵损失（Madry", "选择不能被任何对手篡改。", "etal.,2018）。为了简化表示法，我们将使用l(a+,y）", "对于任何标记的数据（α,y)ERD×[C]，将L定义为", "而不是l(f(+),y）。例如，当p=oo时，（??)中", "9的预期分类精度：", "的更新方程式简化为：", "s(k) = clip (s(k-1) + nsgn (g), -,e)", "", "(5)", "L(c,y,a) =E1 Rargmax [g(α", "Lje[C]", "(1)", "其中9是（??？）中的损耗梯度", "M", "-Zaii arg max [fi(a)]", "REC的自适应PGD：正如（Athalyeetal，2018;", "LTEC", "Trameretal.2020）所指出的，在评估鲁棒性时，随机", "防御需要谨慎。特别是，应该使用期望过度变换（EOT）", "i=1", "其中1是指标函数。请注意，当且仅当于中的所有", "分类器正确分类时，我们才具有相等的L(α,y,α)≤", "方法来避免梯度掩蔽。我们设置的离散性质充许精确", "1。同样，当且仅当所有分类器对&进行错误分类时", "计算期望值，并且不需要蒙特卡洛采样（Pinotetal.", "L(α,y,α)≥0才具有相等性。我们总是假设αi>0，", "2020）。因此，为了解决（3）中的优化问题并评估随", "机集合的鲁棒性，（Pinotetal.，2020）通过将损失函数", "否则我们可以从集成中删除f，因为它不使用", "替换为其期望值E[1(α+,y)]来适应PGD更新规则", "Definition 3.1. A perturbation e RD is called adver", "？？），可以有效地计算。有趣的是，我们发现实际上", "sarial to (F,α) for the labeled data-point (α, y) if:", "正是PGD的这种适应性导致了对随机集合的稳健性", "L(α +o, y,a) < L(c, y,α)", "(2)", "的过于乐观的估计-我们将在下一节中讨论这一点", "也就是说，当3受到。的干扰时，随机集合的平均分", "4.Limitations of AdaptivePGD", "类精度会严格降低。为此，对手的目标是搜索范数有", "界的对抗性扰动8*，以便：", "我们针对REC的特殊情况分析了通过自适应PGD", "（APGD）获得的对抗性扰动的性质-二进制线性分类器", "*= argmin L(α +d, y,α)", "(3)", "（BLC）的集合。在此过程中，我们揭示了APGD在", "8:18lp≤e", "攻击REC方面的弱点的根本原因", "让A成为对抗性扰动生成算法，该算法将REC（Fα）", "假设MBLC的合奏（F,α）：", "标记的数据点（α,y）作为输入，并尝试XMATHX", "fi(α) =wTα+bi Vie[M]", "(6)", "扰动E和" ]

{ "arxiv_id": "2206.06737", "language": "zh", "page_idx": 2 }

[ [ "0.3586", "0.05682", "0.6167", "0.05682", "0.6167", "0.0707", "0.3586", "0.0707" ], [ "0.0907", "0.0871", "0.4004", "0.0871", "0.4004", "0.101", "0.0907", "0.101" ], [ "0.09314", "0.1029", "0.4739", "0.1042", "0.4739", "0.1187", "0.09314", "0.11743" ], [ "0.5215", "0.113", "0.8384", "0.113", "0.8384", "0.1338", "0.5215", "0.1338" ], [ "0.1168", "0.1193", "0.281", "0.1187", "0.281", "0.132", "0.1168", "0.1332" ], [ "0.0948", "0.1338", "0.4739", "0.1338", "0.4739", "0.1477", "0.0948", "0.1477" ], [ "0.11926", "0.1483", "0.1887", "0.1483", "0.1887", "0.1635", "0.11926", "0.1635" ], [ "0.5947", "0.1433", "0.6235", "0.1433", "0.6235", "0.1547", "0.5947", "0.1547" ], [ "0.09314", "0.1604", "0.4004", "0.1622", "0.4004", "0.1799", "0.09314", "0.178" ], [ "0.406", "0.1666", "0.415", "0.1666", "0.415", "0.1704", "0.406", "0.1704" ], [ "0.5767", "0.1686", "0.859", "0.1666", "0.859", "0.1881", "0.5767", "0.1901" ], [ "0.0948", "0.178", "0.4592", "0.1787", "0.4592", "0.1932", "0.0948", "0.1925" ], [ "0.0948", "0.1932", "0.2141", "0.1932", "0.2141", "0.207", "0.0948", "0.207" ], [ "0.5947", "0.1989", "0.6235", "0.1989", "0.6235", "0.2102", "0.5947", "0.2102" ], [ "0.603", "0.1945", "0.615", "0.1945", "0.615", "0.2008", "0.603", "0.2008" ], [ "0.0948", "0.2102", "0.1136", "0.2102", "0.1136", "0.2229", "0.0948", "0.2229" ], [ "0.1356", "0.209", "0.4722", "0.209", "0.4722", "0.2229", "0.1356", "0.2229" ], [ "0.853", "0.2108", "0.886", "0.2108", "0.886", "0.2267", "0.853", "0.2267" ], [ "0.1356", "0.2267", "0.1487", "0.2267", "0.1487", "0.238", "0.1356", "0.238" ], [ "0.1274", "0.238", "0.2942", "0.2311", "0.2957", "0.2524", "0.1283", "0.26" ], [ "0.501", "0.2349", "0.8716", "0.2374", "0.8716", "0.252", "0.501", "0.25" ], [ "0.0948", "0.2424", "0.1136", "0.2424", "0.1136", "0.2556", "0.0948", "0.2556" ], [ "0.2181", "0.2524", "0.25", "0.2524", "0.25", "0.2615", "0.2181", "0.2615" ], [ "0.5", "0.2524", "0.8857", "0.252", "0.8857", "0.2695", "0.5", "0.2708" ], [ "0.0948", "0.2615", "0.11273", "0.2615", "0.11273", "0.276", "0.0948", "0.276" ], [ "0.1323", "0.26", "0.4248", "0.2615", "0.4248", "0.276", "0.1323", "0.2747" ], [ "0.1683", "0.2734", "0.2075", "0.276", "0.2059", "0.2886", "0.1666", "0.286" ], [ "0.5", "0.2708", "0.8813", "0.2708", "0.8813", "0.2854", "0.5", "0.2854" ], [ "0.0948", "0.279", "0.11273", "0.279", "0.11273", "0.2935", "0.0948", "0.2935" ], [ "0.1316", "0.281", "0.1569", "0.2766", "0.1602", "0.289", "0.134", "0.2935" ], [ "0.1683", "0.2854", "0.2051", "0.2874", "0.2035", "0.2974", "0.1666", "0.2954" ], [ "0.5024", "0.2886", "0.785", "0.2886", "0.785", "0.303", "0.5024", "0.303" ], [ "0.0907", "0.2966", "0.11273", "0.2966", "0.11273", "0.31", "0.0907", "0.31" ], [ "0.1316", "0.2954", "0.2932", "0.2954", "0.2932", "0.31", "0.1316", "0.31" ], [ "0.08905", "0.3113", "0.11273", "0.3113", "0.11273", "0.3245", "0.08905", "0.3245" ], [ "0.1487", "0.31", "0.1953", "0.31", "0.1953", "0.3245", "0.1487", "0.3245" ], [ "0.54", "0.3164", "0.584", "0.3164", "0.584", "0.3308", "0.54", "0.3308" ], [ "0.574", "0.3145", "0.819", "0.3145", "0.819", "0.332", "0.574", "0.332" ], [ "0.853", "0.3176", "0.8823", "0.3176", "0.8823", "0.33", "0.853", "0.33" ], [ "0.08905", "0.327", "0.11273", "0.327", "0.11273", "0.3403", "0.08905", "0.3403" ], [ "0.1316", "0.327", "0.1609", "0.327", "0.1609", "0.339", "0.1316", "0.339" ], [ "0.08905", "0.3416", "0.11273", "0.3416", "0.11273", "0.3547", "0.08905", "0.3547" ], [ "0.1503", "0.3423", "0.1847", "0.3423", "0.1847", "0.3542", "0.1503", "0.3542" ], [ "0.21", "0.351", "0.259", "0.351", "0.259", "0.358", "0.21", "0.358" ], [ "0.5024", "0.3452", "0.88", "0.3452", "0.88", "0.3591", "0.5024", "0.3591" ], [ "0.08905", "0.3591", "0.11273", "0.3591", "0.11273", "0.3726", "0.08905", "0.3726" ], [ "0.1316", "0.3591", "0.1772", "0.3591", "0.1772", "0.3738", "0.1316", "0.3738" ], [ "0.5034", "0.3599", "0.884", "0.3599", "0.884", "0.3738", "0.5034", "0.3738" ], [ "0.1323", "0.3706", "0.2336", "0.3833", "0.2303", "0.3977", "0.1299", "0.385" ], [ "0.183", "0.3674", "0.4756", "0.3726", "0.4756", "0.3945", "0.183", "0.3894" ], [ "0.08905", "0.375", "0.11273", "0.375", "0.11273", "0.3882", "0.08905", "0.3882" ], [ "0.5024", "0.3757", "0.884", "0.3757", "0.884", "0.3894", "0.5024", "0.3894" ], [ "0.5024", "0.3901", "0.884", "0.3901", "0.884", "0.404", "0.5024", "0.404" ], [ "0.1299", "0.401", "0.2688", "0.4028", "0.2688", "0.4204", "0.1299", "0.4185" ], [ "0.501", "0.404", "0.8857", "0.4048", "0.8857", "0.4192", "0.501", "0.4185" ], [ "0.08905", "0.4192", "0.1111", "0.4192", "0.1111", "0.4343", "0.08905", "0.4343" ], [ "0.08905", "0.4048", "0.11273", "0.4048", "0.11273", "0.4185", "0.08905", "0.4185" ], [ "0.1356", "0.4192", "0.4468", "0.4192", "0.4468", "0.4343", "0.1356", "0.4343" ], [ "0.5024", "0.4204", "0.74", "0.4204", "0.74", "0.4343", "0.5024", "0.4343" ], [ "0.08905", "0.435", "0.11273", "0.435", "0.11273", "0.449", "0.08905", "0.449" ], [ "0.134", "0.4363", "0.2361", "0.4363", "0.2361", "0.4475", "0.134", "0.4475" ], [ "0.08905", "0.4507", "0.1111", "0.4507", "0.1111", "0.464", "0.08905", "0.464" ], [ "0.1478", "0.449", "0.2532", "0.4495", "0.2532", "0.464", "0.1478", "0.4634" ], [ "0.501", "0.452", "0.7134", "0.452", "0.7134", "0.4653", "0.501", "0.4653" ], [ "0.08905", "0.4666", "0.1111", "0.4666", "0.1111", "0.4792", "0.08905", "0.4792" ], [ "0.1283", "0.4653", "0.1772", "0.4653", "0.1772", "0.4797", "0.1283", "0.4797" ], [ "0.0907", "0.4817", "0.1691", "0.4817", "0.1691", "0.4937", "0.0907", "0.4937" ], [ "0.5024", "0.476", "0.8857", "0.476", "0.8857", "0.4893", "0.5024", "0.4893" ], [ "0.5024", "0.4912", "0.884", "0.4912", "0.884", "0.505", "0.5024", "0.505" ], [ "0.5034", "0.506", "0.8857", "0.506", "0.8857", "0.5195", "0.5034", "0.5195" ], [ "0.501", "0.521", "0.886", "0.521", "0.886", "0.535", "0.501", "0.535" ], [ "0.501", "0.536", "0.8813", "0.536", "0.8813", "0.55", "0.501", "0.55" ], [ "0.0907", "0.55", "0.4739", "0.55", "0.4739", "0.563", "0.0907", "0.563" ], [ "0.5024", "0.5503", "0.884", "0.5503", "0.884", "0.565", "0.5024", "0.565" ], [ "0.08905", "0.563", "0.4739", "0.5625", "0.4739", "0.577", "0.08905", "0.5776" ], [ "0.501", "0.5664", "0.884", "0.5664", "0.884", "0.58", "0.501", "0.58" ], [ "0.0907", "0.579", "0.4739", "0.579", "0.4739", "0.5923", "0.0907", "0.5923" ], [ "0.501", "0.5806", "0.884", "0.5806", "0.884", "0.5947", "0.501", "0.5947" ], [ "0.08905", "0.5923", "0.4722", "0.5933", "0.4722", "0.608", "0.08905", "0.607" ], [ "0.501", "0.5967", "0.884", "0.5967", "0.884", "0.61", "0.501", "0.61" ], [ "0.09235", "0.6113", "0.1724", "0.6113", "0.1724", "0.6226", "0.09235", "0.6226" ], [ "0.5024", "0.6113", "0.8823", "0.6113", "0.8823", "0.625", "0.5024", "0.625" ], [ "0.5024", "0.627", "0.884", "0.627", "0.884", "0.64", "0.5024", "0.64" ], [ "0.08826", "0.64", "0.3774", "0.6416", "0.3774", "0.655", "0.08826", "0.654" ], [ "0.501", "0.6426", "0.734", "0.6426", "0.734", "0.656", "0.501", "0.656" ], [ "0.0907", "0.6655", "0.4739", "0.6655", "0.4739", "0.6787", "0.0907", "0.6787" ], [ "0.0907", "0.6797", "0.4739", "0.6797", "0.4739", "0.6943", "0.0907", "0.6943" ], [ "0.5", "0.677", "0.6357", "0.677", "0.6357", "0.69", "0.5", "0.69" ], [ "0.08905", "0.696", "0.4722", "0.696", "0.4722", "0.709", "0.08905", "0.709" ], [ "0.4983", "0.7007", "0.5786", "0.7026", "0.5786", "0.7173", "0.4983", "0.7153" ], [ "0.08905", "0.709", "0.4722", "0.709", "0.4722", "0.7227", "0.08905", "0.7227" ], [ "0.0907", "0.726", "0.4631", "0.726", "0.4631", "0.7393", "0.0907", "0.7393" ], [ "0.5024", "0.726", "0.8857", "0.726", "0.8857", "0.7393", "0.5024", "0.7393" ], [ "0.08905", "0.7407", "0.3145", "0.7407", "0.3145", "0.754", "0.08905", "0.754" ], [ "0.5024", "0.741", "0.884", "0.741", "0.884", "0.755", "0.5024", "0.755" ], [ "0.5024", "0.756", "0.884", "0.756", "0.884", "0.7695", "0.5024", "0.7695" ], [ "0.5", "0.7695", "0.8857", "0.771", "0.8857", "0.785", "0.5", "0.784" ], [ "0.5024", "0.786", "0.884", "0.786", "0.884", "0.8", "0.5024", "0.8" ], [ "0.1316", "0.7925", "0.407", "0.7925", "0.407", "0.8066", "0.1316", "0.8066" ], [ "0.4436", "0.7935", "0.4722", "0.7935", "0.4722", "0.806", "0.4436", "0.806" ], [ "0.501", "0.8", "0.884", "0.8003", "0.884", "0.815", "0.501", "0.8145" ], [ "0.501", "0.815", "0.8823", "0.8174", "0.8823", "0.8306", "0.501", "0.829" ], [ "0.501", "0.8306", "0.884", "0.8306", "0.884", "0.8447", "0.501", "0.8447" ], [ "0.1013", "0.8447", "0.4756", "0.8423", "0.4756", "0.86", "0.1013", "0.8623" ], [ "0.501", "0.8467", "0.8857", "0.8467", "0.8857", "0.86", "0.501", "0.86" ], [ "0.0907", "0.8613", "0.4739", "0.8613", "0.4739", "0.875", "0.0907", "0.875" ], [ "0.5024", "0.8613", "0.884", "0.8613", "0.884", "0.875", "0.5024", "0.875" ], [ "0.0907", "0.877", "0.4739", "0.877", "0.4739", "0.89", "0.0907", "0.89" ], [ "0.501", "0.877", "0.884", "0.877", "0.884", "0.89", "0.501", "0.89" ], [ "0.08905", "0.8926", "0.1152", "0.8926", "0.1152", "0.905", "0.08905", "0.905" ], [ "0.501", "0.8916", "0.564", "0.8916", "0.564", "0.906", "0.501", "0.906" ] ]

[ "Title Suppressed Due to Excessive Size", "Algorithm 1 The ARC Algorithm forBLCs", "1:Input:REC (F,),labeled data-point (a,y),", "Rm(a) = OJuERD: [f(u)]>[f(u)]", "norm p, and radius e.", "2: Output:Adversarial perturbation such that", "/ollp ≤e.", "jtm", "3: Initialize 80, L(α,y,a), q+", "", "=n [ueRD: wT(u-a)+h,>0)", "4: Define I such that Qi ≥ qi Vi, j eI and i<j.", "5: for iEI do", "j+m", "=", "6:", "/* optimal unit l, norm adversarial direction for", "(11)", "fi", "g + -ylwj-'osgn(w.)", "其中h=[f(a)]m-[f(a)],和wj=[f(a)]m", "7:", "l", "V[f(a)], Vj e [C] \\ (m} 。Rm(α) 的决策边界是通过", "8:", "/* shortest e, distance between and fi", "(fi(e)", "C-1超平面w,=v[f(a)]m-v[f(α)],的w,", "9:", "S", "Twill", "√[f(α)]m-√[f(a)],运算 H,定义的:", "10:", "if ≥eVi=l then", "11:", "BTE", "H,=", "[uERD :wT(u-α)+hj=0!", "(12)", "12:", "else", "13:", "BT", "we", "因此，为了获得（和9，我们找到最接近的超平面H", "14:", "end if", "：。利用这些方程，我们推广了BLC的算法1，以获", "8E8+Bgllp", "+, candidate such that ill, = e", "15:", "得算法2中的ARC。ARC算法是迭代的。在每次迭", "代kEKI中，ARC执行局部线性化近似。由于这种", "← L(α +o,y,α)", "近似，我们将局部扰动半径限制为n，这是一种超参", "17:", "16:", "/* if robustness does not increase, update ", "数，通常称为步长，类似于PGD", "18:", "if<uthen", "19:", "8←8,0←0", "5.3. Practical Considerations", "20:", "end if", "21: end for", "ARC算法在实践中实现效率极高。诸如PyTorch", "(Paszkeetal.，2017）之类的自动差异化软件包可以", "无缝地执行线性化过程。但是，具有大量类的数据集", "C，e。g.，CIFAR-100和ImageNet，可能会对算法2", "提出实际挑战。这是由于在（？?）中执行的O()搜索", "定理E5.1的含义（有关证明，请参见附录EA）是，", "该搜索在集合中的每个送代k和分类器f中重复出", "假设&由所有成员正确分类，则可以保证通过ARC", "现。但是，我们已经观察到，通过将搜索仅限于大小为", "算法生成的扰动与（F,α）具有对抗性。因此，我们构", "GE[C-1]，e的超平面的固定子集，可以有效，准确", "建了一个理论上合理的算法来攻击二进制线性分类器", "地近似（??）。g。，G=4比～14×（对于CIFAR-100）", "的随机集合", "减少了更多的评估时间，而不会影响准确性。我们将", "使用此版本的ARC来评估此类数据集，并请读者参", "5.2. Differentiable Multiclass Classifiers", "考附录C以获取更多详细信息。", "我们现在将算法E1中提出的ARC算法扩展到C元", "可微分分类器的更一般情况，例如DNN。在处理非线", "6. Experiments", "性可微分分类器f时，我们可以“linearize”它在输人", "6.1. Setup", "周围的行为，以便提供到决策边界（的最短e，距离", "以及相应单元l范数方向g的估计。在a上使用-", "我们进行了全面的实验，以证明与APGD相比，ARC", "阶泰勒级数展开，我们近似f：", "在为REC生成范数有界对抗性扰动方面的有效性。具", "体来说，我们尝试了各种网络架构（VGG-16（Simonyan", "& Zisserman, 2014) , ResNet-20, ResNet-18 (He et al.,", "2016) , WideResNet-28-4 (Zagoruyko & Komodakis,", "f(u)~ f(u) = f(α)+f(α)T(u-a)", "(10)", "2016）和MobileNetV1（Howardetal.，2017）），数据集", "(SVHN (Netzer et al.,2011),CIFAR-10 (Krizhevsky", "etal.，2009）和THMAIFAR-。我们使用来自（Rice", "其中 Vf(α) =[V[f(α)]1 ].../[f(α)]c] e RDxc 是", "etal.，2020）的经过试验和测试的PGD公开实现。对", "雅可比矩阵。假设mE[Cl是f分配给&的标签，我", "于我们所有的实验，在训练和评估过程中使用相同的", "们可以在周围构造近似决策区域Rm(α），如下所", "e。范数。有关培训/评估设置的更多详细信息，请参见", "示：", "附录??" ]

{ "arxiv_id": "2206.06737", "language": "zh", "page_idx": 4 }

[ [ "0.3123", "0.0986", "0.693", "0.0986", "0.693", "0.1134", "0.3123", "0.1134" ], [ "0.1301", "0.1312", "0.876", "0.1312", "0.876", "0.1455", "0.1301", "0.1455" ], [ "0.2957", "0.1455", "0.7153", "0.1467", "0.7153", "0.1609", "0.2957", "0.1603" ], [ "0.1914", "0.1592", "0.818", "0.1609", "0.818", "0.1758", "0.1914", "0.1733" ], [ "0.2778", "0.1746", "0.729", "0.1746", "0.729", "0.1882", "0.2778", "0.1882" ], [ "0.3418", "0.1907", "0.6665", "0.1907", "0.6665", "0.2037", "0.3418", "0.2037" ], [ "0.465", "0.2203", "0.5366", "0.2227", "0.5366", "0.2369", "0.465", "0.2351" ], [ "0.12177", "0.2393", "0.881", "0.2393", "0.881", "0.2494", "0.12177", "0.2494" ], [ "0.12006", "0.2505", "0.882", "0.2505", "0.882", "0.2642", "0.12006", "0.2642" ], [ "0.12006", "0.2642", "0.8833", "0.2642", "0.8833", "0.2778", "0.12006", "0.2778" ], [ "0.12006", "0.2756", "0.882", "0.2756", "0.882", "0.2893", "0.12006", "0.2893" ], [ "0.12006", "0.288", "0.882", "0.288", "0.882", "0.3018", "0.12006", "0.3018" ], [ "0.12006", "0.3027", "0.882", "0.3027", "0.882", "0.3164", "0.12006", "0.3164" ], [ "0.1184", "0.3164", "0.8833", "0.3164", "0.8833", "0.33", "0.1184", "0.33" ], [ "0.12177", "0.3308", "0.881", "0.3308", "0.881", "0.3416", "0.12177", "0.3416" ], [ "0.12006", "0.3425", "0.881", "0.3425", "0.881", "0.3528", "0.12006", "0.3528" ], [ "0.12006", "0.357", "0.3586", "0.357", "0.3586", "0.3677", "0.12006", "0.3677" ], [ "0.12006", "0.3813", "0.7036", "0.3813", "0.7036", "0.3943", "0.12006", "0.3943" ], [ "0.5156", "0.4092", "0.881", "0.4092", "0.881", "0.4192", "0.5156", "0.4192" ], [ "0.5146", "0.4216", "0.882", "0.4216", "0.882", "0.4353", "0.5146", "0.4353" ], [ "0.2183", "0.437", "0.382", "0.4382", "0.382", "0.4531", "0.2183", "0.452" ], [ "0.517", "0.437", "0.882", "0.437", "0.882", "0.4478", "0.517", "0.4478" ], [ "0.517", "0.452", "0.882", "0.452", "0.882", "0.462", "0.517", "0.462" ], [ "0.12177", "0.4697", "0.4878", "0.4697", "0.4878", "0.4797", "0.12177", "0.4797" ], [ "0.517", "0.4644", "0.882", "0.4644", "0.882", "0.478", "0.517", "0.478" ], [ "0.1184", "0.4822", "0.4895", "0.4822", "0.4895", "0.4954", "0.1184", "0.4954" ], [ "0.5156", "0.4797", "0.881", "0.4797", "0.881", "0.4905", "0.5156", "0.4905" ], [ "0.1184", "0.4963", "0.4878", "0.4963", "0.4878", "0.5103", "0.1184", "0.5103" ], [ "0.513", "0.4922", "0.8833", "0.4912", "0.8833", "0.506", "0.513", "0.5073" ], [ "0.5146", "0.506", "0.8833", "0.506", "0.8833", "0.5195", "0.5146", "0.5195" ], [ "0.12177", "0.5127", "0.486", "0.5127", "0.486", "0.5225", "0.12177", "0.5225" ], [ "0.12006", "0.525", "0.4878", "0.525", "0.4878", "0.538", "0.12006", "0.538" ], [ "0.5146", "0.5195", "0.659", "0.521", "0.659", "0.5337", "0.5146", "0.533" ], [ "0.1335", "0.539", "0.4895", "0.539", "0.4895", "0.5527", "0.1335", "0.5527" ], [ "0.531", "0.5376", "0.882", "0.538", "0.882", "0.552", "0.531", "0.5503" ], [ "0.12006", "0.5527", "0.4895", "0.5527", "0.4895", "0.5664", "0.12006", "0.5664" ], [ "0.511", "0.5503", "0.882", "0.5503", "0.882", "0.564", "0.511", "0.564" ], [ "0.12177", "0.5684", "0.4878", "0.5684", "0.4878", "0.579", "0.12177", "0.579" ], [ "0.5146", "0.5664", "0.882", "0.5664", "0.882", "0.58", "0.5146", "0.58" ], [ "0.1184", "0.58", "0.4878", "0.58", "0.4878", "0.5933", "0.1184", "0.5933" ], [ "0.513", "0.58", "0.882", "0.579", "0.882", "0.5933", "0.513", "0.594" ], [ "0.12177", "0.5957", "0.4895", "0.5957", "0.4895", "0.6094", "0.12177", "0.6094" ], [ "0.513", "0.594", "0.882", "0.594", "0.882", "0.608", "0.513", "0.608" ], [ "0.1184", "0.608", "0.4878", "0.6094", "0.4878", "0.6235", "0.1184", "0.6226" ], [ "0.5156", "0.6094", "0.882", "0.6094", "0.882", "0.6226", "0.5156", "0.6226" ], [ "0.1184", "0.6235", "0.4902", "0.6235", "0.4902", "0.637", "0.1184", "0.637" ], [ "0.5146", "0.6235", "0.882", "0.6235", "0.882", "0.637", "0.5146", "0.637" ], [ "0.12006", "0.6396", "0.1628", "0.6396", "0.1628", "0.651", "0.12006", "0.651" ], [ "0.517", "0.6396", "0.881", "0.6396", "0.881", "0.6494", "0.517", "0.6494" ], [ "0.1368", "0.652", "0.4895", "0.652", "0.4895", "0.665", "0.1368", "0.665" ], [ "0.5156", "0.653", "0.882", "0.653", "0.882", "0.663", "0.5156", "0.663" ], [ "0.12177", "0.6685", "0.4878", "0.6685", "0.4878", "0.6787", "0.12177", "0.6787" ], [ "0.513", "0.6665", "0.879", "0.6665", "0.879", "0.677", "0.513", "0.677" ], [ "0.1184", "0.681", "0.4878", "0.681", "0.4878", "0.695", "0.1184", "0.695" ], [ "0.513", "0.6797", "0.8833", "0.6797", "0.8833", "0.6934", "0.513", "0.6934" ], [ "0.12006", "0.6953", "0.4895", "0.6953", "0.4895", "0.709", "0.12006", "0.709" ], [ "0.5156", "0.6953", "0.882", "0.6953", "0.882", "0.706", "0.5156", "0.706" ], [ "0.12006", "0.709", "0.4895", "0.709", "0.4895", "0.7227", "0.12006", "0.7227" ], [ "0.5146", "0.709", "0.881", "0.709", "0.881", "0.7197", "0.5146", "0.7197" ], [ "0.12177", "0.7246", "0.486", "0.7246", "0.486", "0.7354", "0.12177", "0.7354" ], [ "0.513", "0.7217", "0.784", "0.7217", "0.784", "0.7354", "0.513", "0.7354" ], [ "0.12177", "0.7393", "0.4844", "0.7393", "0.4844", "0.75", "0.12177", "0.75" ], [ "0.532", "0.741", "0.881", "0.741", "0.881", "0.752", "0.532", "0.752" ], [ "0.12006", "0.7505", "0.4878", "0.7505", "0.4878", "0.764", "0.12006", "0.764" ], [ "0.513", "0.7544", "0.882", "0.7544", "0.882", "0.767", "0.513", "0.767" ], [ "0.12006", "0.7666", "0.4878", "0.7666", "0.4878", "0.78", "0.12006", "0.78" ], [ "0.5146", "0.7686", "0.8774", "0.7686", "0.8774", "0.7793", "0.5146", "0.7793" ], [ "0.12006", "0.782", "0.2544", "0.782", "0.2544", "0.792", "0.12006", "0.792" ], [ "0.5146", "0.782", "0.882", "0.783", "0.882", "0.7964", "0.5146", "0.796" ], [ "0.1351", "0.7944", "0.4878", "0.7935", "0.4878", "0.807", "0.1351", "0.808" ], [ "0.513", "0.796", "0.881", "0.7964", "0.881", "0.81", "0.513", "0.8086" ], [ "0.12006", "0.8086", "0.4895", "0.8086", "0.4895", "0.8223", "0.12006", "0.8223" ], [ "0.511", "0.8086", "0.882", "0.81", "0.882", "0.8237", "0.511", "0.8223" ], [ "0.12177", "0.8247", "0.4878", "0.8247", "0.4878", "0.835", "0.12177", "0.835" ], [ "0.513", "0.8237", "0.881", "0.8247", "0.881", "0.8384", "0.513", "0.8374" ], [ "0.5146", "0.8374", "0.882", "0.8384", "0.882", "0.8525", "0.5146", "0.8516" ], [ "0.1453", "0.8486", "0.4902", "0.8486", "0.4902", "0.8623", "0.1453", "0.8623" ], [ "0.5156", "0.8525", "0.882", "0.8525", "0.882", "0.866", "0.5156", "0.866" ], [ "0.1184", "0.861", "0.4895", "0.861", "0.4895", "0.8745", "0.1184", "0.8745" ], [ "0.5156", "0.8677", "0.882", "0.8677", "0.882", "0.8813", "0.5156", "0.8813" ], [ "0.1184", "0.8745", "0.4895", "0.8745", "0.4895", "0.888", "0.1184", "0.888" ], [ "0.5156", "0.882", "0.882", "0.882", "0.882", "0.8955", "0.5156", "0.8955" ], [ "0.12006", "0.8896", "0.4878", "0.8896", "0.4878", "0.8994", "0.12006", "0.8994" ], [ "0.513", "0.8955", "0.882", "0.8945", "0.882", "0.909", "0.513", "0.91" ], [ "0.12006", "0.902", "0.2301", "0.902", "0.2301", "0.913", "0.12006", "0.913" ] ]

[ "使用语音活动投影的多语言转弯预测", "Koji Inouel,Bing'erJiang?,ErikEkstedt?,TatsuyaKawahara',GabrielSkantze", "1GraduateSchool of Informatics,KyotoUniversity,Japan", "2Division of Speech, Music and Hearing, KTHRoyal Institute of Technology, Sweden", "[email protected], [email protected], [email protected]", "[email protected],[email protected]", "Abstract", "Thispaperinvestigatestheapplicationofvoiceactivityprojection(VAP)apredictiveturn-takingmodelforspoken", "dialogue, on multilingual data, encompassing English, Mandarin, and Japanese. The VAP model continuously", "predicts theupcoming voice activities of participants in dyadic dialogue, leveraging a cross-attention Transformer to", "capture the dynamic interplay between participants.The results show that a monolingual VAP model trained on one", "language does not make good predictions when applied to other languages. However, a multilingual model, trained", "on all three languages,demonstrates predictive performanceon par withmonolingual models across all languages.", "Further analyses show that the multilingual model has learned to discern the language of the input signal. We also", "analyzethe sensitivitytopitch,aprosodic cuethatis thoughtto beimportantforturn-taking.Finally,wecompare", "twodifferentaudioencoders,contrastivepredictivecoding(CPC)pre-trainedonEnglish,witharecentmodelbased", "onmultilingualwav2vec2.0(MMS).", "Keywords: Turn-taking, Multilingual, Spoken Dialogue System, Voice Activity Projection", "用户是否只是暂停（保持）：或者转弯是否产生（移", "位)。。在早期的模型中，特征工程很常见，但现在", "1.Introduction", "将时间序列数据（如韵律特征和词向量表示（词嵌", "入））输入到神经网络（如递归神经网络（RNNs））", "转弯是人类之间口语互动的基本方面，因此是口", "中变得越来越流行。。最近，已经提出了基于变压", "语对话系统中建模的重要功能E(Skantze，2021）。", "器的模型，该模型可以以端到端处理方式接收原", "在人与人之间的对话中，对话地板的过渡顺利进", "始输人文本或音频E(Ekstedt and Skantze,2020;", "Sakuma et al., 2023; Muromachi and Kano, 2023;", "行。在各种语言中已经表明，过渡偏移通常非常", "短暂，大约为100-500毫秒（Stiversetal.，2009)", "Kurata et al., 2023)", "这表明人类在多种模式下使用各种转弯信号，包", "早期模型的另一个局限性是它们仅关注转向保", "括词汇提示，韵律，凝视，呼吸和手势，以协调E", "持与。shift。一个更全面的转弯模型应该涉及更细", "(WtodarczakandHeldner,2016;Kendricketal.,", "微的决定。转弯时，例如，有必要在开始讲话之前", "2023）。此外，鉴于听众也需要一些时间来表达回", "确定适当的等待时间E(RauxandEskenazi,2012;", "应，因此可能涉及一种预测机制，其中听众预测", "Lala et al.,2018;Sakuma etal.,2023)。在预测反", "说话者的话语即将结束E(Garrod and Pickering,", "向通道与。轮班（Lalaetal.，2017）。如上所述，人", "2015; Levinson and Torreira, 2015; Ishimoto et al.,", "类不仅能够对转弯产生的线索做出反应，而且还可", "2017)", "以预测即将到来的转弯。这显然也是口语对话系统", "虽然大型语言模型（LLM）的最新进展使在口语", "的理想属性。此外，没有建立和强大的方法来处理", "对话系统中生成高度复杂的响应变得更加容易，但", "对话中的中断和重叠，这在人与人之间的对话中很", "转弯仍然通常以非常简单的方式处理。在实际的口", "常见。需要更动态的转弯预测模型，以使口语对话", "语对话系统中，通常使用简单的静音超时阈值（通", "系统能够以更人性化的方式处理转弯。对于此类模", "常约为1秒）来实现转弯，以指示转弯的结束。然", "型而言，至关重要的是，它们不会在特定事件中做", "而，沉默并不是一个很好的指标，因为在人与人之", "出轮流决策，而是以连续的方式运行", "间的互动中，转弯内的沉默（停顿）通常比转弯之", "最近提出了几种模型，它们以时间范围的方式连", "间的沉默更长（HeldnerandEdlund，2010）。这意", "续进行更细微的转弯预测E(Skantze，2017;Lala", "味着口语对话系统经常受到长时间响应延迟或停", "etal.2019）。在这样的连续模型中，本研究采用了", "顿频繁中断的困扰", "语音活动投影（VAP）模型（EkstedtandSkantze，", "为了解决这个问题，已经为转弯终点预测模型提", "2022b）。VAP模型使用多层变形金刚，并通过处", "出了许多建议。这些模型考虑了先前用户话语的言", "理来自二元对话中两个说话者的原始音频信号来", "语和非言语提示（例如语言和韵律特征），：以预测", "预测对话参与者的近期语音活动。先前的工作表", "明，VAP模型在预测转弯行为方面优于其他模型，", "This paper has been accepted for presentation at", "包括反向通道预测（EkstedtandSkantze，2022b)", "The 2024 Joint International Conference on Computa-", "。在此模型的最新版本中，在自我关注层之后添加", "tional Linguistics, Language Resources and Evaluation", "了交叉关注转换器层，以分别对来自两个扬声器", "(LREC-COLING2024)andrepresentstheauthorsver-", "的音频进行建模，如第E2节中所述。最近，VAP", "sion of the work." ]

{ "arxiv_id": "2403.06487", "language": "zh", "page_idx": 0 }

[ [ "0.1301", "0.08075", "0.1561", "0.08075", "0.1561", "0.08905", "0.1301", "0.08905" ], [ "0.1301", "0.09143", "0.1545", "0.0932", "0.1528", "0.10156", "0.1293", "0.0998" ], [ "0.6533", "0.0885", "0.6665", "0.0885", "0.6665", "0.09265", "0.6533", "0.09265" ], [ "0.1184", "0.10095", "0.157", "0.10095", "0.157", "0.1164", "0.1184", "0.1164" ], [ "0.636", "0.0986", "0.6724", "0.10095", "0.669", "0.12354", "0.634", "0.1217" ], [ "0.1335", "0.11755", "0.1561", "0.11755", "0.1561", "0.1277", "0.1335", "0.1277" ], [ "0.65", "0.12354", "0.669", "0.12354", "0.669", "0.1318", "0.65", "0.1318" ], [ "0.1377", "0.13", "0.1561", "0.13", "0.1561", "0.1383", "0.1377", "0.1383" ], [ "0.22", "0.1497", "0.3047", "0.152", "0.3047", "0.1663", "0.22", "0.1644" ], [ "0.4836", "0.1526", "0.5557", "0.1526", "0.5557", "0.1644", "0.4836", "0.1644" ], [ "0.733", "0.1484", "0.8213", "0.1509", "0.8213", "0.1687", "0.733", "0.1663" ], [ "0.7734", "0.1561", "0.7925", "0.1561", "0.7925", "0.1603", "0.7734", "0.1603" ], [ "0.4248", "0.1704", "0.574", "0.1704", "0.574", "0.1812", "0.4248", "0.1812" ], [ "0.6616", "0.1704", "0.858", "0.1704", "0.858", "0.1841", "0.6616", "0.1841" ], [ "0.294", "0.1953", "0.7095", "0.1965", "0.7095", "0.2114", "0.294", "0.2102" ], [ "0.157", "0.1729", "0.3267", "0.1729", "0.3267", "0.1829", "0.157", "0.1829" ], [ "0.3972", "0.2351", "0.4106", "0.2351", "0.4106", "0.2393", "0.3972", "0.2393" ], [ "0.3838", "0.253", "0.3955", "0.257", "0.3855", "0.2725", "0.3728", "0.2683" ], [ "0.3972", "0.2559", "0.4106", "0.2559", "0.4106", "0.2588", "0.3972", "0.2588" ], [ "0.1351", "0.2737", "0.1486", "0.2737", "0.1486", "0.2778", "0.1351", "0.2778" ], [ "0.3972", "0.2756", "0.4106", "0.2756", "0.4106", "0.279", "0.3972", "0.279" ], [ "0.6533", "0.2737", "0.6665", "0.2737", "0.6665", "0.2766", "0.6533", "0.2766" ], [ "0.75", "0.2893", "0.812", "0.2915", "0.812", "0.3018", "0.75", "0.2998" ], [ "0.215", "0.304", "0.3032", "0.308", "0.3015", "0.3196", "0.2141", "0.3154" ], [ "0.6826", "0.304", "0.692", "0.304", "0.692", "0.3088", "0.6826", "0.3088" ], [ "0.7246", "0.3027", "0.818", "0.3047", "0.818", "0.3196", "0.7246", "0.3176" ], [ "0.7188", "0.3047", "0.7256", "0.3047", "0.7256", "0.31", "0.7188", "0.31" ], [ "0.4785", "0.3047", "0.561", "0.306", "0.561", "0.3176", "0.4785", "0.3164" ], [ "0.1561", "0.3225", "0.3284", "0.3225", "0.3284", "0.3362", "0.1561", "0.3362" ], [ "0.4224", "0.3218", "0.5767", "0.3218", "0.5767", "0.335", "0.4224", "0.335" ], [ "0.663", "0.3247", "0.8564", "0.3247", "0.8564", "0.335", "0.663", "0.335" ], [ "0.293", "0.3486", "0.7095", "0.3508", "0.7095", "0.364", "0.293", "0.3623" ], [ "0.12006", "0.3896", "0.4878", "0.3896", "0.4878", "0.4026", "0.12006", "0.4026" ], [ "0.529", "0.3896", "0.882", "0.3896", "0.882", "0.4026", "0.529", "0.4026" ], [ "0.12006", "0.4038", "0.4895", "0.4038", "0.4895", "0.4175", "0.12006", "0.4175" ], [ "0.511", "0.4038", "0.882", "0.4038", "0.882", "0.4175", "0.511", "0.4175" ], [ "0.1184", "0.4175", "0.4895", "0.4175", "0.4895", "0.4312", "0.1184", "0.4312" ], [ "0.5146", "0.4175", "0.784", "0.4175", "0.784", "0.4312", "0.5146", "0.4312" ], [ "0.12006", "0.4316", "0.4878", "0.4316", "0.4878", "0.4453", "0.12006", "0.4453" ], [ "0.12006", "0.4478", "0.486", "0.4478", "0.486", "0.4578", "0.12006", "0.4578" ], [ "0.613", "0.4531", "0.659", "0.4531", "0.659", "0.4673", "0.613", "0.4673" ], [ "0.6533", "0.4536", "0.7773", "0.4536", "0.7773", "0.4673", "0.6533", "0.4673" ], [ "0.1184", "0.4602", "0.4358", "0.4602", "0.4358", "0.4739", "0.1184", "0.4739" ], [ "0.5156", "0.4827", "0.8833", "0.4827", "0.8833", "0.4963", "0.5156", "0.4963" ], [ "0.12006", "0.4976", "0.1545", "0.4976", "0.1545", "0.5103", "0.12006", "0.5103" ], [ "0.1704", "0.4954", "0.44", "0.4976", "0.44", "0.5127", "0.1704", "0.5103" ], [ "0.1587", "0.502", "0.1663", "0.502", "0.1663", "0.505", "0.1587", "0.505" ], [ "0.5146", "0.4976", "0.802", "0.4976", "0.802", "0.511", "0.5146", "0.511" ], [ "0.168", "0.5127", "0.266", "0.5127", "0.266", "0.527", "0.168", "0.527" ], [ "0.5146", "0.532", "0.574", "0.532", "0.574", "0.5435", "0.5146", "0.5435" ], [ "0.567", "0.532", "0.8076", "0.532", "0.8076", "0.546", "0.567", "0.546" ], [ "0.12006", "0.549", "0.4895", "0.55", "0.4895", "0.564", "0.12006", "0.5635" ], [ "0.12006", "0.536", "0.4895", "0.536", "0.4895", "0.55", "0.12006", "0.55" ], [ "0.5146", "0.5547", "0.8833", "0.5547", "0.8833", "0.5684", "0.5146", "0.5684" ], [ "0.12006", "0.5654", "0.4844", "0.5654", "0.4844", "0.576", "0.12006", "0.576" ], [ "0.5156", "0.5693", "0.8345", "0.5693", "0.8345", "0.583", "0.5156", "0.583" ], [ "0.12177", "0.5776", "0.486", "0.5776", "0.486", "0.5913", "0.12177", "0.5913" ], [ "0.12177", "0.5933", "0.4878", "0.5933", "0.4878", "0.607", "0.12177", "0.607" ], [ "0.5146", "0.604", "0.58", "0.604", "0.58", "0.6143", "0.5146", "0.6143" ], [ "0.574", "0.6016", "0.657", "0.6016", "0.657", "0.6133", "0.574", "0.6133" ], [ "0.12006", "0.608", "0.4844", "0.608", "0.4844", "0.618", "0.12006", "0.618" ], [ "0.12006", "0.6206", "0.4878", "0.6206", "0.4878", "0.6343", "0.12006", "0.6343" ], [ "0.5156", "0.626", "0.8833", "0.626", "0.8833", "0.6396", "0.5156", "0.6396" ], [ "0.12006", "0.6357", "0.4902", "0.6357", "0.4902", "0.6494", "0.12006", "0.6494" ], [ "0.5146", "0.6396", "0.8833", "0.6396", "0.8833", "0.653", "0.5146", "0.653" ], [ "0.1234", "0.652", "0.4878", "0.652", "0.4878", "0.662", "0.1234", "0.662" ], [ "0.5156", "0.654", "0.8833", "0.654", "0.8833", "0.6675", "0.5156", "0.6675" ], [ "0.12006", "0.6646", "0.3047", "0.6646", "0.3047", "0.6777", "0.12006", "0.6777" ], [ "0.5156", "0.6685", "0.55", "0.6685", "0.55", "0.681", "0.5156", "0.681" ], [ "0.531", "0.683", "0.8833", "0.683", "0.8833", "0.697", "0.531", "0.697" ], [ "0.1696", "0.6987", "0.4475", "0.7007", "0.4475", "0.7153", "0.1696", "0.7134" ], [ "0.5146", "0.6953", "0.8833", "0.6978", "0.8833", "0.7114", "0.5146", "0.709" ], [ "0.12006", "0.7017", "0.1755", "0.7017", "0.1755", "0.712", "0.12006", "0.712" ], [ "0.5146", "0.7104", "0.881", "0.7104", "0.881", "0.7236", "0.5146", "0.7236" ], [ "0.12177", "0.727", "0.4878", "0.727", "0.4878", "0.737", "0.12177", "0.737" ], [ "0.513", "0.7256", "0.8833", "0.7256", "0.8833", "0.7393", "0.513", "0.7393" ], [ "0.12006", "0.7393", "0.4878", "0.7393", "0.4878", "0.753", "0.12006", "0.753" ], [ "0.5146", "0.741", "0.7827", "0.741", "0.7827", "0.752", "0.5146", "0.752" ], [ "0.12006", "0.7544", "0.4878", "0.7544", "0.4878", "0.767", "0.12006", "0.767" ], [ "0.12177", "0.7695", "0.486", "0.7695", "0.486", "0.78", "0.12177", "0.78" ], [ "0.513", "0.7725", "0.764", "0.7725", "0.764", "0.786", "0.513", "0.786" ], [ "0.12177", "0.783", "0.4844", "0.783", "0.4844", "0.7935", "0.12177", "0.7935" ], [ "0.12177", "0.7964", "0.4895", "0.7964", "0.4895", "0.81", "0.12177", "0.81" ], [ "0.5146", "0.7964", "0.8833", "0.7964", "0.8833", "0.81", "0.5146", "0.81" ], [ "0.12006", "0.81", "0.4895", "0.81", "0.4895", "0.8237", "0.12006", "0.8237" ], [ "0.5146", "0.81", "0.881", "0.81", "0.881", "0.8247", "0.5146", "0.8247" ], [ "0.12006", "0.8247", "0.4902", "0.8247", "0.4902", "0.8384", "0.12006", "0.8384" ], [ "0.5156", "0.826", "0.881", "0.826", "0.881", "0.836", "0.5156", "0.836" ], [ "0.12006", "0.8384", "0.4895", "0.8384", "0.4895", "0.8516", "0.12006", "0.8516" ], [ "0.513", "0.8384", "0.882", "0.8384", "0.882", "0.8516", "0.513", "0.8516" ], [ "0.12006", "0.8525", "0.4895", "0.8525", "0.4895", "0.866", "0.12006", "0.866" ], [ "0.5146", "0.8525", "0.8833", "0.8525", "0.8833", "0.866", "0.5146", "0.866" ], [ "0.12177", "0.8677", "0.4878", "0.8677", "0.4878", "0.8813", "0.12177", "0.8813" ], [ "0.5146", "0.8677", "0.882", "0.8677", "0.882", "0.8813", "0.5146", "0.8813" ], [ "0.1234", "0.883", "0.486", "0.883", "0.486", "0.8936", "0.1234", "0.8936" ], [ "0.5146", "0.883", "0.879", "0.883", "0.879", "0.8936", "0.5146", "0.8936" ], [ "0.12006", "0.8955", "0.482", "0.8955", "0.482", "0.909", "0.12006", "0.909" ], [ "0.5146", "0.898", "0.881", "0.898", "0.881", "0.9077", "0.5146", "0.9077" ] ]

[ "25000", "20000", "000", "15000", "14000", "10000", "2000", "5000", "Tum-shit gep[sec]", "Turm-shift gap [se]", "Truma-hit gap[se]", "", "(b)Mandarin (HKUST)", "(c) Japanese (Travel agency)", "Figure 3: Histogram of turn-shift gap in three languages", "(a)English(Switchboard)", "000", "aunoo", "o", "no0", "00", "", "", "Turn-hold gap [sec]", "0", "Turn-hold gap [secl", "1", "Turn-hold gap [sec]", "(a) English (Switchboard)", "(b) Mandarin (HKUST)", "(c)Japanese(Travelagency)", "Figure 4:Histogram of turn-holdgap in three languages", "148个。持续时间6小时。这些已分为758个对话", "鉴于这些差异，重要的是要注意，VAP模型不仅", "的培训(大约130个。1小时)，88个对话的验证", "要考虑语言之间的提示和信号如何不同，还要考虑", "(14。5小时)，以及24次对话的测试(3。9小时)，", "整体分布，这可能会使预测产生偏差", "分别。与Switchboard类似，为了匹配最小数据集", "（即日本数据集）的天小，训练集和验证集已被二", "4.", "Experiments", "次采样为92。5小时和11。分别为5小时", "为了回答E1节中提到的从RQ1到RQ5的研究问", "3.3.", "Travel Agency Task Dialogues", "", "题，我们进行了一系列实验，如下所述", "(Japanese)", "4.1.", "Cross-lingual Performance", "用于日语E(Inabaetal.,2022)的旅行咨询。这些", "旅行社任务对话是一个收集模拟对话数据的项目，", "为了回答RQ1和RQ2，我们比较了多语言模型和", "对话是使用在线会议系统录制的。对话模拟旅行社", "专门为每种语言训练的单语模型之间的性能", "工作人员与客户之间的在线对话，工作人员的角色", "由在旅行社工作的具有实际经验的人扮演。请注", "4.1.1.", "Condition", "意虽然参与者被分配了角色，但他们没有得到脚", "本。总共329个对话（115。5小时)已记录。这些", "多语言模型是在上述三种语言的所有数据上训练", "对话被随机分为会话单元，有263个对话（92。5", "的。此外，还为每种语言分别训练了单语模型。因", "小时）在训练集中，以及33个对话（11。5小时）验", "此，多语言模型的训练数据量是每个单语言模型的", "证集和测试集中的每一个", "三倍", "对于多语言和单语言设置，VAP模型的结构和", "DifferencesAcrossLanguages", "参数相同。训练参数如下：训练时期数为20，批", "3.4.", "量为8，学习率为3。63E-4，并且重量衰减设置为", "如上所述，早期的研究表明，三种语言之间的轮流", "0。001。我们使用了AdamW优化器。我们使用验", "倾向略有不同。因此，我们首先研究了数据集之间", "证集上损失最小的模型来评估测试集", "的间隙和暂停长度的差异。图3显示了转弯过渡", "期间间隙持续时间的直方图，而图4显示了转弯", "4.1.2.TestLoss Performance", "保持期间暂停长度的分布。可以观察到，与英语相", "比，普通话和日语在转弯过渡期间的差距往往略", "作为基本评估指标，我们评估了测试集上的平均损", "短。另一方面，与日语相比，在轮流举行期间，英", "失。评估的兴趣损失是Luap，它在2节中定义。", "语和普通话显示出更短的停顿长度的趋势。此外，", "表1显示了结果。可以看出，虽然单语模型在同", "值得观察的是，日本人的分布更加均匀。日语数据", "一语言上测试时效果很好，但在应用于另一种语言", "集展示了具有明确分配角色的正式对话设置，将其", "时，它们的表现要差得多。从这个结果可以清楚地", "与其他两个数据集区分开来。因此，在日语数据集", "看出，语音活动投影的性质在本研究中使用的三种", "中似乎有一种趋势，即参与者在没有轮班的情况下", "语言数据集中是不同的，为了做出准确的预测，有", "保持更长的时间，从而导致话语之间的差距更长。", "必要为每种语言训练特定的模型。然而，多语言模" ]

{ "arxiv_id": "2403.06487", "language": "zh", "page_idx": 3 }

[ [ "0.3435", "0.07184", "0.4148", "0.0736", "0.4148", "0.0885", "0.3435", "0.0861" ], [ "0.739", "0.07184", "0.81", "0.0736", "0.81", "0.0885", "0.739", "0.0861" ], [ "0.5425", "0.08014", "0.642", "0.082", "0.642", "0.0968", "0.5425", "0.0944" ], [ "0.3552", "0.0873", "0.4014", "0.0873", "0.4014", "0.10156", "0.3552", "0.10156" ], [ "0.7515", "0.0873", "0.796", "0.0873", "0.796", "0.10156", "0.7515", "0.10156" ], [ "0.4282", "0.0885", "0.4685", "0.0885", "0.4685", "0.0998", "0.4282", "0.0998" ], [ "0.288", "0.0885", "0.33", "0.0885", "0.33", "0.10156", "0.288", "0.10156" ], [ "0.6826", "0.0885", "0.7246", "0.0885", "0.7246", "0.10156", "0.6826", "0.10156" ], [ "0.825", "0.0885", "0.864", "0.0885", "0.864", "0.10095", "0.825", "0.10095" ], [ "0.2837", "0.10156", "0.33", "0.10156", "0.33", "0.1164", "0.2837", "0.1164" ], [ "0.1503", "0.08905", "0.2468", "0.08905", "0.2468", "0.10095", "0.1503", "0.10095" ], [ "0.1351", "0.1027", "0.1957", "0.1027", "0.1957", "0.11755", "0.1351", "0.11755" ], [ "0.529", "0.1027", "0.5903", "0.1027", "0.5903", "0.11755", "0.529", "0.11755" ], [ "0.681", "0.10156", "0.727", "0.10156", "0.727", "0.1164", "0.681", "0.1164" ], [ "0.8213", "0.1027", "0.8696", "0.1027", "0.8696", "0.11755", "0.8213", "0.11755" ], [ "0.4265", "0.10394", "0.471", "0.10394", "0.471", "0.1152", "0.4265", "0.1152" ], [ "0.3552", "0.10394", "0.3997", "0.10394", "0.3997", "0.1152", "0.3552", "0.1152" ], [ "0.7515", "0.10394", "0.796", "0.10394", "0.796", "0.1152", "0.7515", "0.1152" ], [ "0.1335", "0.1152", "0.2091", "0.11755", "0.2091", "0.1318", "0.1335", "0.13" ], [ "0.2837", "0.1164", "0.3315", "0.1164", "0.3315", "0.1312", "0.2837", "0.1312" ], [ "0.3535", "0.1164", "0.4014", "0.1164", "0.4014", "0.1312", "0.3535", "0.1312" ], [ "0.4248", "0.1164", "0.471", "0.1164", "0.471", "0.1312", "0.4248", "0.1312" ], [ "0.529", "0.1152", "0.6035", "0.11755", "0.6035", "0.1318", "0.529", "0.13" ], [ "0.679", "0.1164", "0.7256", "0.1164", "0.7256", "0.1312", "0.679", "0.1312" ], [ "0.75", "0.1164", "0.797", "0.1164", "0.797", "0.1312", "0.75", "0.1312" ], [ "0.8213", "0.1164", "0.867", "0.1164", "0.867", "0.1312", "0.8213", "0.1312" ], [ "0.2837", "0.1312", "0.3315", "0.1312", "0.3315", "0.1455", "0.2837", "0.1455" ], [ "0.3535", "0.1312", "0.403", "0.1312", "0.403", "0.1455", "0.3535", "0.1455" ], [ "0.424", "0.1312", "0.471", "0.1312", "0.471", "0.1455", "0.424", "0.1455" ], [ "0.679", "0.1312", "0.727", "0.1312", "0.727", "0.1455", "0.679", "0.1455" ], [ "0.75", "0.1312", "0.797", "0.1312", "0.797", "0.1455", "0.75", "0.1455" ], [ "0.8193", "0.1312", "0.869", "0.1312", "0.869", "0.1455", "0.8193", "0.1455" ], [ "0.532", "0.133", "0.607", "0.133", "0.607", "0.1443", "0.532", "0.1443" ], [ "0.1368", "0.133", "0.2107", "0.133", "0.2107", "0.1443", "0.1368", "0.1443" ], [ "0.1335", "0.1443", "0.2603", "0.1455", "0.2603", "0.1603", "0.1335", "0.1592" ], [ "0.3535", "0.1455", "0.403", "0.1455", "0.403", "0.1603", "0.3535", "0.1603" ], [ "0.424", "0.1455", "0.4727", "0.1455", "0.4727", "0.1603", "0.424", "0.1603" ], [ "0.5283", "0.1443", "0.656", "0.1455", "0.656", "0.1603", "0.5283", "0.1592" ], [ "0.678", "0.1455", "0.729", "0.1455", "0.729", "0.1603", "0.678", "0.1603" ], [ "0.8193", "0.1455", "0.8696", "0.1455", "0.8696", "0.1603", "0.8193", "0.1603" ], [ "0.7515", "0.1467", "0.796", "0.1467", "0.796", "0.158", "0.7515", "0.158" ], [ "0.2854", "0.1467", "0.33", "0.1467", "0.33", "0.158", "0.2854", "0.158" ], [ "0.126", "0.1704", "0.4802", "0.1716", "0.4802", "0.1859", "0.126", "0.1853" ], [ "0.5156", "0.1729", "0.592", "0.1729", "0.592", "0.1841", "0.5156", "0.1841" ], [ "0.5845", "0.1729", "0.8833", "0.1729", "0.8833", "0.1859", "0.5845", "0.1859" ], [ "0.5146", "0.187", "0.8037", "0.187", "0.8037", "0.2007", "0.5146", "0.2007" ], [ "0.1545", "0.2007", "0.2167", "0.2007", "0.2167", "0.2156", "0.1545", "0.2156" ], [ "0.2468", "0.2007", "0.3032", "0.2007", "0.3032", "0.2156", "0.2468", "0.2156" ], [ "0.3225", "0.2007", "0.3787", "0.2007", "0.3787", "0.2156", "0.3225", "0.2156" ], [ "0.4014", "0.2007", "0.4634", "0.2007", "0.4634", "0.2156", "0.4014", "0.2156" ], [ "0.2603", "0.2156", "0.3015", "0.2156", "0.3015", "0.231", "0.2603", "0.231" ], [ "0.3284", "0.2156", "0.378", "0.2156", "0.378", "0.231", "0.3284", "0.231" ], [ "0.4307", "0.2162", "0.4626", "0.2162", "0.4626", "0.231", "0.4307", "0.231" ], [ "0.5156", "0.2257", "0.882", "0.2257", "0.882", "0.2393", "0.5156", "0.2393" ], [ "0.1486", "0.2162", "0.2065", "0.2162", "0.2065", "0.231", "0.1486", "0.231" ], [ "0.147", "0.2299", "0.2225", "0.2299", "0.2225", "0.2446", "0.147", "0.2446" ], [ "0.2703", "0.2299", "0.3015", "0.2299", "0.3015", "0.2446", "0.2703", "0.2446" ], [ "0.339", "0.231", "0.3762", "0.231", "0.3762", "0.2434", "0.339", "0.2434" ], [ "0.424", "0.231", "0.461", "0.231", "0.461", "0.2434", "0.424", "0.2434" ], [ "0.4224", "0.2434", "0.4626", "0.2434", "0.4626", "0.2588", "0.4224", "0.2588" ], [ "0.5156", "0.2411", "0.879", "0.2411", "0.879", "0.2517", "0.5156", "0.2517" ], [ "0.339", "0.2452", "0.3762", "0.2452", "0.3762", "0.2583", "0.339", "0.2583" ], [ "0.147", "0.2465", "0.2225", "0.2465", "0.2225", "0.2583", "0.147", "0.2583" ], [ "0.2612", "0.2452", "0.3005", "0.2452", "0.3005", "0.2583", "0.2612", "0.2583" ], [ "0.5156", "0.2546", "0.882", "0.2546", "0.882", "0.2683", "0.5156", "0.2683" ], [ "0.5146", "0.2673", "0.742", "0.2695", "0.742", "0.284", "0.5146", "0.282" ], [ "0.7363", "0.2683", "0.882", "0.2683", "0.882", "0.282", "0.7363", "0.282" ], [ "0.5156", "0.2832", "0.882", "0.2832", "0.882", "0.2964", "0.5156", "0.2964" ], [ "0.1184", "0.2715", "0.4895", "0.2715", "0.4895", "0.2852", "0.1184", "0.2852" ], [ "0.12006", "0.2852", "0.1948", "0.2852", "0.1948", "0.2998", "0.12006", "0.2998" ], [ "0.517", "0.2976", "0.5825", "0.2976", "0.5825", "0.3088", "0.517", "0.3088" ], [ "0.532", "0.31", "0.882", "0.31", "0.882", "0.3237", "0.532", "0.3237" ], [ "0.12006", "0.3267", "0.4895", "0.3267", "0.4895", "0.3403", "0.12006", "0.3403" ], [ "0.5156", "0.3247", "0.882", "0.3247", "0.882", "0.3384", "0.5156", "0.3384" ], [ "0.12006", "0.3416", "0.34", "0.3416", "0.34", "0.3552", "0.12006", "0.3552" ], [ "0.513", "0.3362", "0.8833", "0.3374", "0.8833", "0.3552", "0.513", "0.354" ], [ "0.5156", "0.3528", "0.882", "0.3528", "0.882", "0.3665", "0.5156", "0.3665" ], [ "0.513", "0.3665", "0.882", "0.3665", "0.882", "0.3801", "0.513", "0.3801" ], [ "0.12006", "0.3735", "0.1814", "0.3735", "0.1814", "0.3855", "0.12006", "0.3855" ], [ "0.1814", "0.3748", "0.3896", "0.3748", "0.3896", "0.3855", "0.1814", "0.3855" ], [ "0.513", "0.3813", "0.881", "0.3813", "0.881", "0.3943", "0.513", "0.3943" ], [ "0.12177", "0.3955", "0.4878", "0.3955", "0.4878", "0.4092", "0.12177", "0.4092" ], [ "0.513", "0.3967", "0.784", "0.3967", "0.784", "0.4104", "0.513", "0.4104" ], [ "0.1184", "0.4104", "0.486", "0.4104", "0.486", "0.4233", "0.1184", "0.4233" ], [ "0.12006", "0.4246", "0.4895", "0.4246", "0.4895", "0.4382", "0.12006", "0.4382" ], [ "0.513", "0.427", "0.7773", "0.427", "0.7773", "0.4412", "0.513", "0.4412" ], [ "0.12006", "0.4395", "0.4895", "0.4395", "0.4895", "0.4531", "0.12006", "0.4531" ], [ "0.12006", "0.4531", "0.4878", "0.4531", "0.4878", "0.466", "0.12006", "0.466" ], [ "0.5146", "0.4495", "0.8833", "0.4495", "0.8833", "0.4631", "0.5146", "0.4631" ], [ "0.12006", "0.4673", "0.4895", "0.4673", "0.4895", "0.481", "0.12006", "0.481" ], [ "0.513", "0.462", "0.882", "0.4631", "0.882", "0.478", "0.513", "0.4768" ], [ "0.1184", "0.4822", "0.4878", "0.4822", "0.4878", "0.4954", "0.1184", "0.4954" ], [ "0.5156", "0.478", "0.882", "0.478", "0.882", "0.4912", "0.5156", "0.4912" ], [ "0.12177", "0.4976", "0.4878", "0.4976", "0.4878", "0.5083", "0.12177", "0.5083" ], [ "0.5156", "0.4922", "0.8833", "0.4922", "0.8833", "0.506", "0.5156", "0.506" ], [ "0.12006", "0.5103", "0.4895", "0.5103", "0.4895", "0.525", "0.12006", "0.525" ], [ "0.5146", "0.5073", "0.8716", "0.5073", "0.8716", "0.521", "0.5146", "0.521" ], [ "0.12006", "0.525", "0.4844", "0.525", "0.4844", "0.538", "0.12006", "0.538" ], [ "0.5146", "0.521", "0.882", "0.521", "0.882", "0.5337", "0.5146", "0.5337" ], [ "0.1351", "0.5376", "0.4878", "0.5376", "0.4878", "0.5503", "0.1351", "0.5503" ], [ "0.5156", "0.5337", "0.8833", "0.5337", "0.8833", "0.5474", "0.5156", "0.5474" ], [ "0.12006", "0.5527", "0.4895", "0.5527", "0.4895", "0.5664", "0.12006", "0.5664" ], [ "0.5146", "0.549", "0.8716", "0.549", "0.8716", "0.5625", "0.5146", "0.5625" ], [ "0.12006", "0.568", "0.4878", "0.568", "0.4878", "0.5806", "0.12006", "0.5806" ], [ "0.5156", "0.5635", "0.882", "0.5635", "0.882", "0.5767", "0.5156", "0.5767" ], [ "0.12177", "0.583", "0.4878", "0.583", "0.4878", "0.5933", "0.12177", "0.5933" ], [ "0.5146", "0.5776", "0.833", "0.5776", "0.833", "0.5913", "0.5146", "0.5913" ], [ "0.12006", "0.5957", "0.4844", "0.5957", "0.4844", "0.6094", "0.12006", "0.6094" ], [ "0.6055", "0.5996", "0.7925", "0.6025", "0.7925", "0.6206", "0.6055", "0.6177" ], [ "0.1351", "0.6094", "0.4844", "0.6094", "0.4844", "0.6226", "0.1351", "0.6226" ], [ "0.86", "0.604", "0.8833", "0.604", "0.8833", "0.616", "0.86", "0.616" ], [ "0.12006", "0.6245", "0.4895", "0.6245", "0.4895", "0.638", "0.12006", "0.638" ], [ "0.511", "0.627", "0.8486", "0.627", "0.8486", "0.64", "0.511", "0.64" ], [ "0.1184", "0.637", "0.4878", "0.637", "0.4878", "0.651", "0.1184", "0.651" ], [ "0.529", "0.6426", "0.8833", "0.6426", "0.8833", "0.6562", "0.529", "0.6562" ], [ "0.12006", "0.653", "0.4895", "0.653", "0.4895", "0.6665", "0.12006", "0.6665" ], [ "0.5146", "0.6562", "0.885", "0.6562", "0.885", "0.6694", "0.5146", "0.6694" ], [ "0.12177", "0.6685", "0.2544", "0.6685", "0.2544", "0.6787", "0.12177", "0.6787" ], [ "0.5156", "0.6704", "0.8506", "0.6704", "0.8506", "0.684", "0.5156", "0.684" ], [ "0.1368", "0.6816", "0.4895", "0.6816", "0.4895", "0.6953", "0.1368", "0.6953" ], [ "0.531", "0.686", "0.881", "0.686", "0.881", "0.697", "0.531", "0.697" ], [ "0.1184", "0.695", "0.4902", "0.697", "0.4902", "0.7104", "0.1184", "0.708" ], [ "0.5146", "0.6997", "0.882", "0.6997", "0.882", "0.7134", "0.5146", "0.7134" ], [ "0.12006", "0.7104", "0.4895", "0.7104", "0.4895", "0.7236", "0.12006", "0.7236" ], [ "0.513", "0.7134", "0.882", "0.7134", "0.882", "0.727", "0.513", "0.727" ], [ "0.1184", "0.7236", "0.4878", "0.7246", "0.4878", "0.7383", "0.1184", "0.737" ], [ "0.5146", "0.727", "0.882", "0.727", "0.882", "0.741", "0.5146", "0.741" ], [ "0.12006", "0.7383", "0.4878", "0.7383", "0.4878", "0.752", "0.12006", "0.752" ], [ "0.5146", "0.741", "0.882", "0.741", "0.882", "0.755", "0.5146", "0.755" ], [ "0.12006", "0.752", "0.378", "0.752", "0.378", "0.7656", "0.12006", "0.7656" ], [ "0.5156", "0.756", "0.882", "0.756", "0.882", "0.7695", "0.5156", "0.7695" ], [ "0.1368", "0.767", "0.4878", "0.767", "0.4878", "0.781", "0.1368", "0.781" ], [ "0.5146", "0.7695", "0.8003", "0.7695", "0.8003", "0.783", "0.5146", "0.783" ], [ "0.12006", "0.782", "0.486", "0.782", "0.486", "0.7964", "0.12006", "0.7964" ], [ "0.1184", "0.7944", "0.4878", "0.796", "0.4878", "0.8086", "0.1184", "0.808" ], [ "0.5146", "0.8003", "0.714", "0.8003", "0.714", "0.814", "0.5146", "0.814" ], [ "0.12177", "0.81", "0.4878", "0.81", "0.4878", "0.8237", "0.12177", "0.8237" ], [ "0.12177", "0.8247", "0.4878", "0.8247", "0.4878", "0.8384", "0.12177", "0.8384" ], [ "0.5146", "0.8237", "0.882", "0.8247", "0.882", "0.8384", "0.5146", "0.8374" ], [ "0.12006", "0.8384", "0.4895", "0.8384", "0.4895", "0.8516", "0.12006", "0.8516" ], [ "0.513", "0.8384", "0.882", "0.8374", "0.882", "0.851", "0.513", "0.8516" ], [ "0.12006", "0.8525", "0.4878", "0.8525", "0.4878", "0.8677", "0.12006", "0.8677" ], [ "0.5146", "0.8525", "0.882", "0.8525", "0.882", "0.866", "0.5146", "0.866" ], [ "0.12006", "0.8677", "0.2217", "0.8677", "0.2217", "0.879", "0.12006", "0.879" ], [ "0.519", "0.8677", "0.8696", "0.8677", "0.8696", "0.8813", "0.519", "0.8813" ], [ "0.1368", "0.8813", "0.4878", "0.8813", "0.4878", "0.8945", "0.1368", "0.8945" ], [ "0.5146", "0.8813", "0.882", "0.8813", "0.882", "0.8945", "0.5146", "0.8945" ], [ "0.12006", "0.8955", "0.4895", "0.8955", "0.4895", "0.91", "0.12006", "0.91" ], [ "0.5156", "0.898", "0.881", "0.898", "0.881", "0.9077", "0.5156", "0.9077" ] ]

[ "Test data", "Test data", "Training data", "MAN", "MAN", "JPN", "ENG", "ENG", "JPN", "2.387", "Training data", "English", "English", "79.59", "59.43", "2.956", "3.401", "68.64", "Mandarin", "2.839", "2.817", "3.098", "Mandarin", "65.31", "84.49", "59.72", "3.306", "4.004", "2.329", "64.46", "67.89", "74.20", "Japanese", "Japanese", "Multi (proposed)", "2.832", "2.265", "Multi (proposed)", "77.16", "76.54", "84.60", "2.396", "Table 1: Test loss on cross-lingual performance", "Table3:", "Cross-lingual turn shift/hold prediction", "performance (balanced accuracy [% ])", "Dataset", "# Shift", "# Hold", "% Shift ", "1253", "11432", "9.9", "说话之前就已经完成了转弯。此外，在普通话示例", "English", "Mandarin", "718", "1807", "28.4", "42.9", "的后半部分，有一个来自蓝色的短反向通道，这是", "1371", "Japanese", "1029", "由pnow预测的。由于该值比pfuture强，因此这说", "明了如何将pnow和pfuture—", "-起用于预测反向通", "道，也用于预测在反向通道开始后蓝色不会产生更", "Table 2: Distribution of samples for turn shift/hold", "prediction", "长的话语", "日本示例（c）还说明了可以有效地预测转弯和", "型的结果表明，它可以预测语音活动，所有语言的", "保持。请注意，在蓝色的转弯开始时，中间有一个", "性能水平与语言匹配模型相同", "长时间的停顿，在此期间，Pnow和pfuture表现出", "谁将是下一个演讲者的不确定性。这些类型的情况", "通常在自然对话中观察到，在文献中被称为自我选", "4.1.3.", "TurnShift/holdPrediction", "择（Sacksetal.，1974）。口语对话系统可以利用", "虽然测试损失揭示了模型的一般性能，但数字很难", "这种信息，使其可以转弯或留给用户", "解释。因此，我们还评估了多语言VAP模型在预", "测班次与。在相互沉默的时期举行。这是一个类似", "4.2.Languageldentification", "于口语对话系统中的回合结束预测的应用程序。此", "评估与先前研究中使用的评估相同E(Ekstedt and", "接下来，为了回答RQ3，我们研究了多语言模型", "Skantze，2022b）。任务是预测当相互静音长于0", "的语言识别能力。根据上一节中多语言模型在所有", "时，前一个扬声器和后一个扬声器是不同的（移位）", "三种语言中都表现良好，而单语模型在其他语言中", "还是相同的（保持）。观察25秒。请注意，前面和", "表现不佳的观察结果，我们假设多语言模型能够识", "后面的话语必须长于一秒。O后的pnow的值。相", "别输人语音的语言并相应地操作。为了研究这一", "互沉默开始后的05秒用于预测下一个说话者是谁", "点，我们在VAP模型的最后一层添加了另一个用", "移位/保持类的分布总结在表2中。由于对话的", "于语言识别的线性层，以及用于VAP和VAD的线", "性质因语言而异，因此轮班与保持的比率也不同。", "性层。由于我们要处理三种语言，因此它变成了", "例如，在日语数据中，停顿（保持）较少，话语更简", "个三类分类问题。然后，我们将基于交叉摘的语言", "洁。普通话数据的不平衡程度第二大，其次是英语。", "识别损失（Llid）添加到训练损失中，作为", "特别是在英语中，保持的频率大约是轮班的10倍", "L= Luap + Luad + Lid .", "尽管先前研究中使用的评估指标是加权的F1分", "(4)", "数，但在本研究中，我们使用了平衡的准确性来减", "我们然后在这个实验中从头开始训练这个模型", "少语言之间类别不平衡的偏差。此外，由于平衡精", "结果，当在测试集上测量语言识别准确性时，它", "度的值将为0。5对于随机或多数类预测，它还具", "达到了加权的F1-得分为99。99%。换句话说，多", "有易于解释的优点", "语言模型几乎能够完美地识别输入语音的语言", "表3显示了预测结果。结果与使用测试损失获", "我们还想看看这种增加的语言识别损失是否会", "得的结果一致。此外，在比较不同语言的准确性时，", "作为多任务损失，从而有可能提高多语言模型的", "很明显，普通话具有最可预测的转弯模式。从这些", "性能。表E4报告了与以前的模型相比，该新模型", "结果中，我们得出结论，单语模型不能直接应用于", "在VAP测试损失（Luap）上的性能以及转弯时换", "其他语言（RQ1），但是多语言模型可以用作所有", "档/保持预测的平衡精度。由于模型之间的结果非", "三种语言（RQ2）的通用转弯模型。", "常相似，我们得出的结论是，模型不需要帮助来学", "图5说明了每种语言数据集中多语言模型的输", "习识别语言，但无论如何它都是隐式的", "出示例。在这些图中，Pnow和Pfuture与用参考", "VAD段着色的输人波形一起示出。在英语示例（a）", "4.3.PitchSensitivity", "中，转弯从橙色参与者转移到蓝色参与者。在此转", "弯过程中，Pnou在蓝色语音之前表现出很高的预", "如E1节中所述，韵律信息是预测转弯的重要因素。", "测值。。此外，在每个发言者的回合（保持）内的", "为了间接评估模型对韵律信息的依赖并回答RQ4", "停顿期间，Pnow和pfuture正确地预测每个发言者", "我们在测试阶段（图E6）对输入语音的音调进行", "的回合的延续", "了平整，并测量了由此产生的性能下降。先前的", "在普通话示例（b）中观察到相同的模式。在这", "项研究进行了类似的测试，发现音调平坦对模型的", "种情况下，我们可以看到pfuture如何在蓝色停止", "性能有较小的整体影响，而在句法歧义的特定上下" ]

{ "arxiv_id": "2403.06487", "language": "zh", "page_idx": 4 }

[ [ "0.3857", "0.11426", "0.615", "0.11426", "0.615", "0.1288", "0.3857", "0.1288" ], [ "0.33", "0.1622", "0.6685", "0.1642", "0.6685", "0.1787", "0.33", "0.1768" ], [ "0.2222", "0.1863", "0.777", "0.1863", "0.777", "0.2008", "0.2222", "0.2008" ], [ "0.3367", "0.2058", "0.6626", "0.2058", "0.6626", "0.2203", "0.3367", "0.2203" ], [ "0.2484", "0.2715", "0.3245", "0.2734", "0.3245", "0.289", "0.2484", "0.2874" ], [ "0.5156", "0.2747", "0.9126", "0.2747", "0.9126", "0.2886", "0.5156", "0.2886" ], [ "0.1177", "0.2942", "0.452", "0.2966", "0.452", "0.3113", "0.1177", "0.3093" ], [ "0.5146", "0.2886", "0.9126", "0.2886", "0.9126", "0.303", "0.5146", "0.303" ], [ "0.513", "0.303", "0.5513", "0.303", "0.5513", "0.3145", "0.513", "0.3145" ], [ "0.1177", "0.3093", "0.452", "0.3113", "0.452", "0.3257", "0.1177", "0.324" ], [ "0.5312", "0.3164", "0.9126", "0.3164", "0.9126", "0.33", "0.5312", "0.33" ], [ "0.1209", "0.3245", "0.392", "0.3245", "0.392", "0.3384", "0.1209", "0.3384" ], [ "0.4004", "0.324", "0.451", "0.3257", "0.451", "0.3372", "0.4004", "0.3352" ], [ "0.5146", "0.33", "0.911", "0.33", "0.911", "0.3447", "0.5146", "0.3447" ], [ "0.11926", "0.3384", "0.452", "0.3384", "0.452", "0.3516", "0.11926", "0.3516" ], [ "0.513", "0.3435", "0.911", "0.3435", "0.911", "0.3567", "0.513", "0.3567" ], [ "0.11926", "0.3516", "0.452", "0.3516", "0.452", "0.3655", "0.11926", "0.3655" ], [ "0.528", "0.358", "0.91", "0.358", "0.91", "0.3718", "0.528", "0.3718" ], [ "0.11926", "0.3662", "0.451", "0.3662", "0.451", "0.3801", "0.11926", "0.3801" ], [ "0.513", "0.3718", "0.911", "0.3718", "0.911", "0.385", "0.513", "0.385" ], [ "0.11926", "0.3801", "0.452", "0.3801", "0.452", "0.3933", "0.11926", "0.3933" ], [ "0.513", "0.385", "0.9126", "0.385", "0.9126", "0.3984", "0.513", "0.3984" ], [ "0.11926", "0.3933", "0.451", "0.3933", "0.451", "0.4072", "0.11926", "0.4072" ], [ "0.5146", "0.3997", "0.911", "0.3997", "0.911", "0.4136", "0.5146", "0.4136" ], [ "0.1177", "0.4072", "0.4495", "0.4072", "0.4495", "0.4204", "0.1177", "0.4204" ], [ "0.5146", "0.4136", "0.6284", "0.4136", "0.6284", "0.4268", "0.5146", "0.4268" ], [ "0.11926", "0.4216", "0.452", "0.4216", "0.452", "0.435", "0.11926", "0.435" ], [ "0.5293", "0.428", "0.9126", "0.428", "0.9126", "0.4414", "0.5293", "0.4414" ], [ "0.11926", "0.435", "0.4534", "0.435", "0.4534", "0.449", "0.11926", "0.449" ], [ "0.5146", "0.4414", "0.911", "0.4414", "0.911", "0.455", "0.5146", "0.455" ], [ "0.11926", "0.449", "0.451", "0.449", "0.451", "0.4622", "0.11926", "0.4622" ], [ "0.5146", "0.4539", "0.911", "0.455", "0.911", "0.4697", "0.5146", "0.4685" ], [ "0.11926", "0.4634", "0.451", "0.4634", "0.451", "0.4766", "0.11926", "0.4766" ], [ "0.5156", "0.4685", "0.911", "0.4685", "0.911", "0.4817", "0.5156", "0.4817" ], [ "0.11926", "0.4766", "0.452", "0.4766", "0.452", "0.4905", "0.11926", "0.4905" ], [ "0.5146", "0.483", "0.9087", "0.483", "0.9087", "0.4968", "0.5146", "0.4968" ], [ "0.11926", "0.4905", "0.4534", "0.4905", "0.4534", "0.504", "0.11926", "0.504" ], [ "0.5146", "0.4968", "0.911", "0.4968", "0.911", "0.5103", "0.5146", "0.5103" ], [ "0.11926", "0.504", "0.4534", "0.504", "0.4534", "0.518", "0.11926", "0.518" ], [ "0.5146", "0.5103", "0.911", "0.5103", "0.911", "0.5244", "0.5146", "0.5244" ], [ "0.1177", "0.5166", "0.452", "0.5176", "0.452", "0.532", "0.1177", "0.531" ], [ "0.5146", "0.5244", "0.911", "0.5244", "0.911", "0.5386", "0.5146", "0.5386" ], [ "0.11926", "0.532", "0.4534", "0.532", "0.4534", "0.5454", "0.11926", "0.5454" ], [ "0.1177", "0.5444", "0.452", "0.5454", "0.452", "0.56", "0.1177", "0.5596" ], [ "0.5156", "0.5386", "0.911", "0.5386", "0.911", "0.552", "0.5156", "0.552" ], [ "0.5146", "0.552", "0.91", "0.552", "0.91", "0.565", "0.5146", "0.565" ], [ "0.1177", "0.558", "0.451", "0.558", "0.451", "0.5713", "0.1177", "0.5713" ], [ "0.513", "0.5645", "0.91", "0.5664", "0.91", "0.5806", "0.513", "0.579" ], [ "0.1177", "0.5737", "0.4534", "0.5737", "0.4534", "0.587", "0.1177", "0.587" ], [ "0.1177", "0.586", "0.452", "0.587", "0.452", "0.6016", "0.1177", "0.601" ], [ "0.5146", "0.58", "0.9126", "0.58", "0.9126", "0.5933", "0.5146", "0.5933" ], [ "0.513", "0.5913", "0.911", "0.5933", "0.911", "0.608", "0.513", "0.606" ], [ "0.1177", "0.601", "0.452", "0.5996", "0.452", "0.6143", "0.1177", "0.6157" ], [ "0.5146", "0.607", "0.911", "0.607", "0.911", "0.6206", "0.5146", "0.6206" ], [ "0.1177", "0.6143", "0.452", "0.6143", "0.452", "0.6274", "0.1177", "0.6274" ], [ "0.513", "0.619", "0.911", "0.6206", "0.911", "0.6353", "0.513", "0.634" ], [ "0.1209", "0.631", "0.451", "0.631", "0.451", "0.6416", "0.1209", "0.6416" ], [ "0.5146", "0.6353", "0.911", "0.6353", "0.911", "0.6484", "0.5146", "0.6484" ], [ "0.11926", "0.6426", "0.451", "0.6426", "0.451", "0.656", "0.11926", "0.656" ], [ "0.513", "0.6484", "0.863", "0.6484", "0.863", "0.662", "0.513", "0.662" ], [ "0.1177", "0.655", "0.4534", "0.656", "0.4534", "0.6704", "0.1177", "0.6694" ], [ "0.5312", "0.6636", "0.911", "0.6636", "0.911", "0.677", "0.5312", "0.677" ], [ "0.11926", "0.6704", "0.452", "0.6704", "0.452", "0.6846", "0.11926", "0.6846" ], [ "0.5156", "0.677", "0.91", "0.677", "0.91", "0.69", "0.5156", "0.69" ], [ "0.11926", "0.6846", "0.452", "0.6846", "0.452", "0.6978", "0.11926", "0.6978" ], [ "0.5156", "0.6914", "0.911", "0.6914", "0.911", "0.705", "0.5156", "0.705" ], [ "0.11926", "0.6987", "0.4355", "0.6987", "0.4355", "0.709", "0.11926", "0.709" ], [ "0.5156", "0.705", "0.9126", "0.705", "0.9126", "0.7183", "0.5156", "0.7183" ], [ "0.513", "0.7173", "0.911", "0.7183", "0.911", "0.733", "0.513", "0.732" ], [ "0.1136", "0.73", "0.2279", "0.731", "0.2279", "0.7456", "0.1136", "0.744" ], [ "0.513", "0.731", "0.911", "0.732", "0.911", "0.747", "0.513", "0.7456" ], [ "0.5146", "0.747", "0.9126", "0.747", "0.9126", "0.7603", "0.5146", "0.7603" ], [ "0.0866", "0.732", "0.1217", "0.732", "0.1217", "0.7437", "0.0866", "0.7437" ], [ "0.08826", "0.7505", "0.4844", "0.7505", "0.4844", "0.7646", "0.08826", "0.7646" ], [ "0.08826", "0.765", "0.4844", "0.765", "0.4844", "0.779", "0.08826", "0.779" ], [ "0.513", "0.7603", "0.824", "0.7603", "0.824", "0.7734", "0.513", "0.7734" ], [ "0.08826", "0.779", "0.4844", "0.779", "0.4844", "0.7935", "0.08826", "0.7935" ], [ "0.5107", "0.7827", "0.665", "0.785", "0.665", "0.799", "0.5107", "0.797" ], [ "0.08826", "0.7925", "0.486", "0.7925", "0.486", "0.806", "0.08826", "0.806" ], [ "0.0866", "0.8047", "0.4844", "0.806", "0.4844", "0.8193", "0.0866", "0.8184" ], [ "0.5146", "0.806", "0.9087", "0.806", "0.9087", "0.8193", "0.5146", "0.8193" ], [ "0.08905", "0.8213", "0.4836", "0.8213", "0.4836", "0.832", "0.08905", "0.832" ], [ "0.513", "0.821", "0.911", "0.821", "0.911", "0.834", "0.513", "0.834" ], [ "0.08826", "0.834", "0.486", "0.834", "0.486", "0.8477", "0.08826", "0.8477" ], [ "0.5146", "0.834", "0.911", "0.834", "0.911", "0.8477", "0.5146", "0.8477" ], [ "0.0866", "0.8477", "0.486", "0.8477", "0.486", "0.8613", "0.0866", "0.8613" ], [ "0.5156", "0.8477", "0.9126", "0.8477", "0.9126", "0.8613", "0.5156", "0.8613" ], [ "0.08826", "0.8623", "0.4844", "0.8623", "0.4844", "0.8755", "0.08826", "0.8755" ], [ "0.511", "0.86", "0.91", "0.8613", "0.91", "0.8755", "0.511", "0.875" ], [ "0.0866", "0.8755", "0.486", "0.8755", "0.486", "0.8896", "0.0866", "0.8896" ], [ "0.513", "0.8755", "0.911", "0.8755", "0.911", "0.8896", "0.513", "0.8896" ] ]

[ "从原始数据中归纳知识", "Wang-Zhou Dai, Stephen H.Muggleton", "Department of Computing, Imperial College London, London, UK", "f w.dai, s.muggleton I @imperial.ac.uk", "Abstract", "分模块来准确地表示一般的关系知识，它可能包含复杂", "For many reasoning-heavy tasks involving raw", "的推理结构，如递归[Glasmachers，2017；Garcezetal.,", "2019", "inputs, it is challenging to design an appro-", "因此，许多研究人员建议将端到端学习管道分开，", "priate end-to-end learning pipeline.", "Neuro-", "并构建一个由较小模块组成的混合模型，其中每个", "Symbolic Learning, divide the process into sub-", "模块仅占一个特定功能E[Glasmachers，2017]。这", "symbolic perception and symbolic reasoning,", "研究领域的一个代表性分支是神经符号（Neuro-", "trying to utilise data-driven machine learn-", "Symbolic)AI[De Raedt et al., 2020; Garcez et al.", "ing and knowledge-driven reasoning simultane-", "2019]，旨在桥接系统1和系统2AI[Kahneman，2011;", "ously. However, they suffer from the exponen-", "Bengio，2017]，i。e。基于神经网络的机器学习和基于", "tial computational complexity within the in", "符号的关系推理", "terface between these two components, where", "但是，基于从原始数据中提取的事实及其真值，在神", "the sub-symbolic learning model lacks direct", "经和符号系统之间的不可微分接口中缺乏监督，导致学", "supervision, and the symbolic model lacks ac-", "习E[Lietal.,2020;Daietal.，2019]的计算复杂度很高。", "curate input facts. Hence, most of them as-", "因此，几乎所有的神经符号模型都假定存在非常强大的", "sume the existence of a strong symbolic knowl-", "预定义领域知识库，并且无法执行程序归纳。这限制了", "edge base and only learn the perception model", "混合结构模型的表达能力，并栖性了符号学习的许多好", "while avoiding a crucial problem: where does", "处（例如。g。，谓词发明，学习递归理论，并将学习的", "the knowledge come from? In this paper, we", "模型重新用作背景知识)。在本文中，我们将神经网络与", "present Abductive Meta-Interpretive Learning", "(MetaAbd ) that unites abduction and induc-", "归纳逻辑编程（ILP)E[MuggletonanddeRaedt，1994]", "集成在一起，以实现从原始数据进行一阶逻辑理论归纳", "tion to learn neural networks and induce logic", "更具体地说，我们提出了归纳元解释学习（MetaAbd）", "theories jointly from raw data. Experimental", "results demonstrate that MetaAbd not only out-", "它通过将逻辑归纳和归纳E[Flachetal.，2000]与元解释", "学习（MIL）E[Muggletonetal，2015]中的神经网络相", "performs the compared systems in predictive", "结合，扩展了归纳学习（ABL）框架E[Daietal.2019", "accuracy and data efficiency but also induces", "Zhou,2019]。MetaAbd利用神经网络从原始数据中提取", "logic programsthat canbe re-used as back-", "概率逻辑事实，并归纳出一个可有效推断事实真值的演", "ground knowledge in subsequent learning tasks.", "绎逻辑程序[Kakasetal.,1992]来训练神经模型", "To the best of our knowledge, MetaAbd is the", "据我们所知，MetaAbd是第一个可以同时（1）从头开", "first system that can jointly learn neural net-", "始训练神经模型，（2）学习递归逻辑理论和（3）从具有", "works from scratch and induce recursive first-", "子符号表示的域中执行谓词发明的系统。在实验中，我", "orderlogictheorieswithpredicateinvention", "们将MetaAbd与比较的最先进的端到端深度学习模型和", "神经符号方法在两个复杂的学习任务上进行了比较。结", "Introduction", "果表明，在背景知识量相同的情况下，MetaAbd在预测", "精度和数据效率方面明显优于对比模型，并学习了可在", "1", "尽管数据驱动的端到端深度学习在许多传统的机器学习", "任务中取得了成功，但事实证明，对于某些复杂的学习问", "后续学习任务中重复使用的人类可解释模型", "题，整合领域知识仍然是必要的E[Dhingraetal.，2020;", "2RelatedWork", "Grover et al.,2019;Trasketal.,2018。为了利用离", "散和关系的复杂领域知识，端到端学习系统需要用可", "解决“系统2”问题需要关系和逻辑推理的能力E[Kah", "嵌入深度学习环境中的可区分模块来表示它。例如，图", "neman，201l；Bengio，2017]。由于其复杂性，许多研", "神经网络（GNN）使用关系图作为外部知识库E[Zhou", "究人员试图将复杂的背景知识嵌人端到端深度学习模型", "etal.，2018]；有些作品甚至考虑了更具体的领域知", "中。。例如，[Trasketal.,2018]提出了可微分神经算术", "识，例如可微的原始谓词和程序E[Dongetal.，2019;", "逻辑单元(NALU）来模拟基本算术函数(e。g。、加法", "Gauntetal.，2017]。然而，很难设计一个统一的可微", "乘法等。）在神经细胞中;E[Groveretal.，2019]用随机" ]

{ "arxiv_id": "2010.03514", "language": "zh", "page_idx": 2 }

[ [ "0.1765", "0.0341", "0.5073", "0.0341", "0.5073", "0.0486", "0.1765", "0.0486" ], [ "0.1765", "0.1042", "0.8203", "0.1048", "0.8203", "0.1225", "0.1765", "0.1218" ], [ "0.1765", "0.1276", "0.3162", "0.1276", "0.3162", "0.1497", "0.1765", "0.1497" ], [ "0.1871", "0.1749", "0.8364", "0.1736", "0.8364", "0.1913", "0.1871", "0.1925" ], [ "0.1854", "0.1913", "0.8423", "0.1913", "0.8423", "0.2058", "0.1854", "0.2058" ], [ "0.1847", "0.207", "0.722", "0.207", "0.722", "0.2203", "0.1847", "0.2203" ], [ "0.1854", "0.2216", "0.6562", "0.2216", "0.6562", "0.2349", "0.1854", "0.2349" ], [ "0.1847", "0.2349", "0.6055", "0.2349", "0.6055", "0.2488", "0.1847", "0.2488" ], [ "0.464", "0.2854", "0.536", "0.2854", "0.536", "0.2998", "0.464", "0.2998" ], [ "0.2345", "0.3208", "0.761", "0.3208", "0.761", "0.334", "0.2345", "0.334" ], [ "0.2336", "0.334", "0.763", "0.3333", "0.763", "0.348", "0.2336", "0.3484" ], [ "0.2336", "0.3467", "0.761", "0.3467", "0.761", "0.361", "0.2336", "0.361" ], [ "0.2345", "0.361", "0.761", "0.361", "0.761", "0.375", "0.2345", "0.375" ], [ "0.2345", "0.3757", "0.7646", "0.3757", "0.7646", "0.3894", "0.2345", "0.3894" ], [ "0.2336", "0.3894", "0.7656", "0.3894", "0.7656", "0.4028", "0.2336", "0.4028" ], [ "0.2345", "0.4016", "0.761", "0.4028", "0.761", "0.4167", "0.2345", "0.4153" ], [ "0.2336", "0.4172", "0.763", "0.4172", "0.763", "0.4312", "0.2336", "0.4312" ], [ "0.2336", "0.4312", "0.7646", "0.4312", "0.7646", "0.4443", "0.2336", "0.4443" ], [ "0.2361", "0.4463", "0.7607", "0.4463", "0.7607", "0.457", "0.2361", "0.457" ], [ "0.2345", "0.459", "0.7656", "0.459", "0.7656", "0.473", "0.2345", "0.473" ], [ "0.2336", "0.473", "0.7656", "0.473", "0.7656", "0.486", "0.2336", "0.486" ], [ "0.2345", "0.486", "0.7646", "0.486", "0.7646", "0.5", "0.2345", "0.5" ], [ "0.2336", "0.5005", "0.66", "0.5005", "0.66", "0.5146", "0.2336", "0.5146" ], [ "0.6543", "0.502", "0.761", "0.502", "0.761", "0.512", "0.6543", "0.512" ], [ "0.232", "0.5146", "0.7656", "0.512", "0.7656", "0.5264", "0.232", "0.529" ], [ "0.2361", "0.528", "0.7656", "0.528", "0.7656", "0.5415", "0.2361", "0.5415" ], [ "0.2345", "0.5415", "0.7646", "0.5415", "0.7646", "0.5547", "0.2345", "0.5547" ], [ "0.232", "0.5547", "0.7646", "0.5537", "0.7646", "0.5684", "0.232", "0.5693" ], [ "0.2345", "0.5693", "0.763", "0.5693", "0.763", "0.5835", "0.2345", "0.5835" ], [ "0.2345", "0.5835", "0.7646", "0.5835", "0.7646", "0.5977", "0.2345", "0.5977" ], [ "0.232", "0.5977", "0.5312", "0.595", "0.5312", "0.6094", "0.232", "0.6113" ], [ "0.1708", "0.6426", "0.3162", "0.646", "0.3162", "0.661", "0.1708", "0.6577" ], [ "0.1765", "0.678", "0.822", "0.678", "0.822", "0.6914", "0.1765", "0.6914" ], [ "0.1772", "0.6924", "0.8228", "0.6924", "0.8228", "0.7056", "0.1772", "0.7056" ], [ "0.1765", "0.7056", "0.8203", "0.7056", "0.8203", "0.72", "0.1765", "0.72" ], [ "0.1765", "0.7197", "0.824", "0.7197", "0.824", "0.733", "0.1765", "0.733" ], [ "0.1765", "0.732", "0.822", "0.7344", "0.822", "0.7485", "0.1765", "0.747" ], [ "0.1765", "0.7476", "0.822", "0.7476", "0.822", "0.761", "0.1765", "0.761" ], [ "0.1772", "0.761", "0.8228", "0.761", "0.8228", "0.7744", "0.1772", "0.7744" ], [ "0.1765", "0.7744", "0.8228", "0.7744", "0.8228", "0.7886", "0.1765", "0.7886" ], [ "0.179", "0.7905", "0.3938", "0.7905", "0.3938", "0.8003", "0.179", "0.8003" ], [ "0.1765", "0.81", "0.822", "0.81", "0.822", "0.8237", "0.1765", "0.8237" ], [ "0.1748", "0.8228", "0.8228", "0.8237", "0.8228", "0.8384", "0.1748", "0.837" ], [ "0.1732", "0.8384", "0.822", "0.8384", "0.822", "0.8516", "0.1732", "0.8516" ], [ "0.1765", "0.8516", "0.812", "0.8516", "0.812", "0.8657", "0.1765", "0.8657" ], [ "0.1765", "0.8657", "0.824", "0.8657", "0.824", "0.879", "0.1765", "0.879" ], [ "0.1765", "0.88", "0.6465", "0.88", "0.6465", "0.893", "0.1765", "0.893" ], [ "0.1953", "0.911", "0.492", "0.911", "0.492", "0.925", "0.1953", "0.925" ], [ "0.4927", "0.9497", "0.5063", "0.9497", "0.5063", "0.961", "0.4927", "0.961" ] ]

[ "Published as a conference paper at ICLR 2024", "通过多种群感知优化检测机器生成的文本，以实现最", "大平均差异", "Shuhai Zhang12* Yiliao Song3+, Jiahao Yang, Yuanqing Li21t, Bo Han5t, Mingkui Tan14t", "South China University of TechnologylPazhou Laboratory2The University of Adelaide3", "Key Laboratory of Big Data and Intelligent Robot, Ministry of Education4", "Department of Computer Science, Hong Kong Baptist Universitys", "[email protected]; [email protected]", "Abstract", "Large language models (LLMs) such as ChatGPT have exhibited remark", "able performance in generating human-like texts.However, machine-", "generated texts (MGTs) may carry critical risks, such as plagiarism issues", "misleading information, or hallucination issues. Therefore, it is very ur", "gent and important to detect MGTs in many situations. Unfortunately, it", "is challenging to distinguish MGTs and human-written texts because the", "distributional discrepancy between them is often very subtle due to the re", "markable performance of LLMs. In this paper, we seek to exploit maximum", "mean discrepancy (MMD) to address this issue in the sense that MMD can", "well identify distributional discrepancies.However,directlytraining a de", "tector with MMD using diverse MGTs will incur a significantly increased", "variance of MMD since MGTs may contain multiple text populations due", "to various LLMs. This will severely impair MMD's ability to measure the", "difference between two samples: To tackle this, we propose", "eanovelmulti", "population aware optimization method for MMD called MMD-MP, which", "can avoid variance increases and thus improve the stability to measure the", "distributional discrepancy. Relying on MMD-MP, we develop two methods", "for paragraph-based and sentence-based detection, respectively. Extensive", "experiments on various LLMs, e.g., GPT2 and ChatGPT, show superion", "detection performance of our MMD-MP. The source code is available at", "https://github.com/ZSHsh98/MMD-MP", "1Introduction", "随着大型语言模型（LLM）的发展，这些模型（例如GPT3（BroWnetal.，2020））生成的", "文本自然，流畅且高质量。这些机器生成的文本（MGT）与人类生成的文本（HWTs）非", "常相似，并且在自然语言处理，e.g.，文本摘要（Liu&Lapata，2019），对话生成（Lietal.", "2016）和机器翻译（Bahdanauetal.，2014）中具有许多有前途的应用。。然而，现有的LLM", "可能会产生虚假新闻（Zellersetal.，2019），垃圾邮件（Guoetal.，2020）和网络钓鱼（Hong，", "2012），遭受事实错误，幻觉和偏见（Zhangetal.，2023b；Lietal.，2023）。这对在线信息的", "可信度和安全性（Liuetal.，2023；Zhouetal.，2023a;b）构成了威胁，需要先进的MGT检", "测技术。不幸的是，区分MGT和HWT是具有挑战性的，因为它们之间的分布差异本质", "上是微妙的（Tianetal.，2023）", "为了检测MGT，现有的基于度量的方法（Gehrmannetal.，2019；Mitchelletal.，2023", "Solaimanetal.，2019）使用一些统计信息（e.g.，对数似然）对测试文本为MGT的概率进", "行评分，当用于训练评分模型的文本与测试的MGT之间存在较大的语言域差距时，这种", "方法的效果较差。另一种策略是基于模型的方法（Solaimanetal.,2019；Guoetal.，2023）", "它严重依赖特定的MGT类型，难以适应其他类型的MGT。这些方法在有效捕获MGT和", "HWT之间的分布差异方面面临挑战，从而限制了它们的检测能力", "*Equal contribution.Corresponding author.", "1" ]

{ "arxiv_id": "2402.16041", "language": "zh", "page_idx": 0 }

[ [ "0.1765", "0.0341", "0.5073", "0.0341", "0.5073", "0.04736", "0.1765", "0.04736" ], [ "0.232", "0.0884", "0.3015", "0.0884", "0.3015", "0.09534", "0.232", "0.09534" ], [ "0.4126", "0.084", "0.4265", "0.084", "0.4265", "0.0871", "0.4126", "0.0871" ], [ "0.2484", "0.09344", "0.3015", "0.0947", "0.3015", "0.1029", "0.2484", "0.1016" ], [ "0.4429", "0.0884", "0.5073", "0.0884", "0.5073", "0.09534", "0.4429", "0.09534" ], [ "0.4558", "0.0947", "0.5073", "0.09534", "0.5073", "0.1042", "0.4558", "0.1029" ], [ "0.2484", "0.101", "0.2932", "0.1042", "0.2917", "0.11554", "0.2467", "0.11237" ], [ "0.3987", "0.101", "0.4307", "0.101", "0.4307", "0.108", "0.3987", "0.108" ], [ "0.4429", "0.1029", "0.451", "0.1029", "0.451", "0.108", "0.4429", "0.108" ], [ "0.4558", "0.1048", "0.468", "0.09973", "0.4739", "0.1067", "0.4631", "0.11237" ], [ "0.2566", "0.1111", "0.2852", "0.1111", "0.2852", "0.11426", "0.2566", "0.11426" ], [ "0.2378", "0.1029", "0.2443", "0.1029", "0.2443", "0.108", "0.2378", "0.108" ], [ "0.4631", "0.1111", "0.486", "0.1111", "0.486", "0.11426", "0.4631", "0.11426" ], [ "0.2532", "0.11743", "0.3022", "0.11743", "0.3022", "0.125", "0.2532", "0.125" ], [ "0.411", "0.1206", "0.429", "0.1206", "0.429", "0.125", "0.411", "0.125" ], [ "0.4534", "0.11743", "0.5146", "0.11743", "0.5146", "0.125", "0.4534", "0.125" ], [ "0.4004", "0.1218", "0.407", "0.1218", "0.407", "0.1276", "0.4004", "0.1276" ], [ "0.2018", "0.1225", "0.2239", "0.1225", "0.2239", "0.13", "0.2018", "0.13" ], [ "0.2336", "0.1218", "0.2524", "0.1276", "0.2484", "0.1351", "0.2296", "0.1288" ], [ "0.25", "0.127", "0.304", "0.127", "0.304", "0.1307", "0.25", "0.1307" ], [ "0.4395", "0.1237", "0.5146", "0.125", "0.5146", "0.1332", "0.4395", "0.132" ], [ "0.4004", "0.1332", "0.411", "0.1332", "0.411", "0.1528", "0.4004", "0.1528" ], [ "0.6855", "0.1338", "0.7344", "0.1338", "0.7344", "0.1414", "0.6855", "0.1414" ], [ "0.4084", "0.137", "0.4314", "0.137", "0.4314", "0.1459", "0.4084", "0.1459" ], [ "0.2009", "0.1364", "0.2255", "0.137", "0.2255", "0.1459", "0.2009", "0.1445" ], [ "0.687", "0.1427", "0.7344", "0.1427", "0.7344", "0.1497", "0.687", "0.1497" ], [ "0.398", "0.1528", "0.433", "0.1528", "0.433", "0.1642", "0.398", "0.1642" ], [ "0.619", "0.1509", "0.625", "0.1509", "0.625", "0.1572", "0.619", "0.1572" ], [ "0.693", "0.1572", "0.7173", "0.1572", "0.7173", "0.161", "0.693", "0.161" ], [ "0.4043", "0.1674", "0.569", "0.1604", "0.5703", "0.1812", "0.406", "0.1881" ], [ "0.671", "0.1686", "0.683", "0.1686", "0.683", "0.1768", "0.671", "0.1768" ], [ "0.6807", "0.1692", "0.7427", "0.1692", "0.7427", "0.1799", "0.6807", "0.1799" ], [ "0.683", "0.1635", "0.7417", "0.1635", "0.7417", "0.1704", "0.683", "0.1704" ], [ "0.2239", "0.1718", "0.2998", "0.1718", "0.2998", "0.1818", "0.2239", "0.1818" ], [ "0.625", "0.1749", "0.642", "0.1749", "0.642", "0.1843", "0.625", "0.1843" ], [ "0.312", "0.1818", "0.3652", "0.1799", "0.3652", "0.1901", "0.312", "0.1925" ], [ "0.528", "0.1818", "0.5728", "0.1799", "0.5728", "0.1901", "0.528", "0.1925" ], [ "0.732", "0.1818", "0.774", "0.1818", "0.774", "0.1925", "0.732", "0.1925" ], [ "0.2443", "0.1831", "0.2712", "0.1831", "0.2712", "0.1913", "0.2443", "0.1913" ], [ "0.2255", "0.1831", "0.246", "0.1831", "0.246", "0.1901", "0.2255", "0.1901" ], [ "0.2737", "0.1831", "0.3145", "0.1831", "0.3145", "0.1989", "0.2737", "0.1989" ], [ "0.6567", "0.1831", "0.6763", "0.1831", "0.6763", "0.1913", "0.6567", "0.1913" ], [ "0.6855", "0.1831", "0.7305", "0.1831", "0.7305", "0.1976", "0.6855", "0.1976" ], [ "0.451", "0.1843", "0.4739", "0.1843", "0.4739", "0.1913", "0.451", "0.1913" ], [ "0.4314", "0.1843", "0.451", "0.1843", "0.451", "0.1881", "0.4314", "0.1881" ], [ "0.4805", "0.1843", "0.523", "0.1843", "0.523", "0.1989", "0.4805", "0.1989" ], [ "0.524", "0.1849", "0.5337", "0.1863", "0.5317", "0.1901", "0.523", "0.1895" ], [ "0.2524", "0.2039", "0.3137", "0.2039", "0.3137", "0.2153", "0.2524", "0.2153" ], [ "0.4126", "0.2039", "0.558", "0.2026", "0.558", "0.214", "0.4126", "0.2153" ], [ "0.613", "0.2026", "0.774", "0.2008", "0.774", "0.214", "0.613", "0.2166" ], [ "0.1748", "0.2203", "0.2573", "0.2197", "0.2573", "0.2311", "0.1748", "0.2317" ], [ "0.255", "0.2197", "0.8228", "0.2197", "0.8228", "0.2329", "0.255", "0.2329" ], [ "0.1748", "0.2343", "0.8228", "0.2343", "0.8228", "0.2474", "0.1748", "0.2474" ], [ "0.1765", "0.2488", "0.819", "0.25", "0.819", "0.2632", "0.1765", "0.262" ], [ "0.1748", "0.2632", "0.824", "0.2632", "0.824", "0.2766", "0.1748", "0.2766" ], [ "0.1748", "0.2778", "0.7876", "0.2778", "0.7876", "0.2917", "0.1748", "0.2917" ], [ "0.1772", "0.3113", "0.8228", "0.3113", "0.8228", "0.3213", "0.1772", "0.3213" ], [ "0.1765", "0.324", "0.8228", "0.324", "0.8228", "0.3372", "0.1765", "0.3372" ], [ "0.1765", "0.3384", "0.818", "0.3384", "0.818", "0.3516", "0.1765", "0.3516" ], [ "0.1732", "0.3499", "0.8228", "0.351", "0.8228", "0.3655", "0.1732", "0.3643" ], [ "0.179", "0.3655", "0.8228", "0.3655", "0.8228", "0.379", "0.179", "0.379" ], [ "0.1772", "0.379", "0.8228", "0.379", "0.8228", "0.3926", "0.1772", "0.3926" ], [ "0.1748", "0.3926", "0.8228", "0.3926", "0.8228", "0.406", "0.1748", "0.406" ], [ "0.1765", "0.406", "0.721", "0.406", "0.721", "0.4192", "0.1765", "0.4192" ], [ "0.1765", "0.428", "0.8203", "0.428", "0.8203", "0.4414", "0.1765", "0.4414" ], [ "0.1765", "0.4414", "0.8228", "0.4414", "0.8228", "0.455", "0.1765", "0.455" ], [ "0.1748", "0.455", "0.8228", "0.4539", "0.8228", "0.4685", "0.1748", "0.4697" ], [ "0.1765", "0.4697", "0.8228", "0.4697", "0.8228", "0.483", "0.1765", "0.483" ], [ "0.1772", "0.483", "0.8228", "0.483", "0.8228", "0.4968", "0.1772", "0.4968" ], [ "0.1765", "0.4976", "0.8228", "0.4976", "0.8228", "0.511", "0.1765", "0.511" ], [ "0.1748", "0.511", "0.822", "0.511", "0.822", "0.5244", "0.1748", "0.5244" ], [ "0.179", "0.5264", "0.8203", "0.5264", "0.8203", "0.537", "0.179", "0.537" ], [ "0.1765", "0.5386", "0.824", "0.5386", "0.824", "0.552", "0.1765", "0.552" ], [ "0.1748", "0.552", "0.7915", "0.552", "0.7915", "0.565", "0.1748", "0.565" ], [ "0.1765", "0.5737", "0.8228", "0.5737", "0.8228", "0.587", "0.1765", "0.587" ], [ "0.1732", "0.586", "0.5522", "0.587", "0.5522", "0.601", "0.1732", "0.5996" ], [ "0.1748", "0.6094", "0.822", "0.6094", "0.822", "0.6226", "0.1748", "0.6226" ], [ "0.1748", "0.6226", "0.6733", "0.6226", "0.6733", "0.636", "0.1748", "0.636" ], [ "0.1748", "0.6445", "0.824", "0.6445", "0.824", "0.6577", "0.1748", "0.6577" ], [ "0.1765", "0.6577", "0.8228", "0.6577", "0.8228", "0.672", "0.1765", "0.672" ], [ "0.1765", "0.6724", "0.273", "0.6724", "0.273", "0.6846", "0.1765", "0.6846" ], [ "0.1765", "0.7183", "0.1896", "0.7183", "0.1896", "0.729", "0.1765", "0.729" ], [ "0.1871", "0.7197", "0.2075", "0.7197", "0.2075", "0.7266", "0.1871", "0.7266" ], [ "0.2075", "0.7173", "0.464", "0.7173", "0.464", "0.731", "0.2075", "0.731" ], [ "0.1732", "0.7524", "0.3145", "0.7524", "0.3145", "0.7666", "0.1732", "0.7666" ], [ "0.1765", "0.782", "0.8228", "0.782", "0.8228", "0.797", "0.1765", "0.797" ], [ "0.1765", "0.797", "0.8228", "0.797", "0.8228", "0.811", "0.1765", "0.811" ], [ "0.1748", "0.8125", "0.601", "0.811", "0.601", "0.8257", "0.1748", "0.827" ], [ "0.1772", "0.8335", "0.824", "0.8335", "0.824", "0.8477", "0.1772", "0.8477" ], [ "0.1765", "0.8477", "0.4412", "0.8477", "0.4412", "0.8613", "0.1765", "0.8613" ], [ "0.1765", "0.869", "0.4968", "0.869", "0.4968", "0.8833", "0.1765", "0.8833" ], [ "0.4902", "0.8706", "0.603", "0.8706", "0.603", "0.881", "0.4902", "0.881" ], [ "0.6167", "0.8706", "0.6685", "0.8706", "0.6685", "0.882", "0.6167", "0.882" ], [ "0.755", "0.8706", "0.778", "0.8706", "0.778", "0.882", "0.755", "0.882" ], [ "0.794", "0.869", "0.8228", "0.869", "0.8228", "0.882", "0.794", "0.882" ], [ "0.2507", "0.881", "0.5483", "0.882", "0.5483", "0.8965", "0.2507", "0.896" ], [ "0.699", "0.872", "0.726", "0.872", "0.726", "0.88", "0.699", "0.88" ], [ "0.7305", "0.8833", "0.755", "0.8833", "0.755", "0.896", "0.7305", "0.896" ], [ "0.789", "0.8833", "0.822", "0.8833", "0.822", "0.896", "0.789", "0.896" ], [ "0.576", "0.884", "0.694", "0.884", "0.694", "0.8945", "0.576", "0.8945" ], [ "0.1772", "0.884", "0.2296", "0.884", "0.2296", "0.896", "0.1772", "0.896" ], [ "0.1765", "0.8965", "0.9985", "0.8965", "0.9985", "0.91", "0.1765", "0.91" ], [ "0.1748", "0.91", "0.58", "0.911", "0.58", "0.9253", "0.1748", "0.925" ], [ "0.492", "0.9478", "0.5073", "0.9478", "0.5073", "0.962", "0.492", "0.962" ] ]

[ "Published as a conference paper at ICLR 2024", "含MMD-D（q=2）", "2", "MMD-D(q=3)", "MMD-D(q=2", "MMD-D(q=3", "MMD-D(q)", "0.020", "4", "MA", "n", "4", "", "MMD-MP(q=)", "0.015", "MMD-MP（9=3）", "t", "0.3", "e", "MMD-MDIa-/", "-MMD-MP（q=4）", "r", "MMD-D (=2)", "0.010", "≥ 0.2", "MMD-D(q=3", " 0.005", "", "MD-ME", "0.00 8886668889888", "中", "MMD-MP (q=4)", "MMD-MP(9=3)", "0000068", "20", "80100", "80100", "80100", "20", "0", "Epochs", "20", "Epochs", "20", "", "Epochs", "", "(a) MMD", "(b)VarianceofMMD", "(c) Test Power of MMD", "Figure l:", "Illustration of MMD values, MMD variances, and the test power of MMD-D and", "our MMD-MP during the optimization process. As the number of St populations (i.e., q", ") increases, MMD-D shows an increase in MMD, accompanied by a sharp rise in variance", "resulting in unstable test power during testing. In contrast, our MMD-MP exhibits minimal", "variance in MMD values, leading to higher and more stable test power during testing.", "在本文中，我们试图利用最大平均差异（MMD）来解决上述问题，因为它具有强大的识别", "分布差异的能力（Liuetal.，2020；2021）。然而，直接应用MMD不能有效地检测MGT。此", "任务通常涉及来自不同人群的数据，e.g.，由不同LLM（e.g.，GPT-3（Brownetal.，2020）", "ChatGPT（OpenAI，2022））或不同LLM设置（e.g：，温度，前k采样（Vilnisetal.，2023）", "）生成的文本。这些人群在语言风格和句法结构上可能存在很大差异，从而导致MGT的", "显着变化。在这种情况下训练深核MMD（MMD-D，Liuetal.（2020））会产生高方差问题", "(i.e.，图1（b）中MMD-D的大方差）。这意味着HWTs和MGTs之间的估计差异波动很", "大，从而导致不可靠和不稳定的检测（图1（c）中MMD-D的低测试功率）", "本文率先探索基于内核的MMD的优化机制。当我们使用来自多个种群的数据训练内核时", "估计的MMD随着其方差的显着增长而增加（参见图1（a）－（b）以及2.3节中的更多解", "释）。这种现象是由于基于内核的MMD的优化目标中MGT内的类内距离而产生的。此距", "离在很大程度上阻碍了旨在聚合MGT的优化过程，从而导致MGT检测的MMD波动很", "大。在本文中，我们为MMD提出了一种新颖的多种群感知优化方法，称为MMD-MP，该", "方法使用多种群代理来消除聚合MGT中所有实例的约束。通过这种方式，我们可以实现", "MGT和HWT之间的MMD的低方差，从而导致更稳定的差异估计和更可靠的检测（参", "见图1（b）-（c）中的MMD-MP结果）。此外，使用经过训练的深度内核，我们分别开", "发了两种基于段落的检测和基于句子的检测方法。ChatGPT，GPT2系列，GPT3系列和", "GPT-Neo系列等各种LLM的经验证据表明了我们方法的优越性。我们的贡献总结为：", "1）我们深入研究了MMD的优化机制，并揭示了在处理来自多个不同种群的训练数据时", "MMD的高方差会导致MGT检测的不稳定差异估计", "2）我们提出了一种新颖的多种群感知优化方法，用于训练基于内核的MMD（称为MMD", "MP），该方法可以缓解MMD-D的不良优化并提高差异度量的稳定性", "3）依靠提出的MMD-MP，我们设计了两种新颖的MGT检测方法。包括ChatGPT，GPT2", "系列，GPT3系列，GPT-Neo系列在内的众多LLM的广泛实验表明，我们的方法始终优", "于现有的基线", "2 ", "", "Preliminaries and Motivations", "2.1 Preliminaries", "双样本测试(2ST)。令P，Q为XcRd上的Borel概率度量。我们观察到独立的同分布", "（IID）数据Sp=[x)\"=1~Pn和SQ={y,)m=1~Qm。2ST旨在确定P和Q是否来自相同的", "分布,i.e.,P=Q(Borgwardt etal.,2006;Liu et al.2020)", "单实例检测(SID)。令P为XcRd和IID观测Sp=(x)-i~Pn上的Borel概率度量，SID", "旨在判断测试实例文是否来自分布P", "最大平均差异。在Gretton", "etal.", "(2012);", "et", "al.", "之后最大平均差异", "Liu", "和", "MA-", "XMAXXMAX1", "(2020)", "MAXXkerXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX", "由于f是固定的，因此kw的参数集是w={e,Φ,,og]", "2" ]

{ "arxiv_id": "2402.16041", "language": "zh", "page_idx": 1 }

[ [ "0.1765", "0.0341", "0.5073", "0.0341", "0.5073", "0.04736", "0.1765", "0.04736" ], [ "0.2695", "0.09784", "0.281", "0.09534", "0.2852", "0.1048", "0.2737", "0.1067" ], [ "0.4346", "0.09973", "0.4395", "0.09973", "0.4395", "0.1048", "0.4346", "0.1048" ], [ "0.589", "0.09973", "0.6094", "0.09973", "0.6094", "0.108", "0.589", "0.108" ], [ "0.6323", "0.101", "0.683", "0.101", "0.683", "0.1111", "0.6323", "0.1111" ], [ "0.6094", "0.1048", "0.6724", "0.1048", "0.6724", "0.1288", "0.6094", "0.1288" ], [ "0.3057", "0.1029", "0.3506", "0.1048", "0.3506", "0.11237", "0.3057", "0.10986" ], [ "0.4355", "0.108", "0.4436", "0.11426", "0.437", "0.1193", "0.429", "0.113" ], [ "0.4756", "0.1061", "0.5", "0.1067", "0.5", "0.10986", "0.4756", "0.1092" ], [ "0.468", "0.10986", "0.5107", "0.11237", "0.509", "0.1225", "0.4673", "0.1206" ], [ "0.273", "0.1092", "0.2852", "0.1092", "0.2852", "0.1193", "0.273", "0.1193" ], [ "0.3064", "0.1111", "0.3513", "0.1111", "0.3513", "0.1218", "0.3064", "0.1218" ], [ "0.5767", "0.10986", "0.5884", "0.10986", "0.5884", "0.1218", "0.5767", "0.1218" ], [ "0.5728", "0.11615", "0.6094", "0.11615", "0.6094", "0.1497", "0.5728", "0.1497" ], [ "0.3245", "0.1193", "0.3367", "0.1193", "0.3367", "0.1237", "0.3245", "0.1237" ], [ "0.4878", "0.1193", "0.5", "0.1193", "0.5", "0.1225", "0.4878", "0.1225" ], [ "0.2974", "0.1225", "0.3555", "0.1193", "0.357", "0.1338", "0.298", "0.137" ], [ "0.2917", "0.113", "0.3022", "0.113", "0.3022", "0.11615", "0.2917", "0.11615" ], [ "0.2573", "0.1237", "0.2852", "0.1237", "0.2852", "0.1604", "0.2573", "0.1604" ], [ "0.4697", "0.1256", "0.5063", "0.1256", "0.5063", "0.1332", "0.4697", "0.1332" ], [ "0.626", "0.1237", "0.669", "0.1237", "0.669", "0.1338", "0.626", "0.1338" ], [ "0.6313", "0.132", "0.6733", "0.132", "0.6733", "0.1427", "0.6313", "0.1427" ], [ "0.308", "0.1338", "0.3555", "0.1338", "0.3555", "0.1414", "0.308", "0.1414" ], [ "0.4697", "0.1338", "0.5146", "0.1338", "0.5146", "0.1414", "0.4697", "0.1414" ], [ "0.2942", "0.1351", "0.3022", "0.1351", "0.3022", "0.1395", "0.2942", "0.1395" ], [ "0.4224", "0.1276", "0.4307", "0.1276", "0.4307", "0.137", "0.4224", "0.137" ], [ "0.619", "0.1401", "0.6274", "0.1401", "0.6274", "0.1445", "0.619", "0.1445" ], [ "0.6323", "0.1401", "0.6914", "0.1401", "0.6914", "0.1477", "0.6323", "0.1477" ], [ "0.3406", "0.1433", "0.3545", "0.1433", "0.3545", "0.1465", "0.3406", "0.1465" ], [ "0.3137", "0.1427", "0.335", "0.1427", "0.335", "0.1459", "0.3137", "0.1459" ], [ "0.5786", "0.1459", "0.5884", "0.1459", "0.5884", "0.1674", "0.5786", "0.1674" ], [ "0.596", "0.1483", "0.6055", "0.1483", "0.6055", "0.1559", "0.596", "0.1559" ], [ "0.3137", "0.1654", "0.3342", "0.1654", "0.3342", "0.1686", "0.3137", "0.1686" ], [ "0.3064", "0.1718", "0.3652", "0.1718", "0.3652", "0.1787", "0.3064", "0.1787" ], [ "0.46", "0.1736", "0.4722", "0.1736", "0.4722", "0.1768", "0.46", "0.1768" ], [ "0.312", "0.1849", "0.3218", "0.1849", "0.3218", "0.1925", "0.312", "0.1925" ], [ "0.3677", "0.1849", "0.38", "0.1849", "0.38", "0.1913", "0.3677", "0.1913" ], [ "0.3962", "0.1843", "0.411", "0.1843", "0.411", "0.1913", "0.3962", "0.1913" ], [ "0.4714", "0.1849", "0.4805", "0.1849", "0.4805", "0.1925", "0.4714", "0.1925" ], [ "0.609", "0.1849", "0.614", "0.1849", "0.614", "0.1901", "0.609", "0.1901" ], [ "0.634", "0.1849", "0.649", "0.1849", "0.649", "0.1925", "0.634", "0.1925" ], [ "0.694", "0.1863", "0.706", "0.1863", "0.706", "0.1925", "0.694", "0.1925" ], [ "0.726", "0.1863", "0.7344", "0.1863", "0.7344", "0.1901", "0.726", "0.1901" ], [ "0.5557", "0.1881", "0.564", "0.1818", "0.5703", "0.1875", "0.562", "0.1932" ], [ "0.327", "0.1913", "0.3652", "0.1913", "0.3652", "0.202", "0.327", "0.202" ], [ "0.4878", "0.1913", "0.524", "0.1913", "0.524", "0.202", "0.4878", "0.202" ], [ "0.528", "0.1863", "0.5376", "0.1863", "0.5376", "0.1913", "0.528", "0.1913" ], [ "0.653", "0.1913", "0.693", "0.1913", "0.693", "0.202", "0.653", "0.202" ], [ "0.286", "0.2084", "0.38", "0.2084", "0.38", "0.2197", "0.286", "0.2197" ], [ "0.4436", "0.2084", "0.54", "0.2084", "0.54", "0.2197", "0.4436", "0.2197" ], [ "0.5767", "0.207", "0.738", "0.207", "0.738", "0.2203", "0.5767", "0.2203" ], [ "0.4697", "0.2361", "0.523", "0.2361", "0.523", "0.2468", "0.4697", "0.2468" ], [ "0.4182", "0.2468", "0.423", "0.2468", "0.423", "0.254", "0.4182", "0.254" ], [ "0.4575", "0.2443", "0.5073", "0.2443", "0.5073", "0.2551", "0.4575", "0.2551" ], [ "0.4575", "0.25", "0.4958", "0.252", "0.4958", "0.26", "0.4575", "0.2583" ], [ "0.4167", "0.2568", "0.4558", "0.2568", "0.4558", "0.2683", "0.4167", "0.2683" ], [ "0.452", "0.2651", "0.5107", "0.2651", "0.5107", "0.276", "0.452", "0.276" ], [ "0.4722", "0.276", "0.5293", "0.276", "0.5293", "0.2827", "0.4722", "0.2827" ], [ "0.4192", "0.2854", "0.4272", "0.2886", "0.423", "0.2954", "0.415", "0.2922" ], [ "0.437", "0.2842", "0.4436", "0.2842", "0.4436", "0.2886", "0.437", "0.2886" ], [ "0.4167", "0.2954", "0.4272", "0.2954", "0.4272", "0.305", "0.4167", "0.305" ], [ "0.523", "0.3037", "0.572", "0.3037", "0.572", "0.3145", "0.523", "0.3145" ], [ "0.478", "0.3193", "0.4844", "0.3245", "0.4795", "0.3289", "0.4722", "0.324" ], [ "0.5034", "0.3193", "0.5186", "0.3225", "0.5146", "0.33", "0.5", "0.327" ], [ "0.5293", "0.3208", "0.5513", "0.3208", "0.5513", "0.3289", "0.5293", "0.3289" ], [ "0.558", "0.318", "0.5767", "0.3213", "0.574", "0.3289", "0.5557", "0.3257" ], [ "0.4927", "0.327", "0.5186", "0.327", "0.5186", "0.334", "0.4927", "0.334" ], [ "0.4126", "0.3403", "0.4805", "0.3384", "0.4805", "0.3528", "0.4126", "0.3547" ], [ "0.482", "0.3452", "0.4885", "0.3484", "0.486", "0.3516", "0.4795", "0.3484" ], [ "0.4927", "0.3423", "0.576", "0.3423", "0.576", "0.3542", "0.4927", "0.3542" ], [ "0.4143", "0.3547", "0.4436", "0.3547", "0.4436", "0.3662", "0.4143", "0.3662" ], [ "0.1748", "0.3655", "0.8228", "0.3655", "0.8228", "0.379", "0.1748", "0.379" ], [ "0.1732", "0.379", "0.822", "0.379", "0.822", "0.3926", "0.1732", "0.3926" ], [ "0.1732", "0.3926", "0.822", "0.3926", "0.822", "0.4072", "0.1732", "0.4072" ], [ "0.1732", "0.4048", "0.8267", "0.4048", "0.8267", "0.4255", "0.1732", "0.4255" ], [ "0.1748", "0.4204", "0.343", "0.4216", "0.343", "0.435", "0.1748", "0.4343" ], [ "0.1765", "0.4475", "0.8228", "0.4475", "0.8228", "0.461", "0.1765", "0.461" ], [ "0.1748", "0.461", "0.822", "0.461", "0.822", "0.476", "0.1748", "0.476" ], [ "0.1765", "0.476", "0.4656", "0.476", "0.4656", "0.4893", "0.1765", "0.4893" ], [ "0.1887", "0.5", "0.8013", "0.5", "0.8013", "0.5176", "0.1887", "0.5176" ], [ "0.1748", "0.529", "0.3381", "0.529", "0.3381", "0.5425", "0.1748", "0.5425" ], [ "0.2917", "0.5474", "0.653", "0.556", "0.652", "0.584", "0.29", "0.5757" ], [ "0.1887", "0.56", "0.3022", "0.556", "0.304", "0.5737", "0.1896", "0.5776" ], [ "0.652", "0.5625", "0.689", "0.5645", "0.687", "0.5806", "0.6504", "0.579" ], [ "0.6787", "0.5566", "0.8228", "0.5566", "0.8228", "0.579", "0.6787", "0.579" ], [ "0.29", "0.5737", "0.4355", "0.5684", "0.437", "0.586", "0.2917", "0.5913" ], [ "0.1765", "0.5996", "0.423", "0.5996", "0.423", "0.613", "0.1765", "0.613" ], [ "0.1724", "0.629", "0.7236", "0.63", "0.7236", "0.6445", "0.1724", "0.643" ], [ "0.1772", "0.656", "0.824", "0.656", "0.824", "0.6694", "0.1772", "0.6694" ], [ "0.1765", "0.6704", "0.822", "0.6704", "0.822", "0.6846", "0.1765", "0.6846" ], [ "0.1765", "0.686", "0.824", "0.686", "0.824", "0.6997", "0.1765", "0.6997" ], [ "0.1732", "0.7007", "0.5317", "0.7007", "0.5317", "0.714", "0.1732", "0.714" ], [ "0.1772", "0.7236", "0.8228", "0.7236", "0.8228", "0.7373", "0.1772", "0.7373" ], [ "0.1765", "0.7363", "0.8228", "0.7363", "0.8228", "0.7505", "0.1765", "0.7505" ], [ "0.1748", "0.7524", "0.8228", "0.7524", "0.8228", "0.7666", "0.1748", "0.7666" ], [ "0.1748", "0.7676", "0.8228", "0.7676", "0.8228", "0.782", "0.1748", "0.782" ], [ "0.1772", "0.784", "0.2737", "0.784", "0.2737", "0.7954", "0.1772", "0.7954" ], [ "0.1765", "0.8047", "0.822", "0.8047", "0.822", "0.8184", "0.1765", "0.8184" ], [ "0.1765", "0.8184", "0.8228", "0.8184", "0.8228", "0.832", "0.1765", "0.832" ], [ "0.1765", "0.834", "0.8228", "0.834", "0.8228", "0.8477", "0.1765", "0.8477" ], [ "0.1765", "0.845", "0.822", "0.8477", "0.822", "0.8623", "0.1765", "0.86" ], [ "0.1765", "0.8623", "0.824", "0.8623", "0.824", "0.8755", "0.1765", "0.8755" ], [ "0.1748", "0.8755", "0.2932", "0.8755", "0.2932", "0.8896", "0.1748", "0.8896" ], [ "0.1765", "0.8965", "0.8228", "0.8965", "0.8228", "0.91", "0.1765", "0.91" ], [ "0.1765", "0.911", "0.826", "0.911", "0.826", "0.925", "0.1765", "0.925" ], [ "0.4927", "0.946", "0.5107", "0.9497", "0.505", "0.9634", "0.4878", "0.959" ] ]

[ "Published as a conference paper at ICLR 2024", "30", "0", "150", "2Cov(H*,kyy", "", "MMD-D_Kx", "2", "Mn.0", "MMD-D_kxy", "25", "MMD-D_kxy", "0", "00", "", "", "FMMD-MPOKO", "", "00t/0)", "MMD-MP1", "Var(kxx)", "Var(2kxy)", "MMD-MP_KXY", "MMD-MP_KXY", "8", "00", "0", "2Cov(kxx,-2kx)", ".kv", "MR-N", "ree", "50", "", "开用", "", "25", "75", "100", "25", "n", "25", "75", "n", "o", "Epochs", "Epochs", "75", "Epochs", "(a) E(k) ,q=l", "(b) E(k),q=3", "(c) Var (MMD-D), q=3", "2Cov(H*,ky)", "10", "Var(kyy)", "VarH", "0100", "_ Var(2kxy)", "2Cov(kxx,-2kx)", "0", "60", "i", "8", "23", "50", "75", "100", "Epoc", "(d).Var", "", "(MMD-MP)", "q=3", "Figure 2: E(k) in MMD and their variances under two optimization methods (MMD-MP", "is ours). Subfigures (a) and (b) depict the value of each E(k) in MMD during training by", "MMD-D and MMD-MP with q=1 and q=3 , respectively. Subfigures (c) and (d) illustrate", "the variances of some terms associated with MMD, i.e., og, when training by MMD-D and", "MMD-MP, respectively", "测试电源。测试能力是当PQ时拒绝零假设（o：P=Q）的概率。Ahighertestpower", "indicates a greater level of certainty regarding the distributional discrepancy.对于相当大", "的n，Liuetal.（2020）发现功率几乎与", "J(P, Q; kw) = MMD’(P,Q; kw)/αsi(P, Q; kw), where o := 4 (E[H,Hi] - E[Hi,]?)", "成正比。我们可以通过", "MIMD(Sn.Soik),where 03:=2(2H)", "J(Sp,SQ;ku)=", "-n4", "(ZZH)(1)", "Vos(Sp,So;ku)+", "用正则化估计器来估计J(P，Q；kw)", "2.2High Variance Problem of Kernel-based MMD in Multiple Populations", "实际上，我们可能会从GPT3（Brownetal.2020），ChatGPT（OpenAI，2022）等不同LLM", "生成的文本混合物中收集具有多个不同人群的训练数据Sur。由于不同的语言模型生成了", "不同的语言风格，这可能会导致生成的文本发生重大变化。在这种情况下，尽管我们可以最", "大化标准（1）来优化内核k，但存在高方差", "为了验证上述现象，我们通过最大化Eq来演示训练期间的MMD值及其方差。（1）下不", "同数量的Str种群（i.e.，）。根据图El（a）-（b），用于MMD-D的Str和Str之间的", "MMD增加，这对于MGT检测是可取的，但是其方差同时增加，这将使检测性能恶化。这", "种现象的影响随着q的增加而恶化。这表明具有较大变化的种群StT使得原始MMD的优", "化更具挑战性", "高方差导致基于内核的MMD的优化效果不佳。尽管基于内核的MMD被广泛用于识别分", "布差异，但有限的文献已经探索了其优化机制，这可能是由于其复杂的数学格式。具体来", "说，当最大化Eqn中的J时。（1），确定MMD值及其方差的个体变化具有挑战性，从而", "导致对MMD中每个术语的复杂分析。为了解决这个问题，我们将MMD的方差分解如下", "并进行实证研究，以证明图E2中训练期间每个组件的趋势及其方差。进一步的详细分析在", "第2.3节中提供", "MMD方差中的分量。我们介绍H*=kw(x,x)-kw(x,y')-kw(y,x）及其方差", "Var(E[H*J)-Var(E[ku (x,x')])-2Cov(E[kw(x,x)], E[2kw(x,y)])+Var(E[2kw (x,y)]) 。在这里,", "3" ]

{ "arxiv_id": "2402.16041", "language": "zh", "page_idx": 2 }

[ [ "0.1765", "0.0341", "0.5063", "0.0341", "0.5063", "0.0486", "0.1765", "0.0486" ], [ "0.1765", "0.1061", "0.8203", "0.1061", "0.8203", "0.1193", "0.1765", "0.1193" ], [ "0.1748", "0.1187", "0.822", "0.1193", "0.822", "0.1338", "0.1748", "0.1332" ], [ "0.1871", "0.132", "0.8228", "0.1332", "0.8228", "0.1477", "0.1871", "0.1465" ], [ "0.183", "0.1395", "0.1929", "0.1395", "0.1929", "0.1427", "0.183", "0.1427" ], [ "0.1765", "0.1477", "0.5317", "0.1477", "0.5317", "0.161", "0.1765", "0.161" ], [ "0.1748", "0.1818", "0.5605", "0.1818", "0.5605", "0.1957", "0.1748", "0.1957" ], [ "0.1765", "0.2102", "0.8228", "0.2102", "0.8228", "0.2235", "0.1765", "0.2235" ], [ "0.1765", "0.2247", "0.8228", "0.2247", "0.8228", "0.238", "0.1765", "0.238" ], [ "0.1765", "0.2393", "0.822", "0.2393", "0.822", "0.2524", "0.1765", "0.2524" ], [ "0.1765", "0.2524", "0.8228", "0.2524", "0.8228", "0.2664", "0.1765", "0.2664" ], [ "0.1748", "0.2651", "0.8228", "0.2664", "0.8228", "0.281", "0.1748", "0.2798" ], [ "0.1748", "0.279", "0.8228", "0.281", "0.8228", "0.2954", "0.1748", "0.2935" ], [ "0.1765", "0.2954", "0.8228", "0.2942", "0.8228", "0.3125", "0.1765", "0.3132" ], [ "0.1765", "0.318", "0.822", "0.318", "0.822", "0.332", "0.1765", "0.332" ], [ "0.1748", "0.332", "0.822", "0.3308", "0.822", "0.3452", "0.1748", "0.3467" ], [ "0.1765", "0.351", "0.8228", "0.351", "0.8228", "0.3643", "0.1765", "0.3643" ], [ "0.1765", "0.3662", "0.822", "0.3662", "0.822", "0.3801", "0.1765", "0.3801" ], [ "0.1772", "0.3818", "0.2712", "0.3818", "0.2712", "0.3933", "0.1772", "0.3933" ], [ "0.1765", "0.4016", "0.8228", "0.4016", "0.8228", "0.4153", "0.1765", "0.4153" ], [ "0.1748", "0.4185", "0.822", "0.4185", "0.822", "0.432", "0.1748", "0.432" ], [ "0.1748", "0.433", "0.822", "0.4343", "0.822", "0.449", "0.1748", "0.4475" ], [ "0.1765", "0.4507", "0.822", "0.4507", "0.822", "0.464", "0.1765", "0.464" ], [ "0.1765", "0.4666", "0.822", "0.4666", "0.822", "0.4797", "0.1765", "0.4797" ], [ "0.1765", "0.4797", "0.8203", "0.4797", "0.8203", "0.4944", "0.1765", "0.4944" ], [ "0.1772", "0.4968", "0.822", "0.4968", "0.822", "0.5103", "0.1772", "0.5103" ], [ "0.1765", "0.5103", "0.822", "0.5103", "0.822", "0.5244", "0.1765", "0.5244" ], [ "0.1765", "0.5264", "0.822", "0.5264", "0.822", "0.5415", "0.1765", "0.5415" ], [ "0.1724", "0.5645", "0.3677", "0.565", "0.3677", "0.58", "0.1724", "0.579" ], [ "0.1724", "0.5967", "0.353", "0.5967", "0.353", "0.61", "0.1724", "0.61" ], [ "0.1765", "0.6255", "0.822", "0.6255", "0.822", "0.6396", "0.1765", "0.6396" ], [ "0.1765", "0.6396", "0.822", "0.6396", "0.822", "0.654", "0.1765", "0.654" ], [ "0.1772", "0.655", "0.8228", "0.655", "0.8228", "0.6685", "0.1772", "0.6685" ], [ "0.1765", "0.6694", "0.2075", "0.6694", "0.2075", "0.6816", "0.1765", "0.6816" ], [ "0.1765", "0.69", "0.8228", "0.69", "0.8228", "0.704", "0.1765", "0.704" ], [ "0.1772", "0.704", "0.8228", "0.704", "0.8228", "0.7183", "0.1772", "0.7183" ], [ "0.1772", "0.7173", "0.822", "0.7173", "0.822", "0.731", "0.1772", "0.731" ], [ "0.1765", "0.732", "0.822", "0.732", "0.822", "0.7456", "0.1765", "0.7456" ], [ "0.1748", "0.744", "0.822", "0.7456", "0.822", "0.7603", "0.1748", "0.759" ], [ "0.1732", "0.759", "0.5195", "0.759", "0.5195", "0.773", "0.1732", "0.773" ], [ "0.1732", "0.7944", "0.4355", "0.7944", "0.4355", "0.808", "0.1732", "0.808" ], [ "0.1772", "0.8228", "0.8228", "0.8228", "0.8228", "0.836", "0.1772", "0.836" ], [ "0.1765", "0.8384", "0.8228", "0.8384", "0.8228", "0.8516", "0.1765", "0.8516" ], [ "0.1732", "0.8516", "0.4805", "0.851", "0.4805", "0.8657", "0.1732", "0.866" ], [ "0.3423", "0.877", "0.6562", "0.8774", "0.6562", "0.8926", "0.3423", "0.8916" ], [ "0.8013", "0.877", "0.826", "0.877", "0.826", "0.8945", "0.8013", "0.8945" ], [ "0.3325", "0.91", "0.6626", "0.911", "0.6626", "0.9253", "0.3325", "0.925" ], [ "0.8013", "0.91", "0.826", "0.91", "0.826", "0.928", "0.8013", "0.928" ], [ "0.492", "0.9487", "0.509", "0.9487", "0.509", "0.962", "0.492", "0.962" ] ]

[ "Published as a conference paper at ICLR 2024", "E表示对从MGT和HWT采样的两个种群进行期望，而Var表示对这些采样种群进行方差", "然后可以将MMD的方差分解为：Var(E[H*J)+Var(E[kw(y,y')])+2Cov(E[H*],E[kw(y,y')])", "通过这种分解，我们发现MMD方差的变化基本上来自ku(x,x），kw(x,y）和kw(y,y')", "", "的方差变化，在图E2中表示为kxx，kxy和kyy", "2.3Optimization Mechanism of Kernel-based MMD", "我们得出的结论是，在优化过程中，我们应该排除Su中的类内距离。为了阐明这一点，我", "们举例说明了一些关键的观察结果，然后解释了这些现象。观察结果：i）在图2（a）-（b）", "中，E[kw(x,x')]和E[kw(y,y')】均表现出总体上升趋势，而E[kw(x,y)]则表现出相对较小的", "变化。i）随着种群数量q的增加，E[kw(y,y)]变得比E[kw(x,x')]小，并且它们之间的间隙", "变宽。ii）在图2（c）中，MMD的方差主要由E[kw(x,x')]和E[kwy,y')】的方差决定，而", "不是其他项。iv附录）当Str中的MGT包含多个不同的种群（图2（c）中的E[kw(y,y')]", "，q=3）时，通过MMD-D优化的E[kw(y,y')]具有显着差异，以及E[kw(x,x\")]中的不同", "说明：首先，i）和ii）表明，随着q的增加，在使用高斯内核同时优化kw(x,x'）和kw(y,y）", "时，在Str（MGT）中聚合实例比在Str（HWTS）中聚合实例更具挑战性，因为同时优", "化。其次，i）和iv）表示目标J（1）以类似的方式影响有关Sfr和Str的每个项的优", "化。因此，通过相同的核函数映射后，它们的分布特征，例如kw的均值和方差，表现出相", "似的变化趋势", "此外，优化的内核函数E[kw(y,y')]不仅会a）将St中的随机成对MGT实例（y,y'）彼此", "靠近地映射，从而使映射的MGT更加均匀；而且b）在StT中执行用于聚合MGT的隐式", "“配对规则”。在整个优化过程中，共享这些规则以配对St\"中的HWT实例。当Str中的", "MGT包含不同的种群时，St对中的差异可能很大。应用配对规则可能会无意中将Str中", "的HWT实例映射到远离其中心的位置，从而导致kw（x,x'）的波动增加，从而导致更大的", "Var(E[kw(x,x')I）。同样，Str中HWT的配对规则对Str中的MGT产生负面影响，但程", "度较小，因为Sp是IID，这意味着Sp实例具有更相似的统计特征。Str中HWT的配对", "规则不需要与Str中聚合非IIDMGT的规则一样强。因此，我们将在整个优化过程中排除", "与E[kw(y,y')]关联的St中的类内距离。我们还探讨了附录G中排除E[k(x,x')]的情况", "3ProposedMethods", "3.1ProblemDefinition", "问题定义。令P为可分离度量空间XCRd和HWT分布P的IID观测Sp={x}\"-1上的", "Borel概率度量，我们旨在判断即将到来的数据SQ={yi}r-1是否来自分布P。请注意，S", "可以是由多个不同LLM生成的HWT或MGT。当m=1时，问题可以被视为单实例检测", "任务", "MGT检测的挑战。HWTs和MGT之间的区别（e.g.，来自GPT-4等LLM的文本）本", "质上很小，尤其是在诸如单个句子之类的较短文本中，因此很难区分它们。此外，LLM的", "多样性导致生成的语言风格发生重大变化，这进一步增加了MGT检测的难度。尽管深核", "MMD（MMD-D）可以有效地测量分布差异，但由多个LLM生成的具有实质性差异的文本", "在训练深核e.g.时面临挑战，MMD的高方差。MMD的这种高方差将导致分布差异的不稳", "定估计，最终导致MGT检测的性能不令人满意", "3.2MMD-MP for MGT Detection", "如前所述，我们不考虑优化S中的类内距离。取而代之的是，我们提出了一种基于代理", "MPP的基于内核的MMD（MMD-MP）的多种群感知优化，方法是最大化一个新目标，如", "Eqn。（2），并显示算法1中的训练算法。", "J(P,Q; kw) = MPP(P, Q; kw)/osr(P,Q; kw),", "(2)", "MPP(P,Q; H) := E[kw (X,X') - 2kw(X,Y)]", "(3)", "4." ]

{ "arxiv_id": "2402.16041", "language": "zh", "page_idx": 3 }

[ [ "0.1748", "0.03284", "0.5073", "0.03284", "0.5073", "0.04736", "0.1748", "0.04736" ], [ "0.1765", "0.0947", "0.4983", "0.0947", "0.4983", "0.1092", "0.1765", "0.1092" ], [ "0.4958", "0.0947", "0.8145", "0.0947", "0.8145", "0.108", "0.4958", "0.108" ], [ "0.1765", "0.1092", "0.2051", "0.1092", "0.2051", "0.1206", "0.1765", "0.1206" ], [ "0.511", "0.1111", "0.8164", "0.1111", "0.8164", "0.125", "0.511", "0.125" ], [ "0.1887", "0.1256", "0.451", "0.1256", "0.451", "0.1395", "0.1887", "0.1395" ], [ "0.513", "0.125", "0.574", "0.125", "0.574", "0.1364", "0.513", "0.1364" ], [ "0.5073", "0.1364", "0.6763", "0.1351", "0.6763", "0.1528", "0.5073", "0.154" ], [ "0.1887", "0.1427", "0.4673", "0.1433", "0.4673", "0.1578", "0.1887", "0.1572" ], [ "0.1887", "0.1572", "0.3652", "0.1572", "0.3652", "0.1718", "0.1887", "0.1718" ], [ "0.527", "0.1547", "0.806", "0.1572", "0.806", "0.1718", "0.527", "0.1692" ], [ "0.2035", "0.1674", "0.437", "0.1686", "0.437", "0.1831", "0.2035", "0.1818" ], [ "0.527", "0.1724", "0.8105", "0.1736", "0.8105", "0.1881", "0.527", "0.1875" ], [ "0.2035", "0.1843", "0.4805", "0.1849", "0.4805", "0.2026", "0.2035", "0.202" ], [ "0.511", "0.1881", "0.5605", "0.1881", "0.5605", "0.1995", "0.511", "0.1995" ], [ "0.2075", "0.1989", "0.2385", "0.2008", "0.2378", "0.2134", "0.2059", "0.2108" ], [ "0.513", "0.2008", "0.5664", "0.2008", "0.5664", "0.2122", "0.513", "0.2122" ], [ "0.589", "0.1995", "0.818", "0.1995", "0.818", "0.2134", "0.589", "0.2134" ], [ "0.1896", "0.214", "0.2378", "0.214", "0.2378", "0.226", "0.1896", "0.226" ], [ "0.5107", "0.2102", "0.584", "0.2122", "0.584", "0.2279", "0.5107", "0.226" ], [ "0.1896", "0.2267", "0.2443", "0.2267", "0.2443", "0.238", "0.1896", "0.238" ], [ "0.367", "0.2291", "0.3782", "0.2291", "0.3782", "0.238", "0.367", "0.238" ], [ "0.4395", "0.226", "0.482", "0.226", "0.482", "0.2374", "0.4395", "0.2374" ], [ "0.2239", "0.2405", "0.2632", "0.2405", "0.2632", "0.252", "0.2239", "0.252" ], [ "0.2524", "0.2361", "0.3733", "0.2374", "0.3733", "0.252", "0.2524", "0.2507" ], [ "0.1929", "0.2468", "0.2222", "0.2468", "0.2222", "0.2551", "0.1929", "0.2551" ], [ "0.1748", "0.2727", "0.5107", "0.2715", "0.5107", "0.286", "0.1748", "0.2874" ], [ "0.3628", "0.2935", "0.3694", "0.2935", "0.3694", "0.3005", "0.3628", "0.3005" ], [ "0.3962", "0.2954", "0.824", "0.2974", "0.824", "0.315", "0.3962", "0.3132" ], [ "0.3325", "0.308", "0.406", "0.3062", "0.406", "0.3208", "0.3325", "0.3225" ], [ "0.1912", "0.2974", "0.3284", "0.2986", "0.3284", "0.3132", "0.1912", "0.3125" ], [ "0.3987", "0.3164", "0.423", "0.3164", "0.423", "0.3289", "0.3987", "0.3289" ], [ "0.1732", "0.3416", "0.4224", "0.3423", "0.4224", "0.3567", "0.1732", "0.356" ], [ "0.4795", "0.3416", "0.822", "0.3423", "0.822", "0.3567", "0.4795", "0.356" ], [ "0.4795", "0.3567", "0.5063", "0.3567", "0.5063", "0.3694", "0.4795", "0.3694" ], [ "0.308", "0.363", "0.4395", "0.3662", "0.4395", "0.3806", "0.308", "0.3774" ], [ "0.179", "0.3718", "0.2852", "0.3738", "0.2852", "0.3882", "0.179", "0.3865" ], [ "0.492", "0.3726", "0.8057", "0.3726", "0.8057", "0.3865", "0.492", "0.3865" ], [ "0.3096", "0.385", "0.4558", "0.385", "0.4558", "0.4028", "0.3096", "0.4028" ], [ "0.4944", "0.3882", "0.6274", "0.3865", "0.6274", "0.3997", "0.4944", "0.4016" ], [ "0.4927", "0.3984", "0.5864", "0.401", "0.5864", "0.414", "0.4927", "0.4124" ], [ "0.601", "0.401", "0.671", "0.3984", "0.671", "0.4136", "0.601", "0.4153" ], [ "0.4436", "0.4077", "0.4631", "0.4077", "0.4631", "0.4204", "0.4436", "0.4204" ], [ "0.509", "0.4136", "0.763", "0.4136", "0.763", "0.4268", "0.509", "0.4268" ], [ "0.3896", "0.4287", "0.4468", "0.4287", "0.4468", "0.4475", "0.3896", "0.4475" ], [ "0.511", "0.4287", "0.8057", "0.4287", "0.8057", "0.4426", "0.511", "0.4426" ], [ "0.509", "0.4426", "0.7876", "0.443", "0.7876", "0.4583", "0.509", "0.457" ], [ "0.1847", "0.433", "0.2133", "0.433", "0.2133", "0.4458", "0.1847", "0.4458" ], [ "0.384", "0.4495", "0.4102", "0.4495", "0.4102", "0.4539", "0.384", "0.4539" ], [ "0.4436", "0.459", "0.464", "0.459", "0.464", "0.4717", "0.4436", "0.4717" ], [ "0.5107", "0.461", "0.774", "0.461", "0.774", "0.476", "0.5107", "0.476" ], [ "0.1748", "0.4792", "0.4631", "0.4792", "0.4631", "0.4937", "0.1748", "0.4937" ], [ "0.5376", "0.4778", "0.675", "0.4778", "0.675", "0.4912", "0.5376", "0.4912" ], [ "0.1772", "0.4937", "0.3245", "0.4937", "0.3245", "0.5083", "0.1772", "0.5083" ], [ "0.4958", "0.4893", "0.5435", "0.4893", "0.5435", "0.5005", "0.4958", "0.5005" ], [ "0.4944", "0.5024", "0.569", "0.5024", "0.569", "0.5146", "0.4944", "0.5146" ], [ "0.1765", "0.5146", "0.824", "0.5146", "0.824", "0.529", "0.1765", "0.529" ], [ "0.1765", "0.534", "0.8145", "0.534", "0.8145", "0.549", "0.1765", "0.549" ], [ "0.1772", "0.552", "0.8203", "0.552", "0.8203", "0.565", "0.1772", "0.565" ], [ "0.1765", "0.5664", "0.8228", "0.5664", "0.8228", "0.58", "0.1765", "0.58" ], [ "0.1765", "0.5835", "0.7036", "0.5835", "0.7036", "0.5967", "0.1765", "0.5967" ], [ "0.1765", "0.603", "0.824", "0.603", "0.824", "0.617", "0.1765", "0.617" ], [ "0.1732", "0.616", "0.4673", "0.617", "0.4673", "0.632", "0.1732", "0.631" ], [ "0.3743", "0.637", "0.621", "0.638", "0.621", "0.659", "0.3743", "0.6577" ], [ "0.8013", "0.64", "0.8267", "0.64", "0.8267", "0.6577", "0.8013", "0.6577" ], [ "0.1724", "0.662", "0.687", "0.662", "0.687", "0.683", "0.1724", "0.683" ], [ "0.1732", "0.688", "0.824", "0.6895", "0.824", "0.704", "0.1732", "0.7026" ], [ "0.1748", "0.704", "0.5474", "0.704", "0.5474", "0.7173", "0.1748", "0.7173" ], [ "0.33", "0.7246", "0.4043", "0.7246", "0.4043", "0.752", "0.33", "0.752" ], [ "0.4265", "0.7197", "0.5923", "0.7227", "0.5923", "0.7407", "0.4265", "0.7373" ], [ "0.614", "0.728", "0.63", "0.728", "0.63", "0.735", "0.614", "0.735" ], [ "0.8022", "0.732", "0.824", "0.732", "0.824", "0.747", "0.8022", "0.747" ], [ "0.482", "0.742", "0.5146", "0.742", "0.5146", "0.7583", "0.482", "0.7583" ], [ "0.589", "0.741", "0.652", "0.741", "0.652", "0.756", "0.589", "0.756" ], [ "0.1683", "0.757", "0.83", "0.757", "0.83", "0.7886", "0.1683", "0.7886" ], [ "0.1748", "0.784", "0.5786", "0.784", "0.5786", "0.7974", "0.1748", "0.7974" ], [ "0.1765", "0.8096", "0.8228", "0.808", "0.8228", "0.8228", "0.1765", "0.8237" ], [ "0.1765", "0.829", "0.824", "0.829", "0.824", "0.8433", "0.1765", "0.8433" ], [ "0.1748", "0.8433", "0.824", "0.8433", "0.824", "0.858", "0.1748", "0.858" ], [ "0.1732", "0.855", "0.2632", "0.857", "0.2632", "0.872", "0.1732", "0.869" ], [ "0.1748", "0.879", "0.8228", "0.879", "0.8228", "0.893", "0.1748", "0.893" ], [ "0.1871", "0.8945", "0.8228", "0.8945", "0.8228", "0.909", "0.1871", "0.909" ], [ "0.1765", "0.91", "0.478", "0.91", "0.478", "0.925", "0.1765", "0.925" ], [ "0.492", "0.9478", "0.5073", "0.9478", "0.5073", "0.962", "0.492", "0.962" ] ]

[ "Published as a conference paper at ICLR 2024", "Algorithm 2 Testing with MMD-MP for", "Algorithm 3 Testing with MMD-MP for SID", "2ST", "Input:Referenced HWT Sre,testing texts", "Input: Testing texts Ste , Ste , f, kw ;", "Ste, Ste :", "for xi,yj in Ste, ste do", "est-MMD(Ste, Ste; ku) using Eqn. (??):", "for i = 1,2,...,nperm do", "PiMMD (Sre,(xi);kw)using Eqn. (9);", "Shufle Ste u ste into Sx and Sy ;", "Qj+-MMD, (Sre, (yil;ku) using Eqn. (9);", "perm,+MMD(Sx, Sy; ku) using Eqn.", "end for", "(??);", "Output:", "AUROC value with two sets", "end for", "(P], (Q;)", "Output:", "p", "-value", "Z", "perm 1(perm,≥est)", "npern", "从经验上讲，我们可以用无偏估计器近似MPP", "1", "Hit,where Ht:=hw(xi,x))-kw(xi,yj)-ku(yi,xj).(4)", "n(n-1) 4", "MPPu(Sp, So; kw)=", "许", "此外，我们可以估计Eqn。（2)by", "Algorithm1TrainingdeepkernelwithMMD", "MP", "MPPu(Sp, SQ; kw)", "J(Se, So; kw) :", "Input: Str , Str , a frozen feature extractor f :", "63(Sp,SQ;kw)+", "w←wo;>←10-8", "for T=1,2,.", ",Tmardo", "(5)", "kw +- kernel function using Eqn. (??):", "H", "M(w) MPPu(Str, Str: kw) using Eqn. (4);", "V(w) +02. (Spr, Sur;kw) using Eqn. (6);", "3*：", "1", "(6)", "Jx(w) M(w)/ Vv(w) as in Eqn. (5);", "接下来，我们提供一些理论分析来阐述", "w+nVAdamJx(w) ;", "目标函数Eqn。（2）", "end for", "Output: ka", "与MMD（Borgwardtetal.，2006；Grettonetal.,2012）同，Eqn中的代理MPP。（4）禾", "包含与So相关的kw（y，y）。但是，MPPu仍然是U-统计量（Serfling，2009），如MMD，", "具有许多理想的统计特性，可方便地进行理论分析。请注意，尽管为内核训练最大化MPP，", "很简单，但它忽略了方差，并可能导致不稳定的差异（请参阅附录E中的更多详细信息）。", "为了解决这个问题，我们分析了MPPu的渐近性，并得出其测试能力如下", "Proposition 1.(Asymptotics of MPPu ) Under the alternative Si : P Q, based on a", "standard centrai limit theorem, we have:", "Vn(MPPu-MPP)→N(0,%:),", "(7)", "where o%: := 4 (E[Hi2Hi3] -E[Hi2]2) , Hi2, Hi3 denote diferent H,", "Corollary 1.(Test power of MPPu)For reasonably large n, the probability of rejecting the", "null hypothesis o : P= Q when P + Q is given by:", "PrMPP-", "Vn(MPP +R(SQ) -", "P", "(8)", "0si", "Vnsi", "where Pr :=Pr(n[MPPu+ R(So)]>) and R(So)=Z+ ku(vy) >0,重", "is the standard normal cumulative distribution function.", "备注2注意，我们不排除Eqn中的术语R(So）。（8）由于nMPPu的不确定收敛性（可", "能与内核kw有关），当P=Q。相反，nMMD=n[MPPu+R(So)I已被证明是收敛的（", "Grettonetal.（2012），定理12)。这使我们能够找到具有拒绝阈值的近似功率，如r（Liu", "et al., 2020)", "推论（1）表明，给定r和s*为常数，对于相当大的n，MPP的测试能力由Φ内部的第", "项控制。正如2.3节所建议的那样，当删除St（i.e.，R(SQ)>0）中的类内距离时，我", "们最后优化了Eqn。（2）用于MGT检测", "5" ]

{ "arxiv_id": "2402.16041", "language": "zh", "page_idx": 4 }

[ [ "0.1356", "0.11426", "0.839", "0.11426", "0.839", "0.1288", "0.1356", "0.1288" ], [ "0.4272", "0.1364", "0.5483", "0.1364", "0.5483", "0.154", "0.4272", "0.154" ], [ "0.4395", "0.1995", "0.5312", "0.1989", "0.5312", "0.214", "0.4395", "0.2153" ], [ "0.0866", "0.2172", "0.1299", "0.2153", "0.1316", "0.2311", "0.08826", "0.2329" ], [ "0.2418", "0.2279", "0.3203", "0.2279", "0.3203", "0.2456", "0.2418", "0.2456" ], [ "0.5024", "0.2311", "0.884", "0.2311", "0.884", "0.2443", "0.5024", "0.2443" ], [ "0.1209", "0.252", "0.4412", "0.252", "0.4412", "0.2651", "0.1209", "0.2651" ], [ "0.5034", "0.2468", "0.8857", "0.2468", "0.8857", "0.26", "0.5034", "0.26" ], [ "0.5", "0.26", "0.884", "0.2615", "0.884", "0.276", "0.5", "0.2747" ], [ "0.1209", "0.2676", "0.4412", "0.2676", "0.4412", "0.281", "0.1209", "0.281" ], [ "0.5024", "0.2766", "0.8857", "0.2766", "0.8857", "0.2903", "0.5024", "0.2903" ], [ "0.1217", "0.2827", "0.4412", "0.2827", "0.4412", "0.2966", "0.1217", "0.2966" ], [ "0.11926", "0.2966", "0.4126", "0.2966", "0.4126", "0.31", "0.11926", "0.31" ], [ "0.501", "0.2922", "0.6274", "0.2922", "0.6274", "0.3062", "0.501", "0.3062" ], [ "0.407", "0.2986", "0.4412", "0.2986", "0.4412", "0.31", "0.407", "0.31" ], [ "0.1209", "0.3132", "0.4412", "0.3132", "0.4412", "0.327", "0.1209", "0.327" ], [ "0.5024", "0.3145", "0.8857", "0.3145", "0.8857", "0.3276", "0.5024", "0.3276" ], [ "0.11926", "0.327", "0.4395", "0.327", "0.4395", "0.3403", "0.11926", "0.3403" ], [ "0.5024", "0.33", "0.8857", "0.33", "0.8857", "0.3435", "0.5024", "0.3435" ], [ "0.1209", "0.3435", "0.4412", "0.3435", "0.4412", "0.3567", "0.1209", "0.3567" ], [ "0.501", "0.3447", "0.8857", "0.3447", "0.8857", "0.358", "0.501", "0.358" ], [ "0.11926", "0.3567", "0.4412", "0.358", "0.4412", "0.3726", "0.11926", "0.3718" ], [ "0.501", "0.3591", "0.886", "0.3591", "0.886", "0.3738", "0.501", "0.3738" ], [ "0.11926", "0.3738", "0.4412", "0.3738", "0.4412", "0.387", "0.11926", "0.387" ], [ "0.501", "0.375", "0.886", "0.375", "0.886", "0.3882", "0.501", "0.3882" ], [ "0.1209", "0.3882", "0.4429", "0.3882", "0.4429", "0.4028", "0.1209", "0.4028" ], [ "0.501", "0.3901", "0.8857", "0.3901", "0.8857", "0.404", "0.501", "0.404" ], [ "0.11926", "0.404", "0.4412", "0.4028", "0.4412", "0.4167", "0.11926", "0.4172" ], [ "0.5024", "0.4048", "0.8857", "0.4048", "0.8857", "0.4192", "0.5024", "0.4192" ], [ "0.1209", "0.4185", "0.4429", "0.4185", "0.4429", "0.432", "0.1209", "0.432" ], [ "0.5", "0.4185", "0.8857", "0.4192", "0.8857", "0.4343", "0.5", "0.433" ], [ "0.1217", "0.435", "0.4395", "0.435", "0.4395", "0.4458", "0.1217", "0.4458" ], [ "0.5024", "0.435", "0.884", "0.435", "0.884", "0.449", "0.5024", "0.449" ], [ "0.1209", "0.449", "0.4429", "0.449", "0.4429", "0.4622", "0.1209", "0.4622" ], [ "0.501", "0.4507", "0.8857", "0.4507", "0.8857", "0.464", "0.501", "0.464" ], [ "0.11926", "0.4622", "0.4412", "0.4634", "0.4412", "0.4778", "0.11926", "0.4766" ], [ "0.501", "0.4666", "0.8857", "0.4666", "0.8857", "0.4797", "0.501", "0.4797" ], [ "0.1209", "0.4792", "0.4412", "0.4792", "0.4412", "0.4924", "0.1209", "0.4924" ], [ "0.5", "0.481", "0.884", "0.481", "0.884", "0.4944", "0.5", "0.4944" ], [ "0.1209", "0.4956", "0.4387", "0.4956", "0.4387", "0.506", "0.1209", "0.506" ], [ "0.501", "0.4956", "0.8857", "0.4956", "0.8857", "0.509", "0.501", "0.509" ], [ "0.1209", "0.509", "0.4412", "0.509", "0.4412", "0.523", "0.1209", "0.523" ], [ "0.5", "0.509", "0.8857", "0.5103", "0.8857", "0.5244", "0.5", "0.5234" ], [ "0.2532", "0.5234", "0.4412", "0.5234", "0.4412", "0.537", "0.2532", "0.537" ], [ "0.501", "0.5244", "0.886", "0.5244", "0.886", "0.539", "0.501", "0.539" ], [ "0.11926", "0.5386", "0.4429", "0.539", "0.4429", "0.5537", "0.11926", "0.553" ], [ "0.1217", "0.526", "0.2336", "0.526", "0.2336", "0.536", "0.1217", "0.536" ], [ "0.501", "0.5415", "0.886", "0.5415", "0.886", "0.5547", "0.501", "0.5547" ], [ "0.1209", "0.5547", "0.4412", "0.5547", "0.4412", "0.5684", "0.1209", "0.5684" ], [ "0.5", "0.556", "0.8857", "0.556", "0.8857", "0.5693", "0.5", "0.5693" ], [ "0.1217", "0.5693", "0.4395", "0.5693", "0.4395", "0.58", "0.1217", "0.58" ], [ "0.501", "0.5713", "0.8774", "0.5713", "0.8774", "0.5854", "0.501", "0.5854" ], [ "0.11926", "0.5835", "0.4429", "0.584", "0.4429", "0.5986", "0.11926", "0.5977" ], [ "0.5", "0.5854", "0.884", "0.5854", "0.884", "0.5986", "0.5", "0.5986" ], [ "0.1209", "0.5996", "0.4412", "0.5996", "0.4412", "0.613", "0.1209", "0.613" ], [ "0.501", "0.601", "0.8857", "0.601", "0.8857", "0.6143", "0.501", "0.6143" ], [ "0.11926", "0.6143", "0.1871", "0.6143", "0.1871", "0.629", "0.11926", "0.629" ], [ "0.5", "0.616", "0.8857", "0.6157", "0.8857", "0.63", "0.5", "0.631" ], [ "0.501", "0.632", "0.8857", "0.632", "0.8857", "0.646", "0.501", "0.646" ], [ "0.501", "0.6465", "0.8857", "0.6465", "0.8857", "0.66", "0.501", "0.66" ], [ "0.08826", "0.657", "0.2222", "0.6577", "0.2222", "0.6724", "0.08826", "0.672" ], [ "0.501", "0.662", "0.8857", "0.662", "0.8857", "0.6753", "0.501", "0.6753" ], [ "0.501", "0.677", "0.8857", "0.677", "0.8857", "0.69", "0.501", "0.69" ], [ "0.08905", "0.6846", "0.4739", "0.6846", "0.4739", "0.6987", "0.08905", "0.6987" ], [ "0.5024", "0.6934", "0.884", "0.6934", "0.884", "0.704", "0.5024", "0.704" ], [ "0.09235", "0.702", "0.4714", "0.702", "0.4714", "0.712", "0.09235", "0.712" ], [ "0.4983", "0.7056", "0.8857", "0.7056", "0.8857", "0.7197", "0.4983", "0.7197" ], [ "0.0907", "0.7153", "0.4722", "0.7153", "0.4722", "0.729", "0.0907", "0.729" ], [ "0.5", "0.7217", "0.884", "0.7217", "0.884", "0.7363", "0.5", "0.7363" ], [ "0.0907", "0.73", "0.4722", "0.73", "0.4722", "0.7437", "0.0907", "0.7437" ], [ "0.4983", "0.7363", "0.884", "0.7363", "0.884", "0.75", "0.4983", "0.75" ], [ "0.0907", "0.744", "0.4697", "0.744", "0.4697", "0.7583", "0.0907", "0.7583" ], [ "0.4983", "0.7505", "0.8857", "0.752", "0.8857", "0.7666", "0.4983", "0.765" ], [ "0.0907", "0.7603", "0.4739", "0.7603", "0.4739", "0.7734", "0.0907", "0.7734" ], [ "0.5", "0.7676", "0.8857", "0.7676", "0.8857", "0.781", "0.5", "0.781" ], [ "0.08905", "0.7744", "0.4722", "0.7744", "0.4722", "0.7886", "0.08905", "0.7886" ], [ "0.501", "0.7827", "0.8857", "0.7827", "0.8857", "0.797", "0.501", "0.797" ], [ "0.0907", "0.7905", "0.4756", "0.7905", "0.4756", "0.8037", "0.0907", "0.8037" ], [ "0.5", "0.797", "0.884", "0.7974", "0.884", "0.8125", "0.5", "0.811" ], [ "0.08905", "0.806", "0.4722", "0.806", "0.4722", "0.8193", "0.08905", "0.8193" ], [ "0.501", "0.813", "0.884", "0.813", "0.884", "0.827", "0.501", "0.827" ], [ "0.1095", "0.8276", "0.4592", "0.8276", "0.4592", "0.8413", "0.1095", "0.8413" ], [ "0.5", "0.827", "0.886", "0.8276", "0.886", "0.8423", "0.5", "0.8413" ], [ "0.08905", "0.8433", "0.4468", "0.8433", "0.4468", "0.854", "0.08905", "0.854" ], [ "0.5", "0.8433", "0.886", "0.8433", "0.886", "0.857", "0.5", "0.857" ], [ "0.5", "0.857", "0.8857", "0.856", "0.8857", "0.8706", "0.5", "0.872" ], [ "0.08496", "0.863", "0.4756", "0.8643", "0.4756", "0.882", "0.08496", "0.881" ], [ "0.5063", "0.8774", "0.5156", "0.8774", "0.5156", "0.8833", "0.5063", "0.8833" ], [ "0.5146", "0.8726", "0.886", "0.8726", "0.886", "0.886", "0.5146", "0.886" ], [ "0.08905", "0.88", "0.4739", "0.88", "0.4739", "0.893", "0.08905", "0.893" ], [ "0.5", "0.886", "0.8", "0.887", "0.8", "0.9014", "0.5", "0.901" ], [ "0.08826", "0.8926", "0.2222", "0.8926", "0.2222", "0.906", "0.08826", "0.906" ], [ "0.482", "0.925", "0.4927", "0.925", "0.4927", "0.9346", "0.482", "0.9346" ] ]

[ "Wrapped Cauchy Distributed Angular Softmax for Long-Tailed Visual", "Recognition", "Boran Han!", "。3in", "Abstract", "布：少数高频类包含许多训练示例，而许多低频类的", "Addressing imbalanced or long-tailed data is", "训练示例不足。这种情况被称为长尾识别（Liuetal.", "2019），用这样的数据集训练的标准方法往往不会产", "a major challenge in visual recognition tasks", "生与平衡的(Liu et al.,2019;Lin et al.,2017;Cui et al.,", "due to disparities between training and testing", "distributions and issues with data noise.", "2021)相同的性能", "We", "propose the Wrapped Cauchy Distributed", "许多研究都集中在长尾识别上，试图通过类平衡采样", "Angular Softmax (WCDAS), a novel softmax", "或类重新加权（Hanetal.，2005;Kangetal.,2020Kubat", "function that incorporates data-wise Gaussian-", "& Matwin, 1997; Huang et al.,2016; 2020; Hong et al.", "based kernels into the angular correlation", "2023）来重新平衡数据分布。但是，在优化（Huang", "between feature representations and classifier", "etal.,2016;2020）期间，它们可能无法代表多数类（Han", "weights, effectively mitigating noise and sparse", "et al.,2005;Kang etal.,2020;Kubat &Matwin,1997)或", "sampling concerns. The class-wise distribution", "使网络不稳定。。Inadditiontodirect sampling，focal", "of angular representation becomes a sum of", "loss (Lin et al., 2017) adopts loss function emphasizing", "these kernels.Our theoretical analysis reveals", "samples with larger loss value.However, it inevitably", "that the wrapped Cauchy distribution excels", "involves hyperparameters tuning by cross-validation. An", "the Gaussian distribution in approximating", "alternative method is to adopt a label-aware correction", "mixed distributions.Additionally, WCDAS", "via introducing a class-wise generalization error bound", "usestrainableconcentrationparametersto", "such as Label-Distribution-Aware Margin Loss (LDAM)", "dynamically adjust the compactness and margin", "(Cao et al., 2019) and Balanced Meta-Softmax (BALMS)", "Empirical results confirm", "(Ren et al., 2020).Cao, et.al.have proved that to", "label-aware behavior in these parameters and", "of each class.", "improve the accuracy in recognizing long-tailed distributed", "demonstrate WCDAS's superiority over other", "data, classes with fewer training examples should have", "state-of-the-artsoftmax-basedmethodsin", "a higher generalization error bound (Cao et al., 2019)", "handling long-tailed visual recognition across", "However, both LDAM and BALMS can be vulnerable", "multiple benchmark datasets. The code is public", "when the number of examples per class is unknown and", " available.", "constantly changing.Therefore, further corrections are", "required for continuous training:Meta-Weight-Net (Shu", "et al., 2019) and Equilibrium loss (Feng et al., 2021) are", "1.Introduction", "developed for class re-weighting and inter-class margin", "correction, which require no visibility to the underlying", "深度卷积神经网络是计算机视觉任务的主要方法，包", "datadistributionHowever,thosemethods can either", "括视觉识别。这种优势在很大程度上归功于它们强大", "be subject to lengthy training time due to the nature of", "的表示学习，这种技术可以将目标图像简化为具有较", "meta-learning (Shu et al., 2019) or high space complexity", "少维数的向量空间。倒数第二层促进了这一关键步骤", "because of the memory module (Feng et al., 2021). Lastly,", "随后将其馈送到最终分类器中，然后是softmax函数", "using angular softmax, Kobayashi has proposed applying", "该函数计算输入图像在i类中的概率：P(y=」α)", "von Mises-Fisher distribution for compact feature space via", "(Bridle，1989;Goodfellowetal，2016）。然而，大多数图", "a user-defined concentration parameter （k )(Kobayashi,", "像识别任务已经在平衡良好的数据集上得到了证明。。", "2021):However, such a method leads to lengthy", "相比之下，大多数现实世界的数据都带有不平衡的分", "hyper-parameter tuning with isotopic k for all classes.", "I Amazon Web Services, AI. Work done while at Shell.", "Meanwhile, their trainable class-wise k approach shows", "Correspondence to: Boran Han < [email protected] >.", "inferior performance compared with the user-defined", "counterpart for an optimal performance (Kobayashi, 2021)", "Proceedings of the 4oth International Conference on Machine", "", "此外，数据噪声也存在于长尾问题（TongWu&Lin,", "Learning, Honolulu, Hawai, USA. PMLR 202, 2023. Copyright", "2021; Cao et al., 2021; Zhang et al., 2023)中", "2023 by the author(s).", "1" ]

{ "arxiv_id": "2305.18732", "language": "zh", "page_idx": 0 }

[ [ "0.2263", "0.05807", "0.746", "0.05807", "0.746", "0.07135", "0.2263", "0.07135" ], [ "0.0907", "0.0884", "0.4722", "0.0884", "0.4722", "0.0985", "0.0907", "0.0985" ], [ "0.4983", "0.0852", "0.8857", "0.0871", "0.8857", "0.101", "0.4983", "0.0985" ], [ "0.08905", "0.1016", "0.4739", "0.1016", "0.4739", "0.11554", "0.08905", "0.11554" ], [ "0.5", "0.1016", "0.8823", "0.1016", "0.8823", "0.11554", "0.5", "0.11554" ], [ "0.0907", "0.11743", "0.4756", "0.11743", "0.4756", "0.1307", "0.0907", "0.1307" ], [ "0.5024", "0.11743", "0.8823", "0.11743", "0.8823", "0.1307", "0.5024", "0.1307" ], [ "0.09235", "0.1338", "0.4722", "0.1338", "0.4722", "0.1445", "0.09235", "0.1445" ], [ "0.5", "0.1307", "0.884", "0.132", "0.884", "0.1459", "0.5", "0.1445" ], [ "0.0907", "0.1465", "0.4756", "0.1465", "0.4756", "0.1604", "0.0907", "0.1604" ], [ "0.501", "0.1477", "0.6274", "0.1477", "0.6274", "0.161", "0.501", "0.161" ], [ "0.0907", "0.1622", "0.4739", "0.1622", "0.4739", "0.1755", "0.0907", "0.1755" ], [ "0.5024", "0.1718", "0.8813", "0.1718", "0.8813", "0.1818", "0.5024", "0.1818" ], [ "0.0907", "0.178", "0.4739", "0.178", "0.4739", "0.1913", "0.0907", "0.1913" ], [ "0.501", "0.1849", "0.884", "0.1849", "0.884", "0.1989", "0.501", "0.1989" ], [ "0.09235", "0.1925", "0.4739", "0.1925", "0.4739", "0.2058", "0.09235", "0.2058" ], [ "0.08905", "0.207", "0.4722", "0.2058", "0.4722", "0.2203", "0.08905", "0.2216" ], [ "0.501", "0.202", "0.8823", "0.202", "0.8823", "0.2122", "0.501", "0.2122" ], [ "0.5", "0.214", "0.8857", "0.2134", "0.8857", "0.2267", "0.5", "0.2279" ], [ "0.0907", "0.2235", "0.12335", "0.2235", "0.12335", "0.2349", "0.0907", "0.2349" ], [ "0.5024", "0.2299", "0.8857", "0.2299", "0.8857", "0.2437", "0.5024", "0.2437" ], [ "0.09235", "0.2468", "0.4722", "0.2468", "0.4722", "0.2568", "0.09235", "0.2568" ], [ "0.501", "0.2456", "0.884", "0.2456", "0.884", "0.2588", "0.501", "0.2588" ], [ "0.08905", "0.26", "0.4722", "0.26", "0.4722", "0.2734", "0.08905", "0.2734" ], [ "0.5024", "0.262", "0.8823", "0.262", "0.8823", "0.2727", "0.5024", "0.2727" ], [ "0.0907", "0.276", "0.4739", "0.276", "0.4739", "0.289", "0.0907", "0.289" ], [ "0.501", "0.276", "0.884", "0.276", "0.884", "0.289", "0.501", "0.289" ], [ "0.0907", "0.2922", "0.4714", "0.2922", "0.4714", "0.303", "0.0907", "0.303" ], [ "0.5034", "0.2922", "0.8823", "0.2922", "0.8823", "0.303", "0.5034", "0.303" ], [ "0.08905", "0.3062", "0.4756", "0.3062", "0.4756", "0.3193", "0.08905", "0.3193" ], [ "0.501", "0.3062", "0.8857", "0.3062", "0.8857", "0.3193", "0.501", "0.3193" ], [ "0.08905", "0.3225", "0.4558", "0.3225", "0.4558", "0.3333", "0.08905", "0.3333" ], [ "0.5034", "0.3208", "0.8857", "0.3208", "0.8857", "0.334", "0.5034", "0.334" ], [ "0.501", "0.3352", "0.884", "0.336", "0.884", "0.3499", "0.501", "0.3484" ], [ "0.0866", "0.3547", "0.2336", "0.3547", "0.2336", "0.3687", "0.0866", "0.3687" ], [ "0.5024", "0.351", "0.8857", "0.351", "0.8857", "0.3643", "0.5024", "0.3643" ], [ "0.5034", "0.3674", "0.8813", "0.3674", "0.8813", "0.3774", "0.5034", "0.3774" ], [ "0.08905", "0.3806", "0.4739", "0.3806", "0.4739", "0.3945", "0.08905", "0.3945" ], [ "0.5", "0.3801", "0.884", "0.3806", "0.884", "0.3958", "0.5", "0.3945" ], [ "0.08905", "0.3965", "0.4722", "0.3965", "0.4722", "0.4104", "0.08905", "0.4104" ], [ "0.5024", "0.3977", "0.6006", "0.3977", "0.6006", "0.4092", "0.5024", "0.4092" ], [ "0.08905", "0.411", "0.4739", "0.411", "0.4739", "0.4248", "0.08905", "0.4248" ], [ "0.5024", "0.4204", "0.884", "0.4204", "0.884", "0.4312", "0.5024", "0.4312" ], [ "0.0907", "0.4268", "0.4722", "0.4268", "0.4722", "0.44", "0.0907", "0.44" ], [ "0.5024", "0.4343", "0.8823", "0.4343", "0.8823", "0.4475", "0.5024", "0.4475" ], [ "0.08905", "0.4414", "0.4722", "0.44", "0.4722", "0.4539", "0.08905", "0.455" ], [ "0.5024", "0.4507", "0.884", "0.4507", "0.884", "0.461", "0.5024", "0.461" ], [ "0.08905", "0.457", "0.4739", "0.457", "0.4739", "0.4702", "0.08905", "0.4702" ], [ "0.501", "0.464", "0.7295", "0.464", "0.7295", "0.4778", "0.501", "0.4778" ], [ "0.08826", "0.4717", "0.4722", "0.4702", "0.4722", "0.4849", "0.08826", "0.486" ], [ "0.08905", "0.4873", "0.4722", "0.4873", "0.4722", "0.5005", "0.08905", "0.5005" ], [ "0.0907", "0.502", "0.4739", "0.502", "0.4739", "0.515", "0.0907", "0.515" ], [ "0.501", "0.5", "0.847", "0.5", "0.847", "0.5146", "0.501", "0.5146" ], [ "0.0907", "0.5176", "0.4739", "0.5176", "0.4739", "0.531", "0.0907", "0.531" ], [ "0.5186", "0.515", "0.6807", "0.5166", "0.6807", "0.531", "0.5186", "0.53" ], [ "0.09235", "0.534", "0.464", "0.534", "0.464", "0.5444", "0.09235", "0.5444" ], [ "0.08905", "0.5474", "0.4739", "0.5474", "0.4739", "0.561", "0.08905", "0.561" ], [ "0.5024", "0.5425", "0.884", "0.5425", "0.884", "0.556", "0.5024", "0.556" ], [ "0.08826", "0.5625", "0.1299", "0.5625", "0.1299", "0.577", "0.08826", "0.577" ], [ "0.5024", "0.5566", "0.884", "0.5566", "0.884", "0.571", "0.5024", "0.571" ], [ "0.5024", "0.5728", "0.884", "0.5728", "0.884", "0.586", "0.5024", "0.586" ], [ "0.0907", "0.5854", "0.4739", "0.5854", "0.4739", "0.5986", "0.0907", "0.5986" ], [ "0.5024", "0.587", "0.886", "0.587", "0.886", "0.601", "0.5024", "0.601" ], [ "0.09235", "0.601", "0.4739", "0.601", "0.4739", "0.6143", "0.09235", "0.6143" ], [ "0.5024", "0.603", "0.8857", "0.603", "0.8857", "0.616", "0.5024", "0.616" ], [ "0.0907", "0.6157", "0.4739", "0.6157", "0.4739", "0.629", "0.0907", "0.629" ], [ "0.5024", "0.617", "0.709", "0.617", "0.709", "0.631", "0.5024", "0.631" ], [ "0.09235", "0.632", "0.4714", "0.632", "0.4714", "0.6426", "0.09235", "0.6426" ], [ "0.08905", "0.646", "0.4697", "0.646", "0.4697", "0.659", "0.08905", "0.659" ], [ "0.501", "0.64", "0.8857", "0.64", "0.8857", "0.654", "0.501", "0.654" ], [ "0.5", "0.655", "0.884", "0.655", "0.884", "0.6685", "0.5", "0.6685" ], [ "0.09235", "0.662", "0.4714", "0.662", "0.4714", "0.6724", "0.09235", "0.6724" ], [ "0.09235", "0.6753", "0.4722", "0.6753", "0.4722", "0.6895", "0.09235", "0.6895" ], [ "0.501", "0.6704", "0.8857", "0.6704", "0.8857", "0.6846", "0.501", "0.6846" ], [ "0.09235", "0.6924", "0.4714", "0.6924", "0.4714", "0.7026", "0.09235", "0.7026" ], [ "0.5024", "0.6875", "0.884", "0.6875", "0.884", "0.6978", "0.5024", "0.6978" ], [ "0.09314", "0.7056", "0.4722", "0.7056", "0.4722", "0.7197", "0.09314", "0.7197" ], [ "0.5034", "0.7026", "0.8813", "0.7026", "0.8813", "0.7134", "0.5034", "0.7134" ], [ "0.09235", "0.7217", "0.4739", "0.7217", "0.4739", "0.735", "0.09235", "0.735" ], [ "0.5024", "0.7163", "0.884", "0.7163", "0.884", "0.73", "0.5024", "0.73" ], [ "0.5", "0.729", "0.8857", "0.73", "0.8857", "0.744", "0.5", "0.7437" ], [ "0.08905", "0.7363", "0.4722", "0.7363", "0.4722", "0.75", "0.08905", "0.75" ], [ "0.08905", "0.7505", "0.4722", "0.7505", "0.4722", "0.7646", "0.08905", "0.7646" ], [ "0.501", "0.7456", "0.8857", "0.7456", "0.8857", "0.759", "0.501", "0.759" ], [ "0.0907", "0.7666", "0.4739", "0.7666", "0.4739", "0.78", "0.0907", "0.78" ], [ "0.501", "0.761", "0.884", "0.761", "0.884", "0.7744", "0.501", "0.7744" ], [ "0.501", "0.7764", "0.8857", "0.7764", "0.8857", "0.7905", "0.501", "0.7905" ], [ "0.09235", "0.7827", "0.4722", "0.7827", "0.4722", "0.7935", "0.09235", "0.7935" ], [ "0.0907", "0.797", "0.4722", "0.797", "0.4722", "0.81", "0.0907", "0.81" ], [ "0.5024", "0.7915", "0.8857", "0.7915", "0.8857", "0.8047", "0.5024", "0.8047" ], [ "0.0907", "0.811", "0.4739", "0.811", "0.4739", "0.8247", "0.0907", "0.8247" ], [ "0.5024", "0.808", "0.8823", "0.808", "0.8823", "0.8184", "0.5024", "0.8184" ], [ "0.0907", "0.827", "0.4722", "0.827", "0.4722", "0.8403", "0.0907", "0.8403" ], [ "0.5024", "0.8213", "0.8857", "0.8213", "0.8857", "0.8354", "0.5024", "0.8354" ], [ "0.09235", "0.8433", "0.4697", "0.8433", "0.4697", "0.854", "0.09235", "0.854" ], [ "0.5034", "0.8384", "0.8813", "0.8384", "0.8813", "0.8486", "0.5034", "0.8486" ], [ "0.5024", "0.8516", "0.8823", "0.8516", "0.8823", "0.8657", "0.5024", "0.8657" ], [ "0.0907", "0.8594", "0.4714", "0.8594", "0.4714", "0.869", "0.0907", "0.869" ], [ "0.08905", "0.8726", "0.4739", "0.8726", "0.4739", "0.886", "0.08905", "0.886" ], [ "0.501", "0.867", "0.8823", "0.867", "0.8823", "0.881", "0.501", "0.881" ], [ "0.08905", "0.887", "0.4739", "0.887", "0.4739", "0.9014", "0.08905", "0.9014" ], [ "0.501", "0.8833", "0.884", "0.8833", "0.884", "0.8965", "0.501", "0.8965" ], [ "0.478", "0.925", "0.492", "0.922", "0.4968", "0.933", "0.4836", "0.9375" ] ]

[ "Wrapped Cauchy Distributed Angular Softmax for Long-Tailed Visual Recognition", "鉴于这些挑战我们提出了用于长尾", "一个简单而有效的解决方案，可以显着改善长尾数据", "视觉识别basedon（Kobayashi，2021）的", "集上的泛化问题。在此两阶段训练期间，表示学习由", "WrappedCauchyDistributedAngularSoftmax(WCDAS)。", "实例平衡采样器（Kangetal.,2020）训练，而分类器由", "我们假设数据式概率分布遵循包裹正态分布，并推断", "类平衡采样器（Kangetal.，2020）和元采样器（Renetal.", "WCDAS可以更好地适合由单个分布组成的混合分布。", "2020)进一步微调", "我们还证明了WCDAS具有几个理想的功能，例如通", "特征分布的参数化建模。尽管深度学习的出现归因于", "过浓度参数对类之间的边距进行自适应调节，从而表", "非参数非线性建模，但在处理某些现实世界的数据集", "现出标签感知行为。在对几个基准长尾图像分类数据", "集进行评估后，WCDAS优于最先进的基于softmax的", "时，有效地训练网络可以证明是具有挑战性的，这些数", "据集存在诸如类不平衡和示例不足等问题。基于某些", "方法", "假设的参数化建模可以在这些不利情况下极大地帮助", "总而言之，我们的贡献包括：1）提出了一个模型，该", "学习（Yangetal.，2021;Hayatetal.2019）。一种这样的", "模型以数据包裹的正态分布内核的形式考虑了噪声引", "方法涉及在少射学习中近似特征表示的高斯分布，以", "起的不确定性；2）证明WCDAS可以更有效地拟合此", "增强（Yangetal.，2021）的通用性。在不平衡类的情况", "类内核的混合分布：3）表明，在特定条件下，我们的", "下，研究表明，特征表示的高斯分布（Hayatetal.，2019）", "方法还显着增强了类间余量，从而产生了紧凑的聚类；", "和vonMises-Fisher分布（Kobayashi，2021）建模或权重", "4）证明浓度参数可以是自适应的，具有更高的浓度", "与特征之间的角度可以显着提高性能。特征空间的参", "数化建模也可以更好地处理数据中噪声引起的不确定", "2. Related works", "性。利用参数模型来解决不确定性的流行方法包括变", "分自动编码器（Kingma&Welling，2013），基于贝叶", "基于角度的Softmax。最近提出了Angularsoftmax（Liu", "斯的dropout（Gal&Ghahramani，2016）和DUL（Chang", "etal.,2016)及其突变方法(Dengetal.,2019;Liuetal.", "etal.,2020)等", "2017)来改善人脸验证任务中的softmax损失。与传统", "受这三种不同方法的启发，我们提出了一种对特征表", "的softmax不同，这些方法允许神经网络通过关注分", "示进行参数化建模的方法。此方法使用数据方面的高", "类器权重和特征之间的余弦相似性来以角度方式学习", "斯内核作为基础，它包括可训练的类参数，为各种类", "特征。其中，大裕度softmax（Liuetal.，2016）直接在点", "型的数据提供了可适应的框架。”", "积相似性上强制执行类间可分性，而SphereFace（Liu", "etal.,2017）和ArcFace（Dengetal.，2019）分别在超球面", "流形上强制执行乘法和加法角裕度。这些边距由超参", "3.Wrapped Cauchy Distributed Angular", "数㎡控制：m的值越大，边距越大。因此，类之间更", "Softmax (WCDAS)", "大的边距可以导致紧的集群，从而比传统的softma", "提高性能。(Liuet al.,2017;Deng et al.,2019;Liu et al.,", "以前的知识。对于角度softmax，在给定样本向量a和", "2016)", "加权向量W的情况下，来自CNN中第j类的线性分", "类器的预测概率表示为：。为了便于编写，我们将α", "长尾识别。具有长尾分布的数据集（Liuetal.，2019）不", "和W之间的角度表示（O）称为“角度特征”。SER", "仅在每个类的示例数量方面具有不平衡的类，而且还", "是经验定义的常数（Dengetal.，2019;Liuetal.,2017）或", "具有只有几个示例的长尾类（<10），i。e。，尾类。解", "可训练参数（Kobayashi,2021)", "决此类问题的两种主要方法是（1）损失函数改进和", "(2）数据重新平衡。前一种方法在尾部类（Linetal.", "WCDAS的直觉和概述。角度softmax函数（方程?？）", "的概率函数（f(O;））描述了表示特征与分类器权重", "2017：JingruTan，2020:Cuietal..2019）中利用了积极", "的学习，或者在类之间强加了很大的余量，尤其是尾", "之间的角度。因此，给定coso；，分类器权重被优化以", "部类(Caoetal..2019:Renetal.,2020:Yeetal.,2020)", "最小化损失函数。但是，这种方法可能会导致过度拟", "。特别是Caoet。al。(Cao etal.，2019）从理论上证明", "合，无其是在使用一些示例进行训练时，如先前基于", "了可以通过增加尾类的边距来最小化泛化误差界。In", "大余量的余弦softmax研究（Kobayashi，2021;Liuetal.", "2016)中所述，或使用其他研究（TongWu&Lin,2021;", "addition to margin correction, Feng et. al. also balances the", "classification via a Feature Memory Module (Feng et al..", "Caoetal.，2021;Zhangetal.2023）报告的包含噪声的数", "2021）：与此同时，少数研究集中在训练过程中的数据", "据。为了解决这些问题，我们的方法寻求，的最佳参", "数概率密度函数，条件是y=j，i。e。,P(「y=j)。", "再平衡，不平衡训练的第二种方法。可以通过数据重新", "采样(Han et al.,2005;Kang et al.,2020;Kubat &Matwin,", "为了实现这一目标，我们最初建议使用基于数据的高", "1997）或类重新加权（Huangetal.2016；2020）来实现", "斯内核作为基础。直观地：给定e：这样的内核可以", "数据重新平衡。然而，基于数据再平衡的策略可能会", "建模由输人噪声或稀疏采样引起的数据不确定性，而", "导致尾部类过拟合，以及对过度代表性的策略的学习", "不是直接的类分布（第3.1节）。通过这样做，我们可", "以通过对单个基础求和来获得类角度特征概率密度函", "效率较低。Thesamplingstrategiesincludefixedsamplers", "(Kang et al., 2020) and meta-based samplers (Ren et al.,", "数（第3.1节）。随后，我们证明了这种按类分布可以", "2020Shuetal.,2019).解耦训练（Kangetal.,2020）是", "通过WrappedCauchy分布f（p,；i）更准确地近似，并", "2" ]

{ "arxiv_id": "2305.18732", "language": "zh", "page_idx": 1 }

[ [ "0.2255", "0.05682", "0.7485", "0.05682", "0.7485", "0.0707", "0.2255", "0.0707" ], [ "0.1259", "0.0884", "0.1569", "0.0884", "0.1569", "0.09973", "0.1259", "0.09973" ], [ "0.501", "0.0859", "0.7974", "0.0859", "0.7974", "0.09973", "0.501", "0.09973" ], [ "0.3325", "0.11237", "0.3489", "0.11554", "0.3447", "0.1256", "0.3284", "0.1225" ], [ "0.7017", "0.108", "0.7256", "0.1111", "0.721", "0.1256", "0.6978", "0.1225" ], [ "0.3464", "0.1193", "0.3555", "0.1193", "0.3555", "0.125", "0.3464", "0.125" ], [ "0.6763", "0.1193", "0.6914", "0.1193", "0.6914", "0.13", "0.6763", "0.13" ], [ "0.697", "0.1206", "0.814", "0.1256", "0.812", "0.1445", "0.6953", "0.1395" ], [ "0.576", "0.127", "0.665", "0.127", "0.665", "0.1414", "0.576", "0.1414" ], [ "0.6685", "0.1276", "0.7305", "0.1427", "0.7236", "0.159", "0.661", "0.1445" ], [ "0.863", "0.127", "0.886", "0.127", "0.886", "0.1401", "0.863", "0.1401" ], [ "0.38", "0.1401", "0.4656", "0.1401", "0.4656", "0.1509", "0.38", "0.1509" ], [ "0.2115", "0.1528", "0.2876", "0.154", "0.2876", "0.1654", "0.2115", "0.1642" ], [ "0.3816", "0.1515", "0.4436", "0.1515", "0.4436", "0.159", "0.3816", "0.159" ], [ "0.3308", "0.1547", "0.3594", "0.1547", "0.3594", "0.1666", "0.3308", "0.1666" ], [ "0.0972", "0.154", "0.1953", "0.154", "0.1953", "0.1654", "0.0972", "0.1654" ], [ "0.3816", "0.1578", "0.4495", "0.1578", "0.4495", "0.1686", "0.3816", "0.1686" ], [ "0.282", "0.1692", "0.3987", "0.1704", "0.3987", "0.1831", "0.282", "0.1818" ], [ "0.501", "0.1654", "0.8857", "0.1654", "0.8857", "0.1799", "0.501", "0.1799" ], [ "0.1405", "0.1718", "0.1772", "0.1718", "0.1772", "0.1818", "0.1405", "0.1818" ], [ "0.5024", "0.1799", "0.884", "0.1799", "0.884", "0.1932", "0.5024", "0.1932" ], [ "0.0866", "0.1957", "0.4756", "0.1945", "0.4756", "0.209", "0.0866", "0.2102" ], [ "0.501", "0.1945", "0.88", "0.1945", "0.88", "0.2084", "0.501", "0.2084" ], [ "0.08826", "0.209", "0.4722", "0.209", "0.4722", "0.2229", "0.08826", "0.2229" ], [ "0.5034", "0.2102", "0.8857", "0.2102", "0.8857", "0.2235", "0.5034", "0.2235" ], [ "0.08826", "0.2216", "0.4739", "0.2216", "0.4739", "0.2349", "0.08826", "0.2349" ], [ "0.5", "0.2247", "0.886", "0.2247", "0.886", "0.2424", "0.5", "0.2424" ], [ "0.08826", "0.2343", "0.4307", "0.2343", "0.4307", "0.2474", "0.08826", "0.2474" ], [ "0.4248", "0.2349", "0.4739", "0.2349", "0.4739", "0.2468", "0.4248", "0.2468" ], [ "0.0866", "0.2468", "0.4739", "0.2456", "0.4739", "0.26", "0.0866", "0.2615" ], [ "0.501", "0.2412", "0.699", "0.2412", "0.699", "0.2551", "0.501", "0.2551" ], [ "0.0866", "0.2588", "0.4756", "0.2583", "0.4756", "0.2727", "0.0866", "0.2734" ], [ "0.5", "0.2588", "0.886", "0.26", "0.886", "0.2747", "0.5", "0.2734" ], [ "0.0866", "0.2727", "0.4756", "0.2727", "0.4756", "0.286", "0.0866", "0.286" ], [ "0.5", "0.276", "0.8857", "0.276", "0.8857", "0.289", "0.5", "0.289" ], [ "0.08826", "0.2854", "0.4756", "0.2854", "0.4756", "0.2986", "0.08826", "0.2986" ], [ "0.5", "0.289", "0.8857", "0.289", "0.8857", "0.3037", "0.5", "0.3037" ], [ "0.0866", "0.2974", "0.304", "0.2974", "0.304", "0.3113", "0.0866", "0.3113" ], [ "0.5", "0.3062", "0.886", "0.305", "0.886", "0.3193", "0.5", "0.3208" ], [ "0.5", "0.3213", "0.886", "0.3213", "0.886", "0.3352", "0.5", "0.3352" ], [ "0.08905", "0.3352", "0.4739", "0.336", "0.4739", "0.351", "0.08905", "0.3499" ], [ "0.5", "0.3384", "0.524", "0.3384", "0.524", "0.3499", "0.5", "0.3499" ], [ "0.08905", "0.3499", "0.4722", "0.351", "0.4722", "0.3655", "0.08905", "0.3643" ], [ "0.6562", "0.3591", "0.6807", "0.3591", "0.6807", "0.3706", "0.6562", "0.3706" ], [ "0.6235", "0.3655", "0.636", "0.3655", "0.636", "0.3757", "0.6235", "0.3757" ], [ "0.7417", "0.3623", "0.7534", "0.3623", "0.7534", "0.3706", "0.7417", "0.3706" ], [ "0.6953", "0.3718", "0.759", "0.3718", "0.759", "0.3865", "0.6953", "0.3865" ], [ "0.6562", "0.3726", "0.669", "0.3726", "0.669", "0.3806", "0.6562", "0.3806" ], [ "0.0907", "0.3674", "0.4722", "0.3674", "0.4722", "0.3806", "0.0907", "0.3806" ], [ "0.7534", "0.375", "0.8423", "0.3718", "0.8433", "0.387", "0.755", "0.3901" ], [ "0.865", "0.3738", "0.8857", "0.3738", "0.8857", "0.3865", "0.865", "0.3865" ], [ "0.6045", "0.377", "0.6787", "0.3865", "0.6763", "0.4016", "0.6006", "0.3926" ], [ "0.513", "0.3757", "0.597", "0.3757", "0.597", "0.387", "0.513", "0.387" ], [ "0.0907", "0.3818", "0.4192", "0.3818", "0.4192", "0.3958", "0.0907", "0.3958" ], [ "0.7363", "0.3914", "0.759", "0.3914", "0.759", "0.3997", "0.7363", "0.3997" ], [ "0.08826", "0.4136", "0.433", "0.4136", "0.433", "0.428", "0.08826", "0.428" ], [ "0.5", "0.4104", "0.886", "0.411", "0.886", "0.4255", "0.5", "0.4248" ], [ "0.1177", "0.4287", "0.2573", "0.4287", "0.2573", "0.4426", "0.1177", "0.4426" ], [ "0.5", "0.4268", "0.8857", "0.4268", "0.8857", "0.44", "0.5", "0.44" ], [ "0.4983", "0.44", "0.886", "0.4414", "0.886", "0.4558", "0.4983", "0.455" ], [ "0.08905", "0.4507", "0.4739", "0.452", "0.4739", "0.4666", "0.08905", "0.4653" ], [ "0.5", "0.4558", "0.886", "0.457", "0.886", "0.4717", "0.5", "0.4702" ], [ "0.08905", "0.4673", "0.4739", "0.4673", "0.4739", "0.481", "0.08905", "0.481" ], [ "0.5024", "0.4749", "0.517", "0.4749", "0.517", "0.4841", "0.5024", "0.4841" ], [ "0.0907", "0.483", "0.4722", "0.483", "0.4722", "0.4968", "0.0907", "0.4968" ], [ "0.7075", "0.4849", "0.7305", "0.4849", "0.7305", "0.4976", "0.7075", "0.4976" ], [ "0.6846", "0.4924", "0.697", "0.4924", "0.697", "0.5024", "0.6846", "0.5024" ], [ "0.705", "0.4944", "0.773", "0.4944", "0.773", "0.518", "0.705", "0.518" ], [ "0.625", "0.4956", "0.6504", "0.4956", "0.6504", "0.507", "0.625", "0.507" ], [ "0.09235", "0.4976", "0.4739", "0.4976", "0.4739", "0.512", "0.09235", "0.512" ], [ "0.614", "0.5005", "0.671", "0.506", "0.6685", "0.521", "0.6113", "0.515" ], [ "0.863", "0.5005", "0.888", "0.5005", "0.888", "0.5146", "0.863", "0.5146" ], [ "0.0907", "0.513", "0.4739", "0.513", "0.4739", "0.5264", "0.0907", "0.5264" ], [ "0.6772", "0.5083", "0.706", "0.5083", "0.706", "0.523", "0.6772", "0.523" ], [ "0.7036", "0.5176", "0.734", "0.5176", "0.734", "0.526", "0.7036", "0.526" ], [ "0.0907", "0.528", "0.4722", "0.528", "0.4722", "0.5415", "0.0907", "0.5415" ], [ "0.4983", "0.5327", "0.888", "0.535", "0.888", "0.553", "0.4983", "0.5503" ], [ "0.0907", "0.5435", "0.4739", "0.5435", "0.4739", "0.5566", "0.0907", "0.5566" ], [ "0.08905", "0.558", "0.4722", "0.558", "0.4722", "0.5713", "0.08905", "0.5713" ], [ "0.5024", "0.5537", "0.589", "0.5537", "0.589", "0.565", "0.5024", "0.565" ], [ "0.08905", "0.5737", "0.4739", "0.5737", "0.4739", "0.587", "0.08905", "0.587" ], [ "0.5024", "0.579", "0.8823", "0.579", "0.8823", "0.5933", "0.5024", "0.5933" ], [ "0.08905", "0.587", "0.4739", "0.5884", "0.4739", "0.603", "0.08905", "0.6016" ], [ "0.501", "0.5933", "0.8857", "0.5947", "0.8857", "0.6094", "0.501", "0.608" ], [ "0.0907", "0.603", "0.4739", "0.603", "0.4739", "0.617", "0.0907", "0.617" ], [ "0.5034", "0.61", "0.884", "0.61", "0.884", "0.6235", "0.5034", "0.6235" ], [ "0.08905", "0.616", "0.1805", "0.6187", "0.1805", "0.6333", "0.08905", "0.631" ], [ "0.5024", "0.625", "0.884", "0.625", "0.884", "0.638", "0.5024", "0.638" ], [ "0.5024", "0.64", "0.884", "0.64", "0.884", "0.654", "0.5024", "0.654" ], [ "0.2157", "0.6484", "0.2222", "0.6484", "0.2222", "0.655", "0.2157", "0.655" ], [ "0.2361", "0.6514", "0.415", "0.654", "0.415", "0.6724", "0.2361", "0.6704" ], [ "0.5", "0.654", "0.884", "0.656", "0.884", "0.6704", "0.5", "0.6685" ], [ "0.415", "0.659", "0.4412", "0.659", "0.4412", "0.6675", "0.415", "0.6675" ], [ "0.1013", "0.655", "0.1929", "0.655", "0.1929", "0.6694", "0.1013", "0.6694" ], [ "0.2059", "0.662", "0.2303", "0.662", "0.2303", "0.675", "0.2059", "0.675" ], [ "0.2852", "0.6724", "0.3057", "0.6724", "0.3057", "0.681", "0.2852", "0.681" ], [ "0.5024", "0.6704", "0.8857", "0.6704", "0.8857", "0.6846", "0.5024", "0.6846" ], [ "0.452", "0.681", "0.4739", "0.681", "0.4739", "0.6934", "0.452", "0.6934" ], [ "0.5063", "0.6895", "0.5156", "0.6895", "0.5156", "0.6943", "0.5063", "0.6943" ], [ "0.5156", "0.685", "0.884", "0.685", "0.884", "0.6987", "0.5156", "0.6987" ], [ "0.0907", "0.6943", "0.4756", "0.6943", "0.4756", "0.709", "0.0907", "0.709" ], [ "0.5024", "0.7007", "0.634", "0.7007", "0.634", "0.714", "0.5024", "0.714" ], [ "0.0907", "0.7104", "0.4722", "0.7104", "0.4722", "0.7236", "0.0907", "0.7236" ], [ "0.08905", "0.726", "0.4739", "0.726", "0.4739", "0.7393", "0.08905", "0.7393" ], [ "0.5", "0.731", "0.869", "0.732", "0.869", "0.747", "0.5", "0.7456" ], [ "0.08905", "0.7407", "0.4739", "0.7407", "0.4739", "0.755", "0.08905", "0.755" ], [ "0.528", "0.7476", "0.7466", "0.7476", "0.7466", "0.762", "0.528", "0.762" ], [ "0.0907", "0.756", "0.4739", "0.756", "0.4739", "0.7695", "0.0907", "0.7695" ], [ "0.08905", "0.7715", "0.4739", "0.7715", "0.4739", "0.785", "0.08905", "0.785" ], [ "0.4983", "0.7695", "0.8857", "0.7715", "0.8857", "0.786", "0.4983", "0.784" ], [ "0.08826", "0.785", "0.2303", "0.786", "0.2303", "0.8003", "0.08826", "0.8" ], [ "0.513", "0.785", "0.8857", "0.786", "0.8857", "0.8003", "0.513", "0.8" ], [ "0.5073", "0.7915", "0.5156", "0.7915", "0.5156", "0.7954", "0.5073", "0.7954" ], [ "0.501", "0.802", "0.8857", "0.802", "0.8857", "0.815", "0.501", "0.815" ], [ "0.08826", "0.8096", "0.4739", "0.8096", "0.4739", "0.8237", "0.08826", "0.8237" ], [ "0.5", "0.8174", "0.886", "0.8174", "0.886", "0.8306", "0.5", "0.8306" ], [ "0.08826", "0.8247", "0.4739", "0.8247", "0.4739", "0.8384", "0.08826", "0.8384" ], [ "0.5", "0.8306", "0.8857", "0.8306", "0.8857", "0.8447", "0.5", "0.8447" ], [ "0.08826", "0.839", "0.4739", "0.839", "0.4739", "0.854", "0.08826", "0.854" ], [ "0.5", "0.8467", "0.8823", "0.8467", "0.8823", "0.8613", "0.5", "0.8613" ], [ "0.0907", "0.854", "0.4739", "0.854", "0.4739", "0.867", "0.0907", "0.867" ], [ "0.5", "0.86", "0.886", "0.86", "0.886", "0.8774", "0.5", "0.8774" ], [ "0.0866", "0.867", "0.406", "0.8687", "0.406", "0.8833", "0.0866", "0.882" ], [ "0.501", "0.877", "0.886", "0.877", "0.886", "0.89", "0.501", "0.89" ], [ "0.08905", "0.8916", "0.4739", "0.8916", "0.4739", "0.906", "0.08905", "0.906" ], [ "0.501", "0.8916", "0.886", "0.8916", "0.886", "0.906", "0.501", "0.906" ], [ "0.4795", "0.922", "0.4958", "0.922", "0.4958", "0.936", "0.4795", "0.936" ] ]

[ "Wrapped Cauchy Distributed Angular Softmax for Long-Tailed Visual Recognition", "Classj", "h(e,0;mj）求和来获得混合分布f(0;）：", "P", "M", "", "1", " h(p,0;m, )", "fmixed(0; j) =", "M,m=", "(2)", "Normal distribution", "Othermothods", "o Cluster centei", "Ours", "Data ,.,..m,..M", " Individual data", "Representationfeatures", "其中M;是j类中的样本总数。fmixed(0;i）描述了以零", "Inputs", "为中心的M，包裹正态分布的混合物。这种思想用于", "Figure 1: Illustration of our method compared with other", "概率密度函数的非参数估计，例如核密度估计（KDE", "methods. Black dot: representation of each data in one class.", "）（Rosenblatt，1956;Parzen,1962）。然而，与KDE不", "Yellow dot with a black edge: centroid of the cluster. Gray solid", "同，Pij，由j类中的所有pm组成的向量的值可以不", "line: Gaussian kernel boundary. (a) Input data 1, 2,...", "...Min", "Class j. (b) parametric modeling of features from each data via", "同，表示每个数据的异构性", "a wrapped Normal kernel. (Hayat et al., 2019; Kobayashi, 2021)", "Theorem 1. Let fmixed(O:) be a mixed distribution", "(c) Left panel: parametric modeling of features from each data", "formed by summing several wrapped Normal distributions", "via wrapped Normal distribution. Right panel: zoomed diagram", "h(e, 0;m,) (Equation 16 ). h(p, 9; m,) is centered at", "of the magenta box in the left panel.", "μm μm follows Normal distribution N(O, o) centered at", "zero, where -→ 0. Then fmixed(0: j) can be approximated", "具有按类可训练的浓度参数pERC（第3.2节）。我", "as:", "们提供了一些见解，说明为什么我们的新颖softmax是", "M,", "1", "8", "1+2", "1", "表示特征建模的更好的参数分布（第3.2节），以及它", "pmcosnom", "(3)", "2元M,m=", "fmixed(0;)~", "如何在特定条件下创建较大的边距（第3.3节）", "2=1", "3.1. Wrapped Normal Basis for Angular Feature", "Corollary 1.1. Let fmixed(o;) be a mixed distribution", "Density Estimation.", "formed by mixing several wrapped Normal distributions", "(Equation 16 and Equation 3 ), then fmixed(0;) is a", "假设。为了减轻表示特征中的过度拟合，我们使用高", "wrapped distribution with cosine moments, Qmixed, given", "斯分布来近似由噪声或稀疏采样引起的不确定性。因", "by", "此，每个数据点的角度特征都遵循圆坐标中正态分布", "M;", "1", "afn", "(n)", "的概率分布，即。e。，包装正态分布或冯·米塞斯-费", "Cmixed", "(4)", "舍尔分布。鉴于后者近似于前一个分布，为了便于讨", "M,", "m=1", "论，我们将两个分布视为等效。这种使用高斯分布的", "where afn) is the n-th cosine trigonometric moment of", "噪声或稀疏采样引起的不确定性模型已广泛用于各种", "研究(Gal &Ghahramani,2016;Rasmussen& Williams.", "h(e,0:m. )", "2005；Abdaretal.2021）。根据此假设，第j类中第m", "我们在附录A中提供了定理1和推论1.1的详细证明，", "个数据点的角度特征的概率分布函数可以以对称包装", "定理1本质上表明，当将包裹的正态分布内核基与远", "稳定（SWS）分布（Jammalamadaka&SenGupta，2001)", "离零的小扰动求和时，结果可以近似为以零为中心的", "的形式表示：", "包裹正态分布之和。推论1.1证明了混合分布的余弦矩", "可以通过平均每个分布的余弦矩来获得。我们注意到，", "1", "(1+2pmcosn(0m-", "Bailey等人已经证明了两个以零为中心的包裹正态分", "Lim", "h(p, 0; m, j) :", "2元", "2=1", "布的混合分布的余弦矩。。al。(Bailey&Codling，2020)", "(1)", "", "我们在这里证明它可以推广到在某些条件下不以零", "其中nEN，PmE[0,1)表示XMATHX类的XMATHX", "为中心的几个函数", "中心参数。当a=1时，方程16返回包裹的柯西分布，", "对于a=2，我们得到包裹的正态分布（Jammalamadaka", "3.2.Angular Feature Probability Approximation via", "&SenGupta,2001）。pm越大，包装的Normal内核越", "Wrapped Cauchy Distribution.", "紧。由于h(p,0;m,)计算了9m属于具有优化分类", "器权重的i类的概率，因此，对于要根据公式？识别", "It is vital to find the optimal presentation of fmixed(0; 3)", "的正确类，μm→0", "One straightforward solution is to use non-parametric", "", "approaches (Rosenblatt, 1956; Parzen, 1962) : However,", "Note that in our proposed method, we approximate the", "those methods usually require large computational costs", "uncertainty of each Om as wrapped Normal distribution", "for large dataset (Holmstrom, 2000) . In our case, those", "parameterized by Pm and μm instead of modeling the", "methods also requires each pm to be calculated separately", "f(0: ) directly (Hayat et al., 2019; Kobayashi, 2021).", "Therefore, we approximate fmixed(0;) with parametric", "Such difference is shown in Figure 1 (b) and (c).", "distribution, denoting f(0, p; i) : According to Theorem", "按类概率分布。随后，可以通过对j类中的所有", "1 , f(0, p; j) should also be an SwS distribution. Among", "3" ]

{ "arxiv_id": "2305.18732", "language": "zh", "page_idx": 2 }

[ [ "0.2279", "0.05936", "0.7466", "0.05936", "0.7466", "0.06946", "0.2279", "0.06946" ], [ "0.3308", "0.089", "0.3381", "0.089", "0.3381", "0.0966", "0.3308", "0.0966" ], [ "0.3743", "0.1029", "0.3962", "0.1029", "0.3962", "0.1067", "0.3743", "0.1067" ], [ "0.621", "0.1016", "0.626", "0.1016", "0.626", "0.108", "0.621", "0.108" ], [ "0.3743", "0.11743", "0.392", "0.11743", "0.392", "0.1256", "0.3743", "0.1256" ], [ "0.6177", "0.11426", "0.63", "0.11426", "0.63", "0.127", "0.6177", "0.127" ], [ "0.3962", "0.1193", "0.407", "0.1193", "0.407", "0.1237", "0.3962", "0.1237" ], [ "0.3628", "0.1193", "0.3694", "0.1193", "0.3694", "0.1237", "0.3628", "0.1237" ], [ "0.8325", "0.1218", "0.8774", "0.1218", "0.8774", "0.1332", "0.8325", "0.1332" ], [ "0.6177", "0.132", "0.6313", "0.132", "0.6313", "0.1433", "0.6177", "0.1433" ], [ "0.831", "0.1401", "0.88", "0.1401", "0.88", "0.1547", "0.831", "0.1547" ], [ "0.6177", "0.1483", "0.63", "0.1483", "0.63", "0.1604", "0.6177", "0.1604" ], [ "0.10046", "0.1686", "0.1274", "0.1686", "0.1274", "0.1799", "0.10046", "0.1799" ], [ "0.601", "0.1686", "0.6274", "0.1642", "0.6313", "0.1755", "0.6045", "0.1799" ], [ "0.1168", "0.1881", "0.1259", "0.1881", "0.1259", "0.1957", "0.1168", "0.1957" ], [ "0.3464", "0.1901", "0.3545", "0.1901", "0.3545", "0.1945", "0.3464", "0.1945" ], [ "0.621", "0.1863", "0.63", "0.1863", "0.63", "0.1932", "0.621", "0.1932" ], [ "0.1168", "0.207", "0.125", "0.207", "0.125", "0.2122", "0.1168", "0.2122" ], [ "0.3506", "0.2039", "0.357", "0.209", "0.3506", "0.2134", "0.343", "0.2084" ], [ "0.6465", "0.252", "0.818", "0.252", "0.818", "0.2588", "0.6465", "0.2588" ], [ "0.2115", "0.2632", "0.2181", "0.2632", "0.2181", "0.2676", "0.2115", "0.2676" ], [ "0.46", "0.2632", "0.4656", "0.2632", "0.4656", "0.2695", "0.46", "0.2695" ], [ "0.739", "0.2583", "0.7485", "0.2664", "0.7417", "0.2715", "0.7305", "0.2644" ], [ "0.08905", "0.2874", "0.2467", "0.286", "0.2467", "0.2974", "0.08905", "0.2986" ], [ "0.2607", "0.2874", "0.3464", "0.286", "0.3464", "0.2974", "0.2607", "0.2986" ], [ "0.3594", "0.2886", "0.8823", "0.2886", "0.8823", "0.2986", "0.3594", "0.2986" ], [ "0.08826", "0.2986", "0.6177", "0.2986", "0.6177", "0.3125", "0.08826", "0.3125" ], [ "0.08905", "0.3403", "0.4722", "0.3403", "0.4722", "0.3542", "0.08905", "0.3542" ], [ "0.5024", "0.3403", "0.8857", "0.3403", "0.8857", "0.3542", "0.5024", "0.3542" ], [ "0.08826", "0.3547", "0.4756", "0.3547", "0.4756", "0.3694", "0.08826", "0.3694" ], [ "0.501", "0.3547", "0.8857", "0.3547", "0.8857", "0.3694", "0.501", "0.3694" ], [ "0.08905", "0.3706", "0.4739", "0.3706", "0.4739", "0.3838", "0.08905", "0.3838" ], [ "0.501", "0.3706", "0.6323", "0.3706", "0.6323", "0.3838", "0.501", "0.3838" ], [ "0.0866", "0.385", "0.282", "0.3838", "0.282", "0.3984", "0.0866", "0.3997" ], [ "0.5024", "0.3926", "0.8857", "0.3926", "0.8857", "0.4072", "0.5024", "0.4072" ], [ "0.08826", "0.4016", "0.4739", "0.4028", "0.4739", "0.4172", "0.08826", "0.4167" ], [ "0.5024", "0.4077", "0.8857", "0.4077", "0.8857", "0.4216", "0.5024", "0.4216" ], [ "0.08826", "0.4185", "0.4739", "0.4185", "0.4739", "0.433", "0.08826", "0.433" ], [ "0.5", "0.4224", "0.886", "0.4224", "0.886", "0.4363", "0.5", "0.4363" ], [ "0.08905", "0.4343", "0.4739", "0.4343", "0.4739", "0.449", "0.08905", "0.449" ], [ "0.5024", "0.4382", "0.8857", "0.4382", "0.8857", "0.452", "0.5024", "0.452" ], [ "0.08905", "0.4495", "0.4739", "0.4495", "0.4739", "0.4634", "0.08905", "0.4634" ], [ "0.5024", "0.4526", "0.884", "0.4526", "0.884", "0.4666", "0.5024", "0.4666" ], [ "0.08905", "0.464", "0.4756", "0.464", "0.4756", "0.4792", "0.08905", "0.4792" ], [ "0.5024", "0.4685", "0.8823", "0.4685", "0.8823", "0.4817", "0.5024", "0.4817" ], [ "0.08905", "0.4797", "0.4739", "0.4797", "0.4739", "0.4944", "0.08905", "0.4944" ], [ "0.501", "0.4817", "0.8823", "0.483", "0.8823", "0.4976", "0.501", "0.4968" ], [ "0.08905", "0.4944", "0.4739", "0.4944", "0.4739", "0.509", "0.08905", "0.509" ], [ "0.5024", "0.4988", "0.8857", "0.4988", "0.8857", "0.512", "0.5024", "0.512" ], [ "0.08826", "0.509", "0.335", "0.509", "0.335", "0.5234", "0.08826", "0.5234" ], [ "0.5034", "0.513", "0.601", "0.513", "0.601", "0.528", "0.5034", "0.528" ], [ "0.1356", "0.532", "0.3594", "0.531", "0.3594", "0.549", "0.1356", "0.55" ], [ "0.5024", "0.536", "0.884", "0.536", "0.884", "0.55", "0.5024", "0.55" ], [ "0.2091", "0.5537", "0.2303", "0.5537", "0.2303", "0.5625", "0.2091", "0.5625" ], [ "0.5063", "0.552", "0.8823", "0.552", "0.8823", "0.565", "0.5063", "0.565" ], [ "0.1772", "0.563", "0.1814", "0.563", "0.1814", "0.5684", "0.1772", "0.5684" ], [ "0.2059", "0.565", "0.2181", "0.565", "0.2181", "0.5757", "0.2059", "0.5757" ], [ "0.2345", "0.561", "0.3987", "0.563", "0.3987", "0.5884", "0.2345", "0.586" ], [ "0.452", "0.5674", "0.4739", "0.5674", "0.4739", "0.5806", "0.452", "0.5806" ], [ "0.5024", "0.565", "0.8823", "0.565", "0.8823", "0.579", "0.5024", "0.579" ], [ "0.156", "0.5737", "0.2035", "0.5757", "0.2018", "0.5903", "0.1544", "0.5884" ], [ "0.2852", "0.5835", "0.312", "0.5835", "0.312", "0.5947", "0.2852", "0.5947" ], [ "0.5024", "0.5806", "0.8857", "0.5806", "0.8857", "0.595", "0.5024", "0.595" ], [ "0.501", "0.5967", "0.8613", "0.5967", "0.8613", "0.61", "0.501", "0.61" ], [ "0.1978", "0.5854", "0.246", "0.5854", "0.246", "0.5923", "0.1978", "0.5923" ], [ "0.2157", "0.6123", "0.343", "0.6113", "0.343", "0.627", "0.2157", "0.6274" ], [ "0.452", "0.613", "0.4756", "0.613", "0.4756", "0.627", "0.452", "0.627" ], [ "0.5", "0.627", "0.767", "0.6274", "0.767", "0.6426", "0.5", "0.6416" ], [ "0.08826", "0.643", "0.4756", "0.643", "0.4756", "0.661", "0.08826", "0.661" ], [ "0.5034", "0.6514", "0.884", "0.6514", "0.884", "0.6655", "0.5034", "0.6655" ], [ "0.08826", "0.6577", "0.3743", "0.66", "0.3743", "0.678", "0.08826", "0.6753" ], [ "0.3694", "0.661", "0.3857", "0.661", "0.3857", "0.6655", "0.3694", "0.6655" ], [ "0.3945", "0.662", "0.4714", "0.6577", "0.4722", "0.674", "0.3962", "0.678" ], [ "0.3733", "0.672", "0.3906", "0.672", "0.3906", "0.675", "0.3733", "0.675" ], [ "0.5024", "0.6665", "0.884", "0.6665", "0.884", "0.681", "0.5024", "0.681" ], [ "0.08417", "0.674", "0.3782", "0.674", "0.3782", "0.702", "0.08417", "0.702" ], [ "0.5024", "0.6816", "0.6665", "0.6816", "0.6665", "0.696", "0.5024", "0.696" ], [ "0.574", "0.711", "0.8105", "0.72", "0.8096", "0.7456", "0.5728", "0.7363" ], [ "0.7017", "0.705", "0.7524", "0.705", "0.7524", "0.72", "0.7017", "0.72" ], [ "0.308", "0.7134", "0.3284", "0.7134", "0.3284", "0.7236", "0.308", "0.7236" ], [ "0.864", "0.7153", "0.886", "0.7153", "0.886", "0.729", "0.864", "0.729" ], [ "0.152", "0.726", "0.282", "0.726", "0.282", "0.7437", "0.152", "0.7437" ], [ "0.501", "0.7476", "0.8857", "0.7476", "0.8857", "0.765", "0.501", "0.765" ], [ "0.5", "0.762", "0.8857", "0.763", "0.8857", "0.778", "0.5", "0.7764" ], [ "0.326", "0.768", "0.3447", "0.768", "0.3447", "0.7764", "0.326", "0.7764" ], [ "0.1487", "0.78", "0.2917", "0.779", "0.2917", "0.7974", "0.1487", "0.799" ], [ "0.452", "0.782", "0.4756", "0.782", "0.4756", "0.7944", "0.452", "0.7944" ], [ "0.4983", "0.779", "0.886", "0.7764", "0.886", "0.7915", "0.4983", "0.7935" ], [ "0.3186", "0.8", "0.3506", "0.8", "0.3506", "0.8066", "0.3186", "0.8066" ], [ "0.501", "0.7954", "0.5513", "0.7954", "0.5513", "0.8066", "0.501", "0.8066" ], [ "0.08905", "0.821", "0.4739", "0.821", "0.4739", "0.8354", "0.08905", "0.8354" ], [ "0.5024", "0.8164", "0.8823", "0.8164", "0.8823", "0.83", "0.5024", "0.83" ], [ "0.08826", "0.834", "0.4722", "0.8354", "0.4722", "0.8496", "0.08826", "0.8486" ], [ "0.5024", "0.8306", "0.8857", "0.8306", "0.8857", "0.8447", "0.5024", "0.8447" ], [ "0.0866", "0.8477", "0.4387", "0.851", "0.4387", "0.8687", "0.0866", "0.8657" ], [ "0.5", "0.845", "0.8857", "0.8447", "0.8857", "0.8594", "0.5", "0.86" ], [ "0.501", "0.8613", "0.8813", "0.8613", "0.8813", "0.875", "0.501", "0.875" ], [ "0.08905", "0.8755", "0.4739", "0.875", "0.4739", "0.8896", "0.08905", "0.89" ], [ "0.501", "0.877", "0.8857", "0.877", "0.8857", "0.89", "0.501", "0.89" ], [ "0.09235", "0.8926", "0.4722", "0.8926", "0.4722", "0.906", "0.09235", "0.906" ], [ "0.5034", "0.893", "0.8823", "0.893", "0.8823", "0.904", "0.5034", "0.904" ], [ "0.4805", "0.925", "0.4927", "0.925", "0.4927", "0.935", "0.4805", "0.935" ] ]

[ "Wrapped CauchyDistributed Angular SoftmaxforLong-TailedVisual Recognition", "b", "", "-", "", "0.2", "", "", "WC", "0.3 ", "WN", "0.4", "p0.5", "p0.5", "0.6", "06", "0.6", "0:7", "2", "0.10.20.30.40.50.60.70.80.9", "C", "o", "6", "Figure 2: Heatmap of △", ".wn (a)and ", ".we(b)withrespecttopandg.(c)BinaryheatmapshowingwhetherwrappedCauchy)", "(WC: black) or wrapped Normal (WN: gray) is preferred for simulated mixed distribution", "the two predominant SWS distributions (Wrapped Cauchy", "布比与包裹的正态分布更好地对齐。尽管证明是分析", "distribution vs Wrapped Normal distribution), wrapped", "性的，但它基于数值假设，即i类中的pm均匀分布在", "Cauchy distribution can fit Equation 3 better than the", "[0,1)中（定理3）", "wrapped Normal distribution.", "我们还为pm～N(μp，o）的更一般情况提供了数值模", "Theorem 2. Let fmixed(0;) be a mixed distribution", "拟。给定pE[0,1)，我们模拟μ和a，其值为0。1、", "formed by mixed several wrapped Normal distributions", "0。2、0. 3、0. 4、0。5、0。6、0。7、0。8,0. 9。[0,1)", "h(p, O; m,j) of j -th class, centred around zero, defined", "域之外的任何p值分别被裁剪为0和1。图2显示了结", "in Equation 2: f(0, P; ) is the approximated distribution", "果，表明当α。较小（α。≤0.1）时，首选包装正态分", "with a choice of wrapped Normal fwr(0, p; i) and wrapped", "布；否则，首选包装柯西分布。这意味着，除非混合分", "Cauchy fwc(0,p;j) : Let Pi min of f(o, p;j) minimize", "布fmixea(0;i）包含具有相似浓度参数的包裹正态分布", "the least square error between f(o,p;) and mixed", "否则包裹的柯西分布由于其重尾而为fmixea(O;i）提供", "distribution fmixed(0; j) of j -th class:", "了更好的近似", "Apj,wC or WN = IIfwN or wc(p, 0; i)", "假设Pm遵循均匀分布是一个理想化的假设，它简化", "Mi", "了学习过程。在实践中，Pm的实际分布可能更复杂，", "-", "", "(1+2㎡cos nom", "(5)", "但是，正如我们对高斯分布pm的模拟所表明的那样", "2元M,", "n=1", "有一种趋势：Pm值的多样性越大，考虑到柯西分布的", "沉重尾巴，柯西分布作为高斯分布的近似值越有利", "m-1", "Pmin = arg min ,", "(6)", "3.3.LargeMarginpand optimization", "Then the least square error of optimal pi. min of j -th class", "重要的是要获得最佳的β。根据几何级数，可以使用", "is correlated with standard deviation of (", "", "1pnor", "", "元素方式计算（Jammalamadaka&SenGupta，2001）以替", "(ZMs-1 pt ) with respet to n e [1, )", "代形式编写方程式16：", "f(p,0) = 2m(1+ p2- 2p cos 0)", "1 - p?", "M,", "(8)", "Apj.mn,WN x SDn=", "其中p是包含Pie[1,C)的所有类的向量，其总数为", "C。We note that this alternative form of Equation 16 is", "M", "or Apijmle,wc α SDn", "(7)", "presented for the ease of calculating the margin between", "=1", "classes.", "Theorem 3. Let pm of individual h(p, ; m, j) distribute", "通过WCDAS的大额保证金。多项研究表明，基于大余", "uniformly across its defined domain [0, 1),wn and wo", "量的softmax方法可以在平衡的（Dengetal.，2019;Liu", "defined in Equation 27,Then ps.m,.wc<psm.wN", "etal.,2017;2016）和不平衡的数据集（Renetal.，2020;", "Hayatetal.,2019Caoetal.,2019)中带来更好的性能", "定理2和定理3的详细证明在附录A中提供。这些证", "我们在这里证明WCDAS可以在的某个域下与那些", "明证明，由SWS分布构成的混合分布与包裹的柯西分", "方法等效地执行。但是，我们注意到，并非WCDAS中", "4" ]

{ "arxiv_id": "2305.18732", "language": "zh", "page_idx": 3 }

[ [ "0.125", "0.11237", "0.8486", "0.113", "0.8486", "0.1307", "0.125", "0.13" ], [ "0.3823", "0.1351", "0.591", "0.1364", "0.591", "0.154", "0.3823", "0.1528" ], [ "0.1365", "0.1957", "0.8325", "0.1963", "0.8325", "0.214", "0.1365", "0.2134" ], [ "0.4143", "0.2134", "0.556", "0.2122", "0.556", "0.2267", "0.4143", "0.2279" ], [ "0.08826", "0.2299", "0.1544", "0.2299", "0.1544", "0.2443", "0.08826", "0.2443" ], [ "0.2418", "0.2393", "0.3218", "0.2412", "0.3218", "0.2588", "0.2418", "0.2568" ], [ "0.5024", "0.2437", "0.8857", "0.2437", "0.8857", "0.2583", "0.5024", "0.2583" ], [ "0.1209", "0.262", "0.4395", "0.262", "0.4395", "0.2766", "0.1209", "0.2766" ], [ "0.501", "0.2583", "0.884", "0.2588", "0.884", "0.2734", "0.501", "0.2727" ], [ "0.1209", "0.279", "0.4412", "0.2766", "0.4412", "0.2917", "0.1209", "0.2935" ], [ "0.5034", "0.2747", "0.8823", "0.2747", "0.8823", "0.2886", "0.5034", "0.2886" ], [ "0.1217", "0.2942", "0.4395", "0.2942", "0.4395", "0.308", "0.1217", "0.308" ], [ "0.5034", "0.2917", "0.8823", "0.2917", "0.8823", "0.3018", "0.5034", "0.3018" ], [ "0.1217", "0.3093", "0.4395", "0.3093", "0.4395", "0.3225", "0.1217", "0.3225" ], [ "0.5024", "0.305", "0.8857", "0.305", "0.8857", "0.318", "0.5024", "0.318" ], [ "0.1217", "0.3245", "0.4395", "0.3245", "0.4395", "0.3384", "0.1217", "0.3384" ], [ "0.5034", "0.3213", "0.8813", "0.3213", "0.8813", "0.332", "0.5034", "0.332" ], [ "0.1209", "0.3384", "0.4387", "0.3384", "0.4387", "0.3516", "0.1209", "0.3516" ], [ "0.5034", "0.3352", "0.8823", "0.3352", "0.8823", "0.3484", "0.5034", "0.3484" ], [ "0.1209", "0.3528", "0.4412", "0.3542", "0.4412", "0.3687", "0.1209", "0.3674" ], [ "0.5034", "0.351", "0.8823", "0.351", "0.8823", "0.3643", "0.5034", "0.3643" ], [ "0.1217", "0.3694", "0.4395", "0.3694", "0.4395", "0.3833", "0.1217", "0.3833" ], [ "0.5034", "0.3655", "0.8823", "0.3655", "0.8823", "0.379", "0.5034", "0.379" ], [ "0.1217", "0.3838", "0.4395", "0.3838", "0.4395", "0.3977", "0.1217", "0.3977" ], [ "0.5024", "0.3801", "0.5522", "0.3801", "0.5522", "0.3945", "0.5024", "0.3945" ], [ "0.1209", "0.3997", "0.4387", "0.3997", "0.4387", "0.4136", "0.1209", "0.4136" ], [ "0.5024", "0.4028", "0.8823", "0.4028", "0.8823", "0.4167", "0.5024", "0.4167" ], [ "0.1209", "0.414", "0.4395", "0.414", "0.4395", "0.428", "0.1209", "0.428" ], [ "0.501", "0.4185", "0.8857", "0.4185", "0.8857", "0.432", "0.501", "0.432" ], [ "0.1217", "0.43", "0.4395", "0.43", "0.4395", "0.443", "0.1217", "0.443" ], [ "0.5024", "0.433", "0.8813", "0.433", "0.8813", "0.4463", "0.5024", "0.4463" ], [ "0.1209", "0.4458", "0.4395", "0.4458", "0.4395", "0.459", "0.1209", "0.459" ], [ "0.501", "0.4475", "0.8857", "0.449", "0.8857", "0.4622", "0.501", "0.461" ], [ "0.11926", "0.459", "0.2279", "0.4602", "0.2279", "0.4749", "0.11926", "0.4734" ], [ "0.5024", "0.4634", "0.8857", "0.4634", "0.8857", "0.4766", "0.5024", "0.4766" ], [ "0.5024", "0.4778", "0.884", "0.4778", "0.884", "0.4912", "0.5024", "0.4912" ], [ "0.5024", "0.4937", "0.8857", "0.4937", "0.8857", "0.507", "0.5024", "0.507" ], [ "0.08905", "0.5024", "0.2222", "0.5024", "0.2222", "0.5166", "0.08905", "0.5166" ], [ "0.5024", "0.5083", "0.8857", "0.5083", "0.8857", "0.523", "0.5024", "0.523" ], [ "0.0907", "0.528", "0.4722", "0.528", "0.4722", "0.5415", "0.0907", "0.5415" ], [ "0.505", "0.526", "0.607", "0.526", "0.607", "0.536", "0.505", "0.536" ], [ "0.0907", "0.5435", "0.4722", "0.5435", "0.4722", "0.5566", "0.0907", "0.5566" ], [ "0.0907", "0.558", "0.4722", "0.558", "0.4722", "0.5713", "0.0907", "0.5713" ], [ "0.509", "0.5547", "0.884", "0.5566", "0.884", "0.5713", "0.509", "0.5693" ], [ "0.0907", "0.5747", "0.4714", "0.5747", "0.4714", "0.5854", "0.0907", "0.5854" ], [ "0.5317", "0.571", "0.835", "0.5713", "0.835", "0.586", "0.5317", "0.5854" ], [ "0.0907", "0.5884", "0.4739", "0.5884", "0.4739", "0.6016", "0.0907", "0.6016" ], [ "0.509", "0.5967", "0.884", "0.5986", "0.884", "0.613", "0.509", "0.6113" ], [ "0.0907", "0.603", "0.4739", "0.603", "0.4739", "0.616", "0.0907", "0.616" ], [ "0.0907", "0.6187", "0.4722", "0.6187", "0.4722", "0.632", "0.0907", "0.632" ], [ "0.5337", "0.613", "0.8857", "0.613", "0.8857", "0.627", "0.5337", "0.627" ], [ "0.0907", "0.634", "0.4739", "0.634", "0.4739", "0.648", "0.0907", "0.648" ], [ "0.535", "0.63", "0.884", "0.63", "0.884", "0.64", "0.535", "0.64" ], [ "0.5317", "0.6416", "0.5786", "0.643", "0.5767", "0.6577", "0.5312", "0.656" ], [ "0.0907", "0.6484", "0.4722", "0.6484", "0.4722", "0.662", "0.0907", "0.662" ], [ "0.0907", "0.664", "0.4739", "0.664", "0.4739", "0.678", "0.0907", "0.678" ], [ "0.509", "0.6685", "0.884", "0.6694", "0.884", "0.683", "0.509", "0.6816" ], [ "0.0907", "0.6787", "0.4714", "0.6787", "0.4714", "0.6924", "0.0907", "0.6924" ], [ "0.5337", "0.685", "0.8857", "0.685", "0.8857", "0.6987", "0.5337", "0.6987" ], [ "0.0989", "0.6943", "0.2646", "0.6943", "0.2646", "0.7085", "0.0989", "0.7085" ], [ "0.535", "0.7007", "0.591", "0.7007", "0.591", "0.712", "0.535", "0.712" ], [ "0.0907", "0.7173", "0.4722", "0.7173", "0.4722", "0.731", "0.0907", "0.731" ], [ "0.5107", "0.7236", "0.884", "0.726", "0.884", "0.7407", "0.5107", "0.738" ], [ "0.1054", "0.7407", "0.2524", "0.7407", "0.2524", "0.754", "0.1054", "0.754" ], [ "0.2607", "0.7407", "0.4722", "0.7407", "0.4722", "0.754", "0.2607", "0.754" ], [ "0.535", "0.741", "0.8857", "0.741", "0.8857", "0.755", "0.535", "0.755" ], [ "0.08905", "0.754", "0.4722", "0.754", "0.4722", "0.768", "0.08905", "0.768" ], [ "0.09235", "0.7676", "0.1299", "0.7676", "0.1299", "0.778", "0.09235", "0.778" ], [ "0.1356", "0.7666", "0.4714", "0.7666", "0.4714", "0.78", "0.1356", "0.78" ], [ "0.08905", "0.78", "0.4739", "0.78", "0.4739", "0.7935", "0.08905", "0.7935" ], [ "0.501", "0.776", "0.7686", "0.776", "0.7686", "0.7905", "0.501", "0.7905" ], [ "0.0907", "0.7935", "0.1683", "0.7935", "0.1683", "0.8037", "0.0907", "0.8037" ], [ "0.1609", "0.7905", "0.4739", "0.7915", "0.4739", "0.806", "0.1609", "0.8047" ], [ "0.0907", "0.8047", "0.4739", "0.8047", "0.4739", "0.8184", "0.0907", "0.8184" ], [ "0.5024", "0.802", "0.8857", "0.802", "0.8857", "0.815", "0.5024", "0.815" ], [ "0.08905", "0.8174", "0.4722", "0.8174", "0.4722", "0.8306", "0.08905", "0.8306" ], [ "0.5024", "0.8164", "0.8823", "0.8164", "0.8823", "0.8306", "0.5024", "0.8306" ], [ "0.08826", "0.83", "0.4739", "0.83", "0.4739", "0.8433", "0.08826", "0.8433" ], [ "0.5024", "0.8306", "0.8823", "0.8306", "0.8823", "0.8447", "0.5024", "0.8447" ], [ "0.08826", "0.8413", "0.2239", "0.8423", "0.2239", "0.856", "0.08826", "0.855" ], [ "0.5024", "0.8467", "0.8823", "0.8467", "0.8823", "0.86", "0.5024", "0.86" ], [ "0.0866", "0.866", "0.4722", "0.8657", "0.4722", "0.88", "0.0866", "0.881" ], [ "0.5034", "0.863", "0.8823", "0.863", "0.8823", "0.8735", "0.5034", "0.8735" ], [ "0.08905", "0.88", "0.4739", "0.88", "0.4739", "0.893", "0.08905", "0.893" ], [ "0.5024", "0.877", "0.8857", "0.877", "0.8857", "0.89", "0.5024", "0.89" ], [ "0.0866", "0.8916", "0.2035", "0.8926", "0.2035", "0.907", "0.0866", "0.906" ], [ "0.5024", "0.8926", "0.8857", "0.8926", "0.8857", "0.906", "0.5024", "0.906" ] ]

[ "Graph Neural Networks for the Prediction of Substrate-Specific Organic", "Reaction Conditions", "Serim Ryou * 1 Michael R. Maser * 2 Alexander Y. Cui * 3 Travis J. DeLano2 Yisong Yue3", "Sarah E. Reisman ?", "???.3in", "Abstract", "获取足够的数据，以在高维（1,000个+）标签空间", "We present a systematic investigation using", "(Gaoetal.，2018）中进行预测。这提供了可以查询的", "graph neural networks (GNNs) to model or-", "反应类型的灵活性，并包括来自所有有机化学的广泛", "ganic chemical reactions. To do so, we pre", "条件空间。然而，考虑到全局数据集的稀疏性，可靠的", "pared a dataset collection of four ubiquitous", "预测可能仅针对每种反应类型的最常见条件获得，而", "reactions from the organic chemistry litera", "不管输入之间的结构差异如何。这对催化反应造成了", "ture. We evaluate seven different GNN ar", "严重的限制，因为最佳条件通常高度依赖于底物结构", "chitectures for classification tasks pertaining", "（Mahatthananchaietal.，2012）。因此，当建议使用适", "to the identification of experimental reagents", "当的条件时，深度网络区分反应类型的各个图是至关", "and conditions.We find that models are", "重要的", "able to identify specific graph features that", "为了解决当前的局限性，我们在单反应水平上解决了", "affect reaction conditions and lead to accu", "这个预测问题。这允许详细学习结构-条件关系，而无", "rate predictions. The results herein show", "需通过学习的反应规则对查询进行分类。我们证明了", "great promise in advancing molecular ma.", "我们的方法在有机化学中四种有价值反应类型的精选", "chine learning.", "数据集上的实用性：Suzuki偶联（Miyaura&Suzuki，", "1995)，C-N偶联(Bariwal&VanderEycken,2013)", "Negishi 偶联(Negishi et al.,1977)和 Pauson-Khand", "1.Introduction", "反应（PKRs）（Khandetal.，1973）。我们的主要贡献", "图形神经网络（GNN）已迅速成为化学领域（Mater", "可以总结如下：", "&Coote，2019）中强大的预测工具。在药物化学方", "面取得了重大进展，其中药物理化性质的预测因子", "1. We apply domain expertise in dataset curation", "(Coleyetal.2019:Withnalletal..2020)和图形生", "and in the construction of the label space.", "成模型（Eltonetal.,2019;Blaschkeetal.,2018）很", "2. We conduct a thorough assessment of modern", "容易获得。有机合成中也报道了几个例子，包括反应", "类型（Schwalleretal.，2019）的分类，反应产物（Sko", "GNN architectures, which, to our knowledge, has", "raczynskietal.,2017;Coleyetal.，2017）的预测以及", "not yet been studied for models of chemical reac", "tivity.", "反合成设计（Segleretal.，2018;Coleyetal.，2018）", "尽管最近取得了进展，但有关预测反应条件的研究相", "3.We achievehigh accuracy in predicting condition", "对较少，可以说是化学家面临的最具挑战性的任务之", "vectors for all four datasets using strict evaluation", "-(Coleyetal.,2017)。", "metrics.", "当前的深度神经网络（NN）依靠数百万反应训练集来", "4. We provide an interpretability analysis to show", "*Equal contribution", "1 Computational Vision Lab.", "the structural features informing our predictions.", "California Institute of Technology, Pasadena, California,", "USA", "2 Division of Chemistry and Chemical Engineer", "ing, California Institute of Technology, Pasadena, Califor-", "2. Approach and Related Work", "nia,USA", "3 Department of Computing and Mathemati-", "cal Sciences, California Institute of Technology, Pasadena,", "本文介绍的工作与Gaoetal.（2018）具有最大的相似", "California, USA : Correspondence to: Michael R. Maser", "性。在本报告中，单个GNN接受了来自完整Reaxys?", "< [email protected] >,Sarah E. Reisman < reis-", "数据库（Reaxys）的1200万次反应的培训，用于预测", "[email protected]>.", "条件。产品和反应指纹（FP）被用作输人，其中后者", "To appear in the ICML 2020 Workshop on Graph Repre", "被定义为产品和反应物图之间的差异。无论反应类型", "sentation Learning and Beyond (GRLB). Copyright 2020", "如何，对所有样品的一种可能的催化剂，两种溶剂和", "by the author(s).", "两种试剂按顺序进行预测。中间预测与FP输人连接" ]

{ "arxiv_id": "2007.04275", "language": "zh", "page_idx": 0 }

[ [ "0.3594", "0.05936", "0.615", "0.05936", "0.615", "0.06946", "0.3594", "0.06946" ], [ "0.09235", "0.0884", "0.4722", "0.0884", "0.4722", "0.0985", "0.09235", "0.0985" ], [ "0.544", "0.09534", "0.6113", "0.09534", "0.6113", "0.1067", "0.544", "0.1067" ], [ "0.607", "0.0947", "0.835", "0.0947", "0.835", "0.108", "0.607", "0.108" ], [ "0.0907", "0.1016", "0.4739", "0.1016", "0.4739", "0.11554", "0.0907", "0.11554" ], [ "0.0907", "0.11743", "0.407", "0.11743", "0.407", "0.1307", "0.0907", "0.1307" ], [ "0.5156", "0.1206", "0.5596", "0.1206", "0.5596", "0.1351", "0.5156", "0.1351" ], [ "0.665", "0.1218", "0.739", "0.1218", "0.739", "0.1332", "0.665", "0.1332" ], [ "0.7534", "0.1206", "0.7983", "0.1206", "0.7983", "0.1351", "0.7534", "0.1351" ], [ "0.8037", "0.1225", "0.8774", "0.1225", "0.8774", "0.1338", "0.8037", "0.1338" ], [ "0.7573", "0.1383", "0.7876", "0.1383", "0.7876", "0.1547", "0.7573", "0.1547" ], [ "0.5825", "0.1225", "0.649", "0.1225", "0.649", "0.1338", "0.5825", "0.1338" ], [ "0.5107", "0.1395", "0.564", "0.1395", "0.564", "0.154", "0.5107", "0.154" ], [ "0.5864", "0.1395", "0.6445", "0.1395", "0.6445", "0.154", "0.5864", "0.154" ], [ "0.681", "0.1395", "0.722", "0.1395", "0.722", "0.1547", "0.681", "0.1547" ], [ "0.8325", "0.1383", "0.8486", "0.1383", "0.8486", "0.1547", "0.8325", "0.1547" ], [ "0.0907", "0.1547", "0.4739", "0.1547", "0.4739", "0.1686", "0.0907", "0.1686" ], [ "0.09235", "0.1414", "0.4714", "0.1414", "0.4714", "0.1515", "0.09235", "0.1515" ], [ "0.7563", "0.1578", "0.7876", "0.1578", "0.7876", "0.1736", "0.7563", "0.1736" ], [ "0.8325", "0.1578", "0.8486", "0.1578", "0.8486", "0.1749", "0.8325", "0.1749" ], [ "0.6807", "0.159", "0.7236", "0.159", "0.7236", "0.1736", "0.6807", "0.1736" ], [ "0.5195", "0.159", "0.5557", "0.159", "0.5557", "0.1718", "0.5195", "0.1718" ], [ "0.591", "0.1604", "0.6396", "0.1604", "0.6396", "0.1718", "0.591", "0.1718" ], [ "0.09235", "0.1718", "0.4722", "0.1718", "0.4722", "0.1818", "0.09235", "0.1818" ], [ "0.5933", "0.1768", "0.6357", "0.1768", "0.6357", "0.1925", "0.5933", "0.1925" ], [ "0.687", "0.1768", "0.7183", "0.1768", "0.7183", "0.1925", "0.687", "0.1925" ], [ "0.759", "0.178", "0.7876", "0.178", "0.7876", "0.1925", "0.759", "0.1925" ], [ "0.8325", "0.1768", "0.8486", "0.1768", "0.8486", "0.1925", "0.8325", "0.1925" ], [ "0.09235", "0.1849", "0.4722", "0.1849", "0.4722", "0.1989", "0.09235", "0.1989" ], [ "0.5107", "0.1787", "0.5645", "0.1787", "0.5645", "0.1901", "0.5107", "0.1901" ], [ "0.517", "0.1963", "0.558", "0.1963", "0.558", "0.2108", "0.517", "0.2108" ], [ "0.5933", "0.1963", "0.636", "0.1963", "0.636", "0.2108", "0.5933", "0.2108" ], [ "0.687", "0.1957", "0.7183", "0.1957", "0.7183", "0.2108", "0.687", "0.2108" ], [ "0.7607", "0.1963", "0.7837", "0.1963", "0.7837", "0.2108", "0.7607", "0.2108" ], [ "0.8325", "0.1963", "0.8486", "0.1963", "0.8486", "0.2108", "0.8325", "0.2108" ], [ "0.0907", "0.2134", "0.4739", "0.214", "0.4739", "0.2279", "0.0907", "0.2267" ], [ "0.09235", "0.202", "0.4714", "0.202", "0.4714", "0.2122", "0.09235", "0.2122" ], [ "0.0907", "0.2299", "0.4739", "0.2299", "0.4739", "0.2437", "0.0907", "0.2437" ], [ "0.09235", "0.2468", "0.4714", "0.2468", "0.4714", "0.2568", "0.09235", "0.2568" ], [ "0.5", "0.2424", "0.884", "0.2424", "0.884", "0.2556", "0.5", "0.2556" ], [ "0.08905", "0.26", "0.4714", "0.26", "0.4714", "0.2734", "0.08905", "0.2734" ], [ "0.501", "0.2583", "0.7573", "0.2583", "0.7573", "0.2715", "0.501", "0.2715" ], [ "0.0907", "0.276", "0.4739", "0.276", "0.4739", "0.289", "0.0907", "0.289" ], [ "0.501", "0.2798", "0.884", "0.2798", "0.884", "0.2935", "0.501", "0.2935" ], [ "0.08826", "0.289", "0.361", "0.2917", "0.361", "0.305", "0.08826", "0.303" ], [ "0.5024", "0.2954", "0.884", "0.2954", "0.884", "0.3093", "0.5024", "0.3093" ], [ "0.5034", "0.3125", "0.884", "0.3125", "0.884", "0.3225", "0.5034", "0.3225" ], [ "0.08905", "0.327", "0.1896", "0.327", "0.1896", "0.3384", "0.08905", "0.3384" ], [ "0.501", "0.3257", "0.8857", "0.3257", "0.8857", "0.339", "0.501", "0.339" ], [ "0.5024", "0.3416", "0.8823", "0.3416", "0.8823", "0.3547", "0.5024", "0.3547" ], [ "0.08905", "0.3499", "0.4739", "0.351", "0.4739", "0.3655", "0.08905", "0.3643" ], [ "0.501", "0.356", "0.8857", "0.356", "0.8857", "0.3694", "0.501", "0.3694" ], [ "0.0907", "0.3662", "0.4739", "0.3662", "0.4739", "0.3801", "0.0907", "0.3801" ], [ "0.501", "0.3706", "0.8857", "0.3706", "0.8857", "0.3838", "0.501", "0.3838" ], [ "0.08905", "0.3806", "0.4739", "0.3806", "0.4739", "0.3945", "0.08905", "0.3945" ], [ "0.501", "0.387", "0.8823", "0.387", "0.8823", "0.3977", "0.501", "0.3977" ], [ "0.0907", "0.3958", "0.4739", "0.3958", "0.4739", "0.4092", "0.0907", "0.4092" ], [ "0.501", "0.401", "0.8857", "0.401", "0.8857", "0.414", "0.501", "0.414" ], [ "0.0907", "0.411", "0.4739", "0.411", "0.4739", "0.4248", "0.0907", "0.4248" ], [ "0.5", "0.4153", "0.884", "0.4167", "0.884", "0.43", "0.5", "0.4287" ], [ "0.09235", "0.4268", "0.1168", "0.4268", "0.1168", "0.44", "0.09235", "0.44" ], [ "0.501", "0.4312", "0.8857", "0.4312", "0.8857", "0.4443", "0.501", "0.4443" ], [ "0.5", "0.4443", "0.884", "0.4458", "0.884", "0.4602", "0.5", "0.459" ], [ "0.10046", "0.4602", "0.4739", "0.4602", "0.4739", "0.4734", "0.10046", "0.4734" ], [ "0.501", "0.461", "0.8857", "0.461", "0.8857", "0.4749", "0.501", "0.4749" ], [ "0.11926", "0.476", "0.4722", "0.476", "0.4722", "0.4893", "0.11926", "0.4893" ], [ "0.501", "0.4766", "0.884", "0.4766", "0.884", "0.4905", "0.501", "0.4905" ], [ "0.1217", "0.4905", "0.335", "0.4905", "0.335", "0.504", "0.1217", "0.504" ], [ "0.5034", "0.4937", "0.8823", "0.4937", "0.8823", "0.504", "0.5034", "0.504" ], [ "0.501", "0.507", "0.8857", "0.507", "0.8857", "0.521", "0.501", "0.521" ], [ "0.10046", "0.518", "0.4739", "0.518", "0.4739", "0.532", "0.10046", "0.532" ], [ "0.5024", "0.5234", "0.786", "0.5234", "0.786", "0.534", "0.5024", "0.534" ], [ "0.1209", "0.5327", "0.4714", "0.5327", "0.4714", "0.547", "0.1209", "0.547" ], [ "0.11926", "0.5474", "0.2255", "0.549", "0.2255", "0.563", "0.11926", "0.5625" ], [ "0.5024", "0.5435", "0.8823", "0.5435", "0.8823", "0.5566", "0.5024", "0.5566" ], [ "0.5034", "0.561", "0.884", "0.561", "0.884", "0.5713", "0.5034", "0.5713" ], [ "0.10046", "0.577", "0.4739", "0.577", "0.4739", "0.5903", "0.10046", "0.5903" ], [ "0.5024", "0.5747", "0.884", "0.5747", "0.884", "0.5884", "0.5024", "0.5884" ], [ "0.12335", "0.5933", "0.4714", "0.5933", "0.4714", "0.604", "0.12335", "0.604" ], [ "0.5034", "0.5913", "0.884", "0.5913", "0.884", "0.6016", "0.5034", "0.6016" ], [ "0.1209", "0.606", "0.4697", "0.606", "0.4697", "0.619", "0.1209", "0.619" ], [ "0.5024", "0.605", "0.8813", "0.605", "0.8813", "0.6187", "0.5024", "0.6187" ], [ "0.501", "0.619", "0.8857", "0.619", "0.8857", "0.6333", "0.501", "0.6333" ], [ "0.5024", "0.6353", "0.8857", "0.6353", "0.8857", "0.6484", "0.5024", "0.6484" ], [ "0.501", "0.651", "0.8594", "0.651", "0.8594", "0.664", "0.501", "0.664" ], [ "0.0907", "0.657", "0.4714", "0.657", "0.4714", "0.6675", "0.0907", "0.6675" ], [ "0.6562", "0.662", "0.7817", "0.688", "0.7686", "0.726", "0.642", "0.6997" ], [ "0.0907", "0.672", "0.392", "0.672", "0.392", "0.6816", "0.0907", "0.6816" ], [ "0.6167", "0.678", "0.71", "0.6987", "0.701", "0.7246", "0.607", "0.704" ], [ "0.0907", "0.6934", "0.4697", "0.6934", "0.4697", "0.707", "0.0907", "0.707" ], [ "0.8613", "0.6895", "0.888", "0.6895", "0.888", "0.7056", "0.8613", "0.7056" ], [ "0.09235", "0.7104", "0.4714", "0.7104", "0.4714", "0.72", "0.09235", "0.72" ], [ "0.0907", "0.7236", "0.4722", "0.7236", "0.4722", "0.7373", "0.0907", "0.7373" ], [ "0.5024", "0.73", "0.8857", "0.73", "0.8857", "0.7437", "0.5024", "0.7437" ], [ "0.0907", "0.738", "0.4739", "0.738", "0.4739", "0.7524", "0.0907", "0.7524" ], [ "0.5", "0.7437", "0.8774", "0.7456", "0.8774", "0.759", "0.5", "0.757" ], [ "0.08826", "0.7524", "0.4722", "0.754", "0.4722", "0.7676", "0.08826", "0.7666" ], [ "0.5024", "0.768", "0.8823", "0.768", "0.8823", "0.782", "0.5024", "0.782" ], [ "0.1111", "0.778", "0.4739", "0.778", "0.4739", "0.7915", "0.1111", "0.7915" ], [ "0.5024", "0.7827", "0.884", "0.7827", "0.884", "0.797", "0.5024", "0.797" ], [ "0.0907", "0.7925", "0.4722", "0.7925", "0.4722", "0.803", "0.0907", "0.803" ], [ "0.08905", "0.8037", "0.4722", "0.8037", "0.4722", "0.8174", "0.08905", "0.8174" ], [ "0.5024", "0.8", "0.884", "0.8", "0.884", "0.81", "0.5024", "0.81" ], [ "0.08905", "0.8184", "0.4739", "0.8184", "0.4739", "0.829", "0.08905", "0.829" ], [ "0.5024", "0.813", "0.8857", "0.813", "0.8857", "0.827", "0.5024", "0.827" ], [ "0.09235", "0.8306", "0.4722", "0.8306", "0.4722", "0.8413", "0.09235", "0.8413" ], [ "0.5034", "0.83", "0.884", "0.83", "0.884", "0.8403", "0.5034", "0.8403" ], [ "0.0907", "0.8433", "0.4722", "0.8433", "0.4722", "0.854", "0.0907", "0.854" ], [ "0.501", "0.8433", "0.884", "0.8433", "0.884", "0.857", "0.501", "0.857" ], [ "0.08905", "0.855", "0.433", "0.855", "0.433", "0.8687", "0.08905", "0.8687" ], [ "0.10864", "0.8657", "0.4739", "0.866", "0.4739", "0.881", "0.10864", "0.88" ], [ "0.5186", "0.8657", "0.8857", "0.866", "0.8857", "0.881", "0.5186", "0.88" ], [ "0.0907", "0.88", "0.4722", "0.88", "0.4722", "0.893", "0.0907", "0.893" ], [ "0.501", "0.88", "0.886", "0.88", "0.886", "0.893", "0.501", "0.893" ], [ "0.08826", "0.8916", "0.4207", "0.8926", "0.4207", "0.906", "0.08826", "0.905" ], [ "0.5024", "0.893", "0.5923", "0.893", "0.5923", "0.905", "0.5024", "0.905" ] ]

[ "Title Suppressed Due to ExcessiveSize", "为独热向量，并传递到每个后续层，所有这些都是完", "Table 1.", " Summary of reaction sets studied.", "全连接的。尽管在许多情况下需要包含前10名排名和", "“接近匹配”标签，但仍获得了合理的准确性", "name", "raw labels", "bins", "categories", "118", "reactions", "Suzuki", "145,413", "3,315", "5", "优势。我们对聚焦反应集进行建模，以获得细粒度、特", "我们的方法在多个帐户上提供了优于现有技术的显着", "205", "5", "1,528", "C-N", "36,519", "定于图形的预测。我们的网络将所有反应组件的完整", "6,391", "492", "105", "5", "图作为输入，以最大程度地减少编码结构中的信息丢", "Negishi", "PKR", "2,749", "335", "83", "8", "于反应的角色，以确保输出向量的适当化学背景。我", "失。我们的预测使用直接从数据集分析中确定的特定", "们仅分析前1名和前3名的预测排名，以增加通过实", "验测试建议结果的可行性。我们明确对待相对于朴素", "的，该预处理器对从数据集SMILES计算的RDKitmol", "模型的准确性，以提供严格的评估框架。即使使用这", "对象（Open-Source，2006）进行操作", "些严格的指标，我们也可以通过对每个任务上的一套", "我们的建模研究测试了ChainerChem库中的七个", "GNN架构进行系统测试来获得高精度", "GNN架构。每个模型包含两个子网络，这两个子网络", "针对总体任务进行了联合训练。第一个子网是图形处", "3.Methods", "理网络（GPN），它在体系结构之间有所不同，并构", "成了它们相对性能的基础。GPN将输人图转换为学习", "反应数据。文献数据集从Reaxys?数据库中获取，并", "的分子嵌入，这些分子嵌人被连接以形成整体反应向", "经过预处理以删除不完整的记录。准备好的数据集的", "量。这些作为输入传递给第二个子网，即多层感知器", "摘要包含在表1中。详细的处理步骤和数据分析可以", "（MLP）：用于最终预测。在这项工作中探索的GPN包", "在补充材料（SM）中找到，包括分子特性，反应产率", "括神经指纹网络（NFP）（Duvenaudetal，2015），门", "和试剂频率的分布。1数据集准备的一般工作流程如", "控图序列NNs（GGNNs）（Lietal.，2017），消息传", "下：", "递NNs（MPNNs）（Gilmeretal.，2017），编织模块", "NNs（Weave）（Kearnesetal.，2016），关系图注意网", "1. From Reaxys? exports, SMILES string encodings", "络（R-GAT）XMATHX-CN2-7，。3在所有实验中使", "(Weininger, 1988) of reactants and products are", "用Adam优化器（Kingma&Ba，2017），S形交叉熵", "extracted for each data point.", "损失和80/10/10训练/验证/测试拆分来训练100个", "时期的模型。表2和3中的SM中包含了每个模型的", "2. Full condition vectors including reagents, cata-", "其他通用建模参数和详细的超参数设置。", "lysts, solvents, temperatures, etc. are extracted", "for each entry.", "模型输出和评估。我们根据模型在预测每个反应角色", "的真实标签方面的准确性来分析模型的成功。在实践", "3. Dataset conditions are enumerated into dictionar", "中，输出只是对应于完整反应字典的概率向量。通过", "ies by reaction roles, which we term categories.", "分类子学典对它们进行后处理，最终输出是每个类别", "and ground-truth vectors are binned accordingly", "的标签列表，按其概率分数排序。如果在模型的top", "k预测排名中确定了地面实况标签，则类别的预测被", "分类为准确。在这里，我们考虑前1名和前3名的预", "测，尽管这符合偏好。分类精度（Ac）定义如下：", "通过此过程，实现了对标签空间的重大修整（请参见", "ZPnY", "表1），同时仍保留了深度和代表性的字典", "A=N", "学习任务&模型设置。与Gaoetal.（2018）类似", "(1)", "我们将学习问题构造为多标签分类之一。反应物和", "乘积图作为输人馈送到GNN，GNN经过训练以输", "其中Pi和Yi分别是第i个样本的预测和地面实况标", "出二进制条件向量。这些图是使用ChainerChemistry", "签集，N是测试集中的样本数（Wu&Zhou，2017）。", "（ChainerChem）（Tokuietal.，2015）的预处理器构造", "我们直接将模型性能与虚拟预测器（虚拟）进行比较", "ISince Reaxys is a subscription database, we are not", "该虚拟预测器始终建议数据集每个类别中最常出现的", "permitted to publish exported data.We have compiled de", "tailed procedures to prepare each dataset such that those", "标签。由于类别（Cuietal.2019）之间存在可变的类", "with access can replicate our results. Most academic insti-", "别不平衡（有关完整分布，请参见SM），因此我们不", "tutionshavefull-accessReaxys?subscriptions.Wemake", "是对反应模型的Ac值进行平均，而是从基线计算其", "full reaction label dictionaries and all modeling code avail", "平均误差降低（AER）。我们使用AER来比较每个任", "able at https://github.com/slryou41/reaction-gcnn", "2 For consistency, individual reagents and conditions are", "3Abbreviations used here are true to the original re-", "referred to simply as labels, regardless of their identity.", "ports of each architecture; some differ from those in Chain-", "The terms label and bin are used interchangeably.", "erChem code." ]

{ "arxiv_id": "2007.04275", "language": "zh", "page_idx": 1 }

[ [ "0.3635", "0.11426", "0.636", "0.11426", "0.636", "0.1288", "0.3635", "0.1288" ], [ "0.1609", "0.1445", "0.831", "0.1445", "0.831", "0.159", "0.1609", "0.159" ], [ "0.1168", "0.1635", "0.8823", "0.1642", "0.8823", "0.1787", "0.1168", "0.178" ], [ "0.2402", "0.1787", "0.755", "0.1799", "0.755", "0.1976", "0.2402", "0.1963" ], [ "0.2361", "0.1963", "0.7607", "0.1989", "0.7607", "0.2166", "0.2361", "0.214" ], [ "0.339", "0.2134", "0.6504", "0.2153", "0.6504", "0.2329", "0.339", "0.2311" ], [ "0.1854", "0.2343", "0.814", "0.2329", "0.814", "0.2474", "0.1854", "0.2488" ], [ "0.25", "0.2727", "0.3245", "0.2727", "0.3245", "0.2874", "0.25", "0.2874" ], [ "0.722", "0.2734", "0.7607", "0.2734", "0.7607", "0.2854", "0.722", "0.2854" ], [ "0.8057", "0.2734", "0.8364", "0.2734", "0.8364", "0.286", "0.8057", "0.286" ], [ "0.721", "0.2886", "0.7563", "0.2886", "0.7563", "0.2998", "0.721", "0.2998" ], [ "0.7656", "0.2886", "0.7915", "0.2886", "0.7915", "0.2998", "0.7656", "0.2998" ], [ "0.808", "0.2886", "0.831", "0.2886", "0.831", "0.2998", "0.808", "0.2998" ], [ "0.846", "0.2886", "0.9067", "0.2886", "0.9067", "0.2998", "0.846", "0.2998" ], [ "0.7573", "0.2917", "0.7656", "0.2917", "0.7656", "0.2954", "0.7573", "0.2954" ], [ "0.6914", "0.289", "0.7256", "0.289", "0.7256", "0.2986", "0.6914", "0.2986" ], [ "0.11926", "0.2974", "0.452", "0.2974", "0.452", "0.3113", "0.11926", "0.3113" ], [ "0.5186", "0.3018", "0.6626", "0.3018", "0.6626", "0.3125", "0.5186", "0.3125" ], [ "0.697", "0.303", "0.709", "0.303", "0.709", "0.3145", "0.697", "0.3145" ], [ "0.7363", "0.303", "0.7485", "0.303", "0.7485", "0.3132", "0.7363", "0.3132" ], [ "0.773", "0.303", "0.7876", "0.303", "0.7876", "0.3352", "0.773", "0.3352" ], [ "0.8145", "0.303", "0.826", "0.303", "0.826", "0.3145", "0.8145", "0.3145" ], [ "0.8545", "0.3037", "0.865", "0.3037", "0.865", "0.3145", "0.8545", "0.3145" ], [ "0.8965", "0.2998", "0.9067", "0.308", "0.8965", "0.3164", "0.8857", "0.308" ], [ "0.732", "0.31", "0.751", "0.31", "0.751", "0.3958", "0.732", "0.3958" ], [ "0.8096", "0.3093", "0.8267", "0.3093", "0.8267", "0.3945", "0.8096", "0.3945" ], [ "0.8857", "0.31", "0.9043", "0.31", "0.9043", "0.3965", "0.8857", "0.3965" ], [ "0.851", "0.3113", "0.8667", "0.3113", "0.8667", "0.3547", "0.851", "0.3547" ], [ "0.11926", "0.3125", "0.4534", "0.3125", "0.4534", "0.3257", "0.11926", "0.3257" ], [ "0.517", "0.3125", "0.671", "0.3125", "0.671", "0.3225", "0.517", "0.3225" ], [ "0.6953", "0.3113", "0.71", "0.3113", "0.71", "0.334", "0.6953", "0.334" ], [ "0.11926", "0.3257", "0.4534", "0.3257", "0.4534", "0.339", "0.11926", "0.339" ], [ "0.5156", "0.3213", "0.661", "0.3213", "0.661", "0.3352", "0.5156", "0.3352" ], [ "0.6953", "0.33", "0.71", "0.33", "0.71", "0.3547", "0.6953", "0.3547" ], [ "0.773", "0.332", "0.7876", "0.332", "0.7876", "0.3945", "0.773", "0.3945" ], [ "0.1177", "0.339", "0.452", "0.339", "0.452", "0.3528", "0.1177", "0.3528" ], [ "0.517", "0.3333", "0.636", "0.3333", "0.636", "0.3435", "0.517", "0.3435" ], [ "0.517", "0.3416", "0.6787", "0.3416", "0.6787", "0.3547", "0.517", "0.3547" ], [ "0.517", "0.3528", "0.6504", "0.3528", "0.6504", "0.363", "0.517", "0.363" ], [ "0.6953", "0.3516", "0.71", "0.3516", "0.71", "0.3865", "0.6953", "0.3865" ], [ "0.851", "0.3528", "0.8667", "0.3528", "0.8667", "0.3958", "0.851", "0.3958" ], [ "0.5186", "0.363", "0.6504", "0.363", "0.6504", "0.3738", "0.5186", "0.3738" ], [ "0.11926", "0.3542", "0.452", "0.3542", "0.452", "0.3674", "0.11926", "0.3674" ], [ "0.11926", "0.3674", "0.4534", "0.3674", "0.4534", "0.3806", "0.11926", "0.3806" ], [ "0.5186", "0.3738", "0.6587", "0.3738", "0.6587", "0.3838", "0.5186", "0.3838" ], [ "0.11926", "0.3833", "0.452", "0.3833", "0.452", "0.3933", "0.11926", "0.3933" ], [ "0.517", "0.3818", "0.6772", "0.3833", "0.6772", "0.3945", "0.517", "0.3933" ], [ "0.699", "0.3838", "0.706", "0.3838", "0.706", "0.3914", "0.699", "0.3914" ], [ "0.11926", "0.3958", "0.451", "0.3965", "0.451", "0.4077", "0.11926", "0.4072" ], [ "0.5195", "0.3945", "0.6094", "0.3945", "0.6094", "0.4016", "0.5195", "0.4016" ], [ "0.517", "0.4028", "0.6724", "0.4028", "0.6724", "0.4136", "0.517", "0.4136" ], [ "0.7344", "0.404", "0.749", "0.404", "0.749", "0.4224", "0.7344", "0.4224" ], [ "0.773", "0.404", "0.7876", "0.404", "0.7876", "0.4224", "0.773", "0.4224" ], [ "0.812", "0.404", "0.8286", "0.404", "0.8286", "0.4236", "0.812", "0.4236" ], [ "0.853", "0.404", "0.865", "0.404", "0.865", "0.414", "0.853", "0.414" ], [ "0.8896", "0.404", "0.9004", "0.404", "0.9004", "0.414", "0.8896", "0.414" ], [ "0.1209", "0.4104", "0.4495", "0.4104", "0.4495", "0.4204", "0.1209", "0.4204" ], [ "0.517", "0.411", "0.5513", "0.4136", "0.55", "0.4248", "0.5156", "0.4224" ], [ "0.6953", "0.4092", "0.7036", "0.4028", "0.71", "0.4077", "0.7017", "0.414" ], [ "0.699", "0.4153", "0.706", "0.4153", "0.706", "0.4204", "0.699", "0.4204" ], [ "0.8545", "0.4153", "0.864", "0.4153", "0.864", "0.4204", "0.8545", "0.4204" ], [ "0.8896", "0.4124", "0.903", "0.4124", "0.903", "0.4224", "0.8896", "0.4224" ], [ "0.1177", "0.4204", "0.4534", "0.4224", "0.4534", "0.4375", "0.1177", "0.435" ], [ "0.511", "0.4255", "0.91", "0.428", "0.91", "0.4414", "0.511", "0.4395" ], [ "0.1177", "0.4363", "0.4495", "0.4363", "0.4495", "0.4495", "0.1177", "0.4495" ], [ "0.1177", "0.4507", "0.451", "0.4507", "0.451", "0.464", "0.1177", "0.464" ], [ "0.5312", "0.4526", "0.911", "0.4526", "0.911", "0.4666", "0.5312", "0.4666" ], [ "0.11926", "0.4634", "0.4534", "0.4634", "0.4534", "0.4766", "0.11926", "0.4766" ], [ "0.5156", "0.4666", "0.911", "0.4666", "0.911", "0.4797", "0.5156", "0.4797" ], [ "0.11926", "0.4778", "0.451", "0.4778", "0.451", "0.4912", "0.11926", "0.4912" ], [ "0.5156", "0.4797", "0.9126", "0.4797", "0.9126", "0.4944", "0.5156", "0.4944" ], [ "0.11926", "0.4912", "0.452", "0.4912", "0.452", "0.506", "0.11926", "0.506" ], [ "0.513", "0.4937", "0.9126", "0.4924", "0.9126", "0.507", "0.513", "0.5083" ], [ "0.1177", "0.506", "0.452", "0.506", "0.452", "0.5195", "0.1177", "0.5195" ], [ "0.5146", "0.5083", "0.911", "0.5083", "0.911", "0.5215", "0.5146", "0.5215" ], [ "0.11926", "0.5195", "0.4207", "0.5195", "0.4207", "0.5327", "0.11926", "0.5327" ], [ "0.5146", "0.5215", "0.9126", "0.5215", "0.9126", "0.536", "0.5146", "0.536" ], [ "0.5146", "0.534", "0.9126", "0.535", "0.9126", "0.55", "0.5146", "0.549" ], [ "0.11926", "0.5547", "0.2279", "0.556", "0.2279", "0.571", "0.11926", "0.5693" ], [ "0.513", "0.55", "0.5317", "0.55", "0.5317", "0.5625", "0.513", "0.5625" ], [ "0.08826", "0.558", "0.10455", "0.558", "0.10455", "0.5674", "0.08826", "0.5674" ], [ "0.5293", "0.5625", "0.9126", "0.5625", "0.9126", "0.5757", "0.5293", "0.5757" ], [ "0.0866", "0.5747", "0.4836", "0.5747", "0.4836", "0.589", "0.0866", "0.589" ], [ "0.5146", "0.577", "0.9126", "0.577", "0.9126", "0.5903", "0.5146", "0.5903" ], [ "0.08905", "0.5913", "0.482", "0.5913", "0.482", "0.6016", "0.08905", "0.6016" ], [ "0.5146", "0.5913", "0.911", "0.5913", "0.911", "0.605", "0.5146", "0.605" ], [ "0.08826", "0.604", "0.4722", "0.604", "0.4722", "0.6187", "0.08826", "0.6187" ], [ "0.513", "0.604", "0.911", "0.604", "0.911", "0.617", "0.513", "0.617" ], [ "0.08826", "0.6187", "0.4844", "0.6187", "0.4844", "0.632", "0.08826", "0.632" ], [ "0.5146", "0.6187", "0.9126", "0.6187", "0.9126", "0.6333", "0.5146", "0.6333" ], [ "0.0907", "0.632", "0.4836", "0.632", "0.4836", "0.646", "0.0907", "0.646" ], [ "0.511", "0.631", "0.911", "0.632", "0.911", "0.6465", "0.511", "0.646" ], [ "0.08826", "0.6465", "0.4844", "0.6465", "0.4844", "0.66", "0.08826", "0.66" ], [ "0.5146", "0.6465", "0.91", "0.6465", "0.91", "0.66", "0.5146", "0.66" ], [ "0.0866", "0.659", "0.4844", "0.659", "0.4844", "0.6724", "0.0866", "0.6724" ], [ "0.513", "0.6577", "0.9126", "0.659", "0.9126", "0.674", "0.513", "0.6724" ], [ "0.0866", "0.672", "0.486", "0.6724", "0.486", "0.6875", "0.0866", "0.686" ], [ "0.5146", "0.6724", "0.6807", "0.6724", "0.6807", "0.686", "0.5146", "0.686" ], [ "0.0866", "0.6875", "0.2418", "0.6875", "0.2418", "0.7007", "0.0866", "0.7007" ], [ "0.528", "0.6875", "0.911", "0.686", "0.911", "0.7007", "0.528", "0.702" ], [ "0.10455", "0.702", "0.4844", "0.702", "0.4844", "0.7153", "0.10455", "0.7153" ], [ "0.5156", "0.702", "0.9126", "0.702", "0.9126", "0.7153", "0.5156", "0.7153" ], [ "0.08826", "0.714", "0.4836", "0.714", "0.4836", "0.728", "0.08826", "0.728" ], [ "0.511", "0.714", "0.911", "0.7153", "0.911", "0.73", "0.511", "0.729" ], [ "0.08826", "0.729", "0.486", "0.729", "0.486", "0.742", "0.08826", "0.742" ], [ "0.5146", "0.73", "0.91", "0.73", "0.91", "0.7437", "0.5146", "0.7437" ], [ "0.0866", "0.741", "0.4836", "0.741", "0.4836", "0.755", "0.0866", "0.755" ], [ "0.513", "0.742", "0.911", "0.742", "0.911", "0.756", "0.513", "0.756" ], [ "0.0866", "0.757", "0.4844", "0.757", "0.4844", "0.771", "0.0866", "0.771" ], [ "0.513", "0.7583", "0.661", "0.7583", "0.661", "0.7715", "0.513", "0.7715" ], [ "0.08826", "0.771", "0.486", "0.771", "0.486", "0.784", "0.08826", "0.784" ], [ "0.08826", "0.784", "0.4844", "0.784", "0.4844", "0.7974", "0.08826", "0.7974" ], [ "0.5146", "0.789", "0.527", "0.789", "0.527", "0.7974", "0.5146", "0.7974" ], [ "0.524", "0.789", "0.544", "0.789", "0.544", "0.797", "0.524", "0.797" ], [ "0.5435", "0.7905", "0.5483", "0.7905", "0.5483", "0.7954", "0.5435", "0.7954" ], [ "0.544", "0.787", "0.71", "0.787", "0.71", "0.8003", "0.544", "0.8003" ], [ "0.08826", "0.799", "0.4844", "0.799", "0.4844", "0.8125", "0.08826", "0.8125" ], [ "0.5146", "0.806", "0.911", "0.806", "0.911", "0.8193", "0.5146", "0.8193" ], [ "0.08826", "0.8125", "0.4844", "0.8125", "0.4844", "0.8257", "0.08826", "0.8257" ], [ "0.511", "0.8193", "0.911", "0.8184", "0.911", "0.8335", "0.511", "0.834" ], [ "0.08905", "0.827", "0.4844", "0.827", "0.4844", "0.8403", "0.08905", "0.8403" ], [ "0.5146", "0.834", "0.911", "0.834", "0.911", "0.8477", "0.5146", "0.8477" ], [ "0.08826", "0.8413", "0.1365", "0.8413", "0.1365", "0.853", "0.08826", "0.853" ], [ "0.5146", "0.8477", "0.902", "0.8477", "0.902", "0.8613", "0.5146", "0.8613" ], [ "0.1136", "0.863", "0.4844", "0.863", "0.4844", "0.8774", "0.1136", "0.8774" ], [ "0.5146", "0.863", "0.9067", "0.863", "0.9067", "0.8735", "0.5146", "0.8735" ], [ "0.08826", "0.877", "0.1356", "0.877", "0.1356", "0.888", "0.08826", "0.888" ], [ "0.5146", "0.8755", "0.911", "0.8755", "0.911", "0.8896", "0.5146", "0.8896" ] ]

[ "具有连续时间的多代理寻路", "Anton Andreychukl,3,Konstantin Yakovlevl,?, Dor Atzmon4 and Roni Stern4", "1Federal Research Center \"Computer Science and Control\" of Russian Academy of Sciences", "2National Research University “Higher School of Economics\"", "3Peoples'Friendship University of Russia (RUDN University)", "4Ben-Gurion University of the Negev", "[email protected],[email protected],/dorat,sternron [email protected]", "Abstract", "Actions", "Agent", "Cont.", "Ang.", "Vol.", "Opt.Dist", "", "N.U.", "Multi-Agent Pathfinding (MAPF) is the prob-", "CBS-CL [Walker et al., 2017]", "1", "x", "xVy", "x", "x", "+", "×××/////", "X", "××××>>>×", "1X", "lem of finding paths for multiple agents such", "M* [Wagner and Choset, 2015]", "", "that every agent reaches its goal and the agents", "E-ICTS [Walker et al., 2018]", "x>", "XV", "do not collide.Most prior work on MAPF", "MCCBS [Liet al., 2019]", "POST-MAPF [Ma et al., 2017a]", "ORCA [Snape et al., 2011]", "", "Xxxx", "ALAN [Godoy et al., 2018]", "was on grids, assumed agents' actions have uni-", "form duration, and that time is discretized into", "dRRT*[Dobsonetal.,2017]", "timesteps.WeproposeaMAPFalgorithmthat", "AA-SIPP(m）[Yakovlevand", "", "does not rely on these assumptions,is com-", "Andrevchuk.2017", "TP-SIPPwRT [Liu et al., 2019]", ">y", "", "s", "x", "x", "plete,andprovidesprovablyoptimalsolutions", "CCBS", "", "", "", "X", "This algorithm is based on a novel adaptation", "Table 1:Overview:MAPF researchbeyond the basic setting", "of Safe interval path planning (SIPP), a con", "tinuous time single-agent planning algorithm.", "我们提出了Continuous-time conflict-based search", "and a modified version of Conflict-based search", "(CCBS），这是一种MAPF算法，它不依赖于这些假设", "(CBS), a state of the art multi-agent pathfind-", "中的任何一个，并且是合理的，完整的和最佳的。CCBS", "ing algorithm. We analyze this algorithm, dis-", "基于 Safe interval path planning (SIPP) E[Phillips and", "cuss its pros and cons, and evaluate it experi-", "Likhachev，2011]的定制版本，这是一种连续时间的单代", "mentally on several standard benchmarks.", "理寻路算法，并改编了Conflict-based search（CBS）E", "[Sharonetal.，2015]，这是一种最先进的多代理寻路算", "Introduction", "法", "1.", "CCBS依赖于准确检测代理之间的冲突并计算每个代", "Multi-AgentPathfinding（MAPF）是为多个代理查找路", "理的安全间隔的能力，即代理可以开始在边缘上移动", "径的问题，以便每个代理都能达到且标并目代理不会发", "而不会发生冲突的最小时间。在我们的实验中，我们使", "生冲突。MAPF在仓库管理E[Wurmanet al.，2008]", "用闭环公式进行碰撞检测，并使用基于离散化的方法进", "机场牵引E[Morrisetal.，2016]，自动驾驶汽车，机器", "行安全间隔计算。结果表明，与以前提出的算法相比，", "人E[Velosoetal.,2015]和数字娱乐E[Maetal.,2017b]", "CCBS是可行的，并输出较低成本的解决方案。但是，由", "中具有局部应用。虽然可以在多项式时间E[Kornhauser", "于CCBS考虑了代理的几何形状和连续时间，因此它可", "etal.，1984|中找到MAPF的解，但在几个常见假设F", "能比基于网格的解决方案要慢，从而引入了自然的计划", "[Surynek，2010；YuandLaValle，2013]下，最优地求解", "成本与计划时间的权衡", "MAPF是NP困难的", "我们不是第一个研究MAPF超出其基本设置的人。但", "尽管如此，过去几年的AI研究人员在为越来越多的场", "是，据我们所知，CCBS是第一个MAPF算法，可以处", "景和一百多个代理寻找最佳解决方案方面取得了实质性", "理非单位动作持续时间，连续时间，非网格域，具有体", "进展E[Sharon et al., 2015; Sharon et al., 2013; Wagner", "积的代理，并且仍然是最佳和完整的。表1概述了以前", "and Choset, 2015: Standley. 2010: Felner et al., 2018", "关于MAPF的工作与CCBS的关系。更详细的讨论在", "Bartak et al.,2017;Yu and LaValle,2013]。然而，大多", "相关的工作部分给出", "数先前的工作假设（1）时间被离散化为时间步长，（2）", "每个动作的持续时间是一个时间步长，并且（3）在每个", "2", "", "", "Problem Definition", "时间步长中，每个代理恰好占据一个位置。这些简化的", "我们考虑在2D工作空间中针对非点平移非旋转代理进", "假设限制了MAPF算法在实际应用中的适用性。实际", "行合作寻路。假设所有代理（1）具有相同的形状和大", "上，大多数先前的工作仅在4个连接的电网上进行了经", "小，（2）以相同（恒定）的速度移动，以及（3）被约", "验评估", "束到相同的环境路线图，i。e。有一个图G=(V,E)", "Camera-ready version of the paper as submitted to IJ-", "其顶点对应于代理可以占用（并在其中等待）的位置", "CAI19", "而边对应于代理从一个位置移动到另一个位置时遍历的" ]

{ "arxiv_id": "1901.05506", "language": "zh", "page_idx": 0 }

[ [ "0.09235", "0.0707", "0.4805", "0.0707", "0.4805", "0.0808", "0.09235", "0.0808" ], [ "0.6006", "0.0745", "0.6113", "0.07764", "0.6055", "0.0884", "0.5947", "0.0852" ], [ "0.625", "0.07574", "0.6323", "0.0821", "0.626", "0.0871", "0.6177", "0.0808" ], [ "0.08905", "0.084", "0.482", "0.084", "0.482", "0.0947", "0.08905", "0.0947" ], [ "0.08826", "0.09784", "0.4844", "0.09784", "0.4844", "0.1111", "0.08826", "0.1111" ], [ "0.596", "0.09344", "0.6094", "0.09534", "0.6055", "0.108", "0.5923", "0.1061" ], [ "0.0866", "0.1111", "0.3015", "0.1111", "0.3015", "0.1256", "0.0866", "0.1256" ], [ "0.10297", "0.125", "0.4844", "0.1237", "0.4844", "0.1383", "0.10297", "0.1395" ], [ "0.597", "0.1193", "0.6045", "0.1187", "0.607", "0.1256", "0.6006", "0.1276" ], [ "0.286", "0.1395", "0.4844", "0.137", "0.4844", "0.1515", "0.286", "0.154" ], [ "0.597", "0.1364", "0.609", "0.1364", "0.609", "0.1509", "0.597", "0.1509" ], [ "0.08826", "0.1528", "0.4844", "0.1528", "0.4844", "0.1666", "0.08826", "0.1666" ], [ "0.0907", "0.1401", "0.2852", "0.1401", "0.2852", "0.1509", "0.0907", "0.1509" ], [ "0.615", "0.1604", "0.638", "0.1604", "0.638", "0.1724", "0.615", "0.1724" ], [ "0.683", "0.1604", "0.697", "0.1604", "0.697", "0.1692", "0.683", "0.1692" ], [ "0.0866", "0.1666", "0.4126", "0.1666", "0.4126", "0.1799", "0.0866", "0.1799" ], [ "0.10455", "0.1799", "0.486", "0.1799", "0.486", "0.1945", "0.10455", "0.1945" ], [ "0.625", "0.1787", "0.638", "0.1849", "0.6313", "0.1932", "0.6177", "0.1875" ], [ "0.6763", "0.1812", "0.714", "0.1812", "0.714", "0.1913", "0.6763", "0.1913" ], [ "0.7114", "0.1818", "0.732", "0.1818", "0.732", "0.1895", "0.7114", "0.1895" ], [ "0.7466", "0.1818", "0.7935", "0.1818", "0.7935", "0.1895", "0.7466", "0.1895" ], [ "0.671", "0.1863", "0.6763", "0.1812", "0.681", "0.1849", "0.675", "0.1895" ], [ "0.6504", "0.1831", "0.665", "0.1787", "0.669", "0.1881", "0.6543", "0.1925" ], [ "0.08826", "0.1945", "0.4756", "0.1945", "0.4756", "0.2084", "0.08826", "0.2084" ], [ "0.08905", "0.2084", "0.4844", "0.2084", "0.4844", "0.2216", "0.08905", "0.2216" ], [ "0.513", "0.2026", "0.911", "0.202", "0.911", "0.2153", "0.513", "0.2166" ], [ "0.08826", "0.2216", "0.486", "0.2216", "0.486", "0.2349", "0.08826", "0.2349" ], [ "0.513", "0.2172", "0.558", "0.2172", "0.558", "0.2279", "0.513", "0.2279" ], [ "0.08905", "0.2361", "0.4844", "0.2361", "0.4844", "0.25", "0.08905", "0.25" ], [ "0.08905", "0.25", "0.4844", "0.25", "0.4844", "0.2632", "0.08905", "0.2632" ], [ "0.511", "0.25", "0.9126", "0.2507", "0.9126", "0.2651", "0.511", "0.2644" ], [ "0.0866", "0.2632", "0.4844", "0.2632", "0.4844", "0.2766", "0.0866", "0.2766" ], [ "0.5146", "0.2651", "0.911", "0.2651", "0.911", "0.279", "0.5146", "0.279" ], [ "0.08905", "0.2766", "0.486", "0.2766", "0.486", "0.2917", "0.08905", "0.2917" ], [ "0.513", "0.279", "0.911", "0.279", "0.911", "0.2922", "0.513", "0.2922" ], [ "0.08905", "0.2917", "0.4844", "0.2917", "0.4844", "0.305", "0.08905", "0.305" ], [ "0.5156", "0.2922", "0.911", "0.2922", "0.911", "0.3062", "0.5156", "0.3062" ], [ "0.08826", "0.3037", "0.4844", "0.3037", "0.4844", "0.3176", "0.08826", "0.3176" ], [ "0.517", "0.308", "0.91", "0.308", "0.91", "0.318", "0.517", "0.318" ], [ "0.08905", "0.318", "0.486", "0.318", "0.486", "0.332", "0.08905", "0.332" ], [ "0.513", "0.318", "0.91", "0.3193", "0.91", "0.334", "0.513", "0.3333" ], [ "0.0866", "0.3308", "0.486", "0.332", "0.486", "0.3467", "0.0866", "0.3452" ], [ "0.513", "0.3333", "0.911", "0.334", "0.911", "0.3484", "0.513", "0.348" ], [ "0.08826", "0.3467", "0.486", "0.3467", "0.486", "0.3599", "0.08826", "0.3599" ], [ "0.5146", "0.3484", "0.911", "0.3484", "0.911", "0.3623", "0.5146", "0.3623" ], [ "0.08826", "0.3599", "0.4028", "0.3599", "0.4028", "0.3738", "0.08826", "0.3738" ], [ "0.5156", "0.3623", "0.911", "0.3623", "0.911", "0.3757", "0.5156", "0.3757" ], [ "0.5156", "0.3757", "0.91", "0.3757", "0.91", "0.3894", "0.5156", "0.3894" ], [ "0.08496", "0.3894", "0.3635", "0.3894", "0.3635", "0.4028", "0.08496", "0.4028" ], [ "0.528", "0.3901", "0.911", "0.3901", "0.911", "0.404", "0.528", "0.404" ], [ "0.0866", "0.4077", "0.4878", "0.4072", "0.4878", "0.4204", "0.0866", "0.4216" ], [ "0.5146", "0.404", "0.7812", "0.404", "0.7812", "0.4172", "0.5146", "0.4172" ], [ "0.08826", "0.4204", "0.4844", "0.4216", "0.4844", "0.4363", "0.08826", "0.435" ], [ "0.511", "0.428", "0.7197", "0.428", "0.7197", "0.4414", "0.511", "0.4414" ], [ "0.08905", "0.4363", "0.4844", "0.4363", "0.4844", "0.4495", "0.08905", "0.4495" ], [ "0.513", "0.4443", "0.9126", "0.4458", "0.9126", "0.4602", "0.513", "0.459" ], [ "0.0907", "0.4507", "0.4844", "0.4507", "0.4844", "0.464", "0.0907", "0.464" ], [ "0.08826", "0.464", "0.4844", "0.464", "0.4844", "0.4778", "0.08826", "0.4778" ], [ "0.5146", "0.4602", "0.5376", "0.4602", "0.5376", "0.473", "0.5146", "0.473" ], [ "0.08826", "0.4778", "0.4844", "0.4778", "0.4844", "0.4912", "0.08826", "0.4912" ], [ "0.5254", "0.4766", "0.9126", "0.4792", "0.9126", "0.4937", "0.5254", "0.4912" ], [ "0.08905", "0.4937", "0.4836", "0.4937", "0.4836", "0.504", "0.08905", "0.504" ], [ "0.5415", "0.4937", "0.763", "0.4937", "0.763", "0.507", "0.5415", "0.507" ], [ "0.0866", "0.505", "0.4844", "0.506", "0.4844", "0.5195", "0.0866", "0.518" ], [ "0.5254", "0.509", "0.9126", "0.512", "0.9126", "0.526", "0.5254", "0.523" ], [ "0.0907", "0.5195", "0.4844", "0.5195", "0.4844", "0.534", "0.0907", "0.534" ], [ "0.5415", "0.5244", "0.8105", "0.526", "0.8105", "0.539", "0.5415", "0.5386" ], [ "0.08826", "0.5327", "0.4844", "0.5327", "0.4844", "0.547", "0.08826", "0.547" ], [ "0.0866", "0.5474", "0.4844", "0.5474", "0.4844", "0.5625", "0.0866", "0.5625" ], [ "0.527", "0.5425", "0.9126", "0.5425", "0.9126", "0.556", "0.527", "0.556" ], [ "0.0866", "0.5596", "0.4844", "0.56", "0.4844", "0.5747", "0.0866", "0.5737" ], [ "0.544", "0.558", "0.853", "0.558", "0.853", "0.5713", "0.544", "0.5713" ], [ "0.08905", "0.5747", "0.4844", "0.5747", "0.4844", "0.5884", "0.08905", "0.5884" ], [ "0.527", "0.577", "0.914", "0.577", "0.914", "0.5903", "0.527", "0.5903" ], [ "0.08905", "0.5903", "0.4844", "0.5903", "0.4844", "0.601", "0.08905", "0.601" ], [ "0.544", "0.5903", "0.857", "0.5903", "0.857", "0.604", "0.544", "0.604" ], [ "0.08826", "0.6016", "0.4844", "0.6016", "0.4844", "0.616", "0.08826", "0.616" ], [ "0.513", "0.6094", "0.744", "0.6094", "0.744", "0.6226", "0.513", "0.6226" ], [ "0.08905", "0.616", "0.4844", "0.616", "0.4844", "0.63", "0.08905", "0.63" ], [ "0.08826", "0.631", "0.3162", "0.631", "0.3162", "0.6445", "0.08826", "0.6445" ], [ "0.5195", "0.629", "0.7183", "0.629", "0.7183", "0.6426", "0.5195", "0.6426" ], [ "0.09235", "0.648", "0.4844", "0.6484", "0.4844", "0.662", "0.09235", "0.661" ], [ "0.5146", "0.646", "0.911", "0.646", "0.911", "0.659", "0.5146", "0.659" ], [ "0.08905", "0.662", "0.486", "0.662", "0.486", "0.6753", "0.08905", "0.6753" ], [ "0.513", "0.659", "0.911", "0.659", "0.911", "0.6724", "0.513", "0.6724" ], [ "0.08905", "0.678", "0.4836", "0.678", "0.4836", "0.688", "0.08905", "0.688" ], [ "0.513", "0.6724", "0.911", "0.6724", "0.911", "0.686", "0.513", "0.686" ], [ "0.08826", "0.69", "0.4844", "0.69", "0.4844", "0.704", "0.08826", "0.704" ], [ "0.5156", "0.688", "0.91", "0.688", "0.91", "0.6987", "0.5156", "0.6987" ], [ "0.0866", "0.704", "0.4844", "0.704", "0.4844", "0.7173", "0.0866", "0.7173" ], [ "0.513", "0.6987", "0.91", "0.6997", "0.91", "0.714", "0.513", "0.7134" ], [ "0.08826", "0.7173", "0.486", "0.7173", "0.486", "0.731", "0.08826", "0.731" ], [ "0.513", "0.7134", "0.911", "0.714", "0.911", "0.729", "0.513", "0.728" ], [ "0.08826", "0.732", "0.486", "0.732", "0.486", "0.7456", "0.08826", "0.7456" ], [ "0.511", "0.726", "0.9126", "0.728", "0.9126", "0.744", "0.511", "0.742" ], [ "0.08905", "0.747", "0.4844", "0.747", "0.4844", "0.7603", "0.08905", "0.7603" ], [ "0.5146", "0.742", "0.6284", "0.742", "0.6284", "0.756", "0.5146", "0.756" ], [ "0.08905", "0.7603", "0.4844", "0.7603", "0.4844", "0.7734", "0.08905", "0.7734" ], [ "0.5312", "0.756", "0.911", "0.756", "0.911", "0.7695", "0.5312", "0.7695" ], [ "0.08905", "0.7734", "0.4844", "0.7734", "0.4844", "0.787", "0.08905", "0.787" ], [ "0.5146", "0.7695", "0.911", "0.7695", "0.911", "0.7827", "0.5146", "0.7827" ], [ "0.0866", "0.786", "0.4844", "0.787", "0.4844", "0.802", "0.0866", "0.8003" ], [ "0.5146", "0.782", "0.911", "0.784", "0.911", "0.7974", "0.5146", "0.7954" ], [ "0.08905", "0.802", "0.4836", "0.802", "0.4836", "0.815", "0.08905", "0.815" ], [ "0.5254", "0.797", "0.91", "0.7974", "0.91", "0.8125", "0.5254", "0.811" ], [ "0.08826", "0.815", "0.4844", "0.815", "0.4844", "0.829", "0.08826", "0.829" ], [ "0.511", "0.81", "0.911", "0.8125", "0.911", "0.827", "0.511", "0.8247" ], [ "0.08905", "0.83", "0.4143", "0.83", "0.4143", "0.8433", "0.08905", "0.8433" ], [ "0.5146", "0.8257", "0.6626", "0.8257", "0.6626", "0.839", "0.5146", "0.839" ], [ "0.09235", "0.8477", "0.4836", "0.8477", "0.4836", "0.8613", "0.09235", "0.8613" ], [ "0.5186", "0.845", "0.726", "0.845", "0.726", "0.86", "0.5186", "0.86" ], [ "0.0866", "0.86", "0.486", "0.8613", "0.486", "0.8755", "0.0866", "0.875" ], [ "0.511", "0.86", "0.911", "0.8613", "0.911", "0.8755", "0.511", "0.875" ], [ "0.08905", "0.8755", "0.4844", "0.8755", "0.4844", "0.8896", "0.08905", "0.8896" ], [ "0.513", "0.8755", "0.911", "0.8755", "0.911", "0.8896", "0.513", "0.8896" ] ]

[ "直线轨迹。惯性效应被勿略，代理立即开始/停止移动", "5", "P", "移动的持续时间是平移速度乘以边的长度。等待操作的", "持续时间可以是任何正实数。先前的工作将此设置称为", "4", "MAPFR [Walker et al., 2018]", "请注意，我们在本工作中提出的CCBS算法不限于上", "3", "e。g。，它可以处理以不同速", "2", "度移动的代理，使用不同的路线图，并且具有复杂的形", "述假设（1）：（2）和（3）：", "G", "H", "状和大小。我们做这些假设只是为了简化说明", "代理i的计划是一系列操作元i，因此，如果i执行此", "1", "3", "4", "56", "", "2", "操作序列，则它将达到其目标。一套计划，每个代理人", "个，称为联合计划。MAPFR的解决方案是一个联合计", "Figure1:Our running example:aMAPFRproblem with3", "划，这样，如果所有代理都开始同时执行各自的计划，那", "agents", "么所有代理都将到达其目标位置而不会相互冲突。在这", "项工作中，我们专注于寻找成本最优的解决方案。要定义", "这些CBS约束一致的联合计划。对于CT节点N，我", "MAPFR解决方案的成本最优性，我们首先将计划元；的", "们分别用N.constraints和N.I表示其约束和联合计划。", "成本定义为其组成操作的持续时间之和。在MAPF研究", "CT节点N是通过首先设置其约束（N.constraints）然", "中讨论了几种形式的解决方案成本最优性。最值得注意", "后通过运行低级求解器来计算N.II而生成的，该求解器", "的是makespan和 sum of costs(SOC)，其中 makespan", "为每个受N.constraints中与其相关的约束约束纳束的", "是组成计划成本的最大值，SOC是它们的总和。我们在", "代理找到一个计划。如果N.I不包含任何冲突，则N", "这项工作中解决的问题是找到一个给定的MAPFR问题", "是目标。否则，选择N.II中的CBS冲突i,j,a,t）（其", "的解决方案，该解决方案是最优的。r。它的SOC，也", "中是顶点或边）之一，并生成两个新的CT节点N,和", "就是说，没有其他解决方案具有较低的SOC", "Ni。两个节点都具有与N相同的约束集，外加一个新", "的约束：N,添加了约束《i,a,t），N,添加了约束《i,a,t)", "3CBS with Continuous Times", "CBS以最佳优先的方式搜索CT，在每次选代中扩展", "接下来，我们介绍CCBS并提供相关背景。CBSE", "具有最低成本联合计划的CT节点N", "[Sharonetal.2015]是完整且最佳的MAPF求解器，", "3.1 From CBS to CCBS", "设计用于标准MAPF，i。e。，其中时间是离散化的，并", "CCBS遵循CBS框架。CCBS和CBS之间的主要区别", "且所有操作都具有相同的持续时间。它通过分别查找每", "个代理的计划，检测这些计划之间的冲突，并通过重新", "是：", "规划受特定约束的各个代理来解决给定的MAPF问题。", ". To detect conflicts, CCBS uses a geometry-aware", "典型的CBS实现考虑两种类型的冲突：顶点冲突和边", " collision detection mechanism.", "冲突。计划元和元之间的顶点冲突由元组i，j,u,t）定", ": To resolve conflicts, CCBS uses a geometry-aware", "义，这意味着根据这些计划，代理i和计划同时占用", "unsafe-interval detection mechanism.", "ut。元组（i,j,e,t）类似地定义了边冲突，这意味着根", "据元和元，两个代理都计划从相反的方向同时遍历边", ". CCBS adds constraints over pairs of actions and", "eEE。CBS顶点约束由元组《i,u,t）定义，表示禁止", "time ranges, instead of location-time pairs.", "代理i在t处占用顶点。CBS边缘约束类似地由元组", ": For the low-level search, CCBS uses a version of", "(i.e.t）定义，其中eEE。为了保证完整性和最优性", "SIPP adapted to handle CCBS constraints.", "CBS运行两种搜索算法：一种低级搜索算法，用于查找", "接下来，我们详细解释这些区别", "受一组给定约束约束的各个代理的路径；另一种高级搜", "索算法，用于选择要添加的约束", "Conflict Detection in CCBS", "CBS：Low-LevelSearch.CBS中的低级搜索可以是", "在CCBS中，代理可以具有任何几何形状，并且代理的", "任何寻路算法，该算法可以为与给定的CBS约束集一致", "操作可以具有任何持续时间。因此，在遍历不同边的代", "的代理找到最佳计划。为了使诸如A*之类的单代理寻", "理之间可能会发生冲突，以及顶点-边冲突，这是在沿边", "路算法适合考虑CBS约束，搜索空间还必须考虑时间", "移动的代理与在顶点ELietal..2019处等待的代理发", "维度，因为CBS约束（i,u,t）仅在特定时间t阻止位置", "生冲突时发生的。因此，CCBS冲突被定义为操作之间", "U。对于时间离散化的MAPF问题，这意味着此单代理", "的冲突。形式上，CCBS冲突是元组(ai,ti,aj,ti），表", "搜索空间中的状态是一对（u,t），表示代理在时间t位", "示如果代理i在t执行ai，而代理i在t，执行ai，则", "于位置。扩展这样的状态会生成（,t+1）形式的状", "它们将发生冲突", "态，其中等于，表示等待操作，或者等于与相", "任意形状运动物体的碰撞检测是计算机图形学，计", "邻的位置之一。由违反给定CBS约束集的操作生成的状", "算儿何和机器人技术中广泛研究的一个重要问题E", "态将被修剪。在此搜索空间上运行A*将根据需要返回", "[Jimenezetal.,2001]。对于我们实验中使用的设置，有", "与给定的CBS约束集一致的代理目标的最低成本路径。", "个快速的闭环碰撞检测机制E[GuyandKaramouzas,", "教材A*的这个改编很简单，确实大多数关于CBS的论", "2015]。通常，作为MAPF算法的CCBS与所使用的确", "文都没有报道，只是说CBS的低级搜索是A", "切碰撞检测过程无关", "CBS：High-LevelSearch.CBS中的高级搜索算法在", "Resolving Conflicts in CCBS", "ConstraintTree（CT）上工作，该CT是一棵二叉树，其", "CCBS中的高级搜索会像常规CBS一样运行最佳优先", "中每个节点表示一组施加在代理上的CBS约束以及与", "搜索，在每次迭代中选择具有最小成本的联合计划的CT" ]

{ "arxiv_id": "1901.05506", "language": "zh", "page_idx": 1 }

[ [ "0.08496", "0.06946", "0.2695", "0.06946", "0.2695", "0.0833", "0.08496", "0.0833" ], [ "0.6396", "0.06946", "0.6685", "0.06946", "0.6685", "0.0821", "0.6396", "0.0821" ], [ "0.763", "0.0707", "0.834", "0.0707", "0.834", "0.0808", "0.763", "0.0808" ], [ "0.5596", "0.0859", "0.58", "0.0859", "0.58", "0.0947", "0.5596", "0.0947" ], [ "0.6323", "0.0859", "0.6436", "0.0859", "0.6436", "0.09534", "0.6323", "0.09534" ], [ "0.706", "0.0859", "0.7183", "0.0859", "0.7183", "0.09534", "0.706", "0.09534" ], [ "0.778", "0.0859", "0.789", "0.0859", "0.789", "0.0966", "0.778", "0.0966" ], [ "0.859", "0.0833", "0.869", "0.09155", "0.859", "0.09973", "0.847", "0.09155" ], [ "0.74", "0.0871", "0.7534", "0.0871", "0.7534", "0.09534", "0.74", "0.09534" ], [ "0.596", "0.0884", "0.603", "0.0884", "0.603", "0.0947", "0.596", "0.0947" ], [ "0.6724", "0.089", "0.6772", "0.089", "0.6772", "0.0947", "0.6724", "0.0947" ], [ "0.8145", "0.089", "0.8228", "0.0859", "0.8267", "0.09216", "0.819", "0.09534" ], [ "0.626", "0.09973", "0.652", "0.09973", "0.652", "0.11237", "0.626", "0.11237" ], [ "0.09235", "0.09155", "0.4836", "0.09155", "0.4836", "0.1048", "0.09235", "0.1048" ], [ "0.661", "0.09973", "0.687", "0.09973", "0.687", "0.1111", "0.661", "0.1111" ], [ "0.738", "0.09973", "0.7607", "0.09973", "0.7607", "0.1111", "0.738", "0.1111" ], [ "0.773", "0.09973", "0.7974", "0.09973", "0.7974", "0.1111", "0.773", "0.1111" ], [ "0.808", "0.09973", "0.834", "0.09973", "0.834", "0.1111", "0.808", "0.1111" ], [ "0.8433", "0.09973", "0.8677", "0.09973", "0.8677", "0.1111", "0.8433", "0.1111" ], [ "0.6978", "0.101", "0.721", "0.09784", "0.7236", "0.1092", "0.701", "0.11237" ], [ "0.562", "0.1029", "0.569", "0.1029", "0.569", "0.108", "0.562", "0.108" ], [ "0.584", "0.101", "0.609", "0.09784", "0.6113", "0.1092", "0.5864", "0.11237" ], [ "0.585", "0.1067", "0.613", "0.10986", "0.6094", "0.1237", "0.5825", "0.1206" ], [ "0.661", "0.108", "0.689", "0.1111", "0.6855", "0.1225", "0.6587", "0.1193" ], [ "0.699", "0.1092", "0.7236", "0.1092", "0.7236", "0.1225", "0.699", "0.1225" ], [ "0.626", "0.10986", "0.652", "0.10986", "0.652", "0.1225", "0.626", "0.1225" ], [ "0.08826", "0.1048", "0.4836", "0.1048", "0.4836", "0.1187", "0.08826", "0.1187" ], [ "0.7363", "0.10986", "0.7607", "0.10986", "0.7607", "0.1218", "0.7363", "0.1218" ], [ "0.773", "0.10986", "0.7974", "0.10986", "0.7974", "0.1218", "0.773", "0.1218" ], [ "0.808", "0.10986", "0.834", "0.10986", "0.834", "0.1218", "0.808", "0.1218" ], [ "0.8433", "0.10986", "0.869", "0.10986", "0.869", "0.1218", "0.8433", "0.1218" ], [ "0.699", "0.11615", "0.7236", "0.1193", "0.721", "0.1332", "0.6978", "0.1307" ], [ "0.7363", "0.11615", "0.761", "0.1193", "0.759", "0.1332", "0.734", "0.13" ], [ "0.5605", "0.11237", "0.569", "0.11237", "0.569", "0.1218", "0.5605", "0.1218" ], [ "0.0866", "0.11743", "0.4836", "0.1187", "0.4836", "0.1332", "0.0866", "0.132" ], [ "0.584", "0.1187", "0.609", "0.11554", "0.613", "0.13", "0.5884", "0.1332" ], [ "0.6274", "0.1187", "0.6504", "0.1187", "0.6504", "0.132", "0.6274", "0.132" ], [ "0.661", "0.1193", "0.687", "0.1193", "0.687", "0.132", "0.661", "0.132" ], [ "0.7715", "0.1193", "0.796", "0.1193", "0.796", "0.1307", "0.7715", "0.1307" ], [ "0.808", "0.1193", "0.834", "0.1193", "0.834", "0.1307", "0.808", "0.1307" ], [ "0.8447", "0.1187", "0.871", "0.1218", "0.8677", "0.132", "0.8423", "0.1288" ], [ "0.585", "0.13", "0.613", "0.13", "0.613", "0.1427", "0.585", "0.1427" ], [ "0.562", "0.1218", "0.568", "0.1218", "0.568", "0.1288", "0.562", "0.1288" ], [ "0.626", "0.13", "0.652", "0.13", "0.652", "0.1433", "0.626", "0.1433" ], [ "0.6626", "0.13", "0.689", "0.13", "0.689", "0.1433", "0.6626", "0.1433" ], [ "0.699", "0.13", "0.7236", "0.13", "0.7236", "0.1427", "0.699", "0.1427" ], [ "0.7363", "0.13", "0.761", "0.13", "0.761", "0.1433", "0.7363", "0.1433" ], [ "0.7715", "0.13", "0.7974", "0.13", "0.7974", "0.1427", "0.7715", "0.1427" ], [ "0.808", "0.1307", "0.834", "0.1307", "0.834", "0.1427", "0.808", "0.1427" ], [ "0.8433", "0.1307", "0.869", "0.1307", "0.869", "0.1427", "0.8433", "0.1427" ], [ "0.5596", "0.1307", "0.572", "0.1307", "0.572", "0.1427", "0.5596", "0.1427" ], [ "0.08826", "0.1332", "0.4836", "0.1332", "0.4836", "0.1465", "0.08826", "0.1465" ], [ "0.5596", "0.1401", "0.572", "0.1401", "0.572", "0.1509", "0.5596", "0.1509" ], [ "0.585", "0.1401", "0.613", "0.1401", "0.613", "0.1528", "0.585", "0.1528" ], [ "0.626", "0.1401", "0.6504", "0.1401", "0.6504", "0.154", "0.626", "0.154" ], [ "0.7363", "0.137", "0.761", "0.1401", "0.759", "0.154", "0.734", "0.1509" ], [ "0.6724", "0.1445", "0.6772", "0.1445", "0.6772", "0.1622", "0.6724", "0.1622" ], [ "0.7715", "0.1401", "0.7974", "0.1401", "0.7974", "0.1528", "0.7715", "0.1528" ], [ "0.08905", "0.1465", "0.4836", "0.1465", "0.4836", "0.1604", "0.08905", "0.1604" ], [ "0.5864", "0.1509", "0.613", "0.1509", "0.613", "0.1635", "0.5864", "0.1635" ], [ "0.6274", "0.1509", "0.652", "0.1509", "0.652", "0.1642", "0.6274", "0.1642" ], [ "0.7363", "0.1509", "0.761", "0.1509", "0.761", "0.1635", "0.7363", "0.1635" ], [ "0.7695", "0.1497", "0.7974", "0.1515", "0.796", "0.1642", "0.767", "0.161" ], [ "0.709", "0.1547", "0.7134", "0.1547", "0.7134", "0.161", "0.709", "0.161" ], [ "0.5596", "0.1515", "0.572", "0.1515", "0.572", "0.1635", "0.5596", "0.1635" ], [ "0.08826", "0.1604", "0.4844", "0.1604", "0.4844", "0.1736", "0.08826", "0.1736" ], [ "0.5596", "0.161", "0.572", "0.161", "0.572", "0.1718", "0.5596", "0.1718" ], [ "0.584", "0.161", "0.614", "0.161", "0.614", "0.1724", "0.584", "0.1724" ], [ "0.626", "0.161", "0.652", "0.161", "0.652", "0.1749", "0.626", "0.1749" ], [ "0.7344", "0.161", "0.761", "0.161", "0.761", "0.1736", "0.7344", "0.1736" ], [ "0.7695", "0.161", "0.7974", "0.161", "0.7974", "0.1736", "0.7695", "0.1736" ], [ "0.709", "0.1654", "0.714", "0.1654", "0.714", "0.1736", "0.709", "0.1736" ], [ "0.6724", "0.1666", "0.6772", "0.1666", "0.6772", "0.1749", "0.6724", "0.1749" ], [ "0.7344", "0.1718", "0.7607", "0.1718", "0.7607", "0.1843", "0.7344", "0.1843" ], [ "0.5825", "0.1718", "0.614", "0.1718", "0.614", "0.1843", "0.5825", "0.1843" ], [ "0.1259", "0.1724", "0.4836", "0.1724", "0.4836", "0.1863", "0.1259", "0.1863" ], [ "0.5596", "0.1724", "0.5728", "0.1724", "0.5728", "0.1831", "0.5596", "0.1831" ], [ "0.709", "0.1736", "0.7134", "0.1736", "0.7134", "0.1812", "0.709", "0.1812" ], [ "0.58", "0.1818", "0.614", "0.1818", "0.614", "0.1945", "0.58", "0.1945" ], [ "0.732", "0.1818", "0.7607", "0.1818", "0.7607", "0.1945", "0.732", "0.1945" ], [ "0.558", "0.1831", "0.574", "0.1831", "0.574", "0.1925", "0.558", "0.1925" ], [ "0.09314", "0.178", "0.11273", "0.178", "0.11273", "0.1849", "0.09314", "0.1849" ], [ "0.08826", "0.1881", "0.4844", "0.1881", "0.4844", "0.202", "0.08826", "0.202" ], [ "0.08905", "0.202", "0.4836", "0.202", "0.4836", "0.2153", "0.08905", "0.2153" ], [ "0.5513", "0.2084", "0.8716", "0.209", "0.8716", "0.2203", "0.5513", "0.2197" ], [ "0.08826", "0.2153", "0.4836", "0.214", "0.4836", "0.2279", "0.08826", "0.2291" ], [ "0.08826", "0.2299", "0.4836", "0.2299", "0.4836", "0.2437", "0.08826", "0.2437" ], [ "0.513", "0.2361", "0.5254", "0.2361", "0.5254", "0.2443", "0.513", "0.2443" ], [ "0.536", "0.2393", "0.5435", "0.2393", "0.5435", "0.2437", "0.536", "0.2437" ], [ "0.5483", "0.2343", "0.7305", "0.2343", "0.7305", "0.2474", "0.5483", "0.2474" ], [ "0.08905", "0.2437", "0.4836", "0.2437", "0.4836", "0.2568", "0.08905", "0.2568" ], [ "0.0907", "0.2588", "0.482", "0.2588", "0.482", "0.2695", "0.0907", "0.2695" ], [ "0.5146", "0.2568", "0.911", "0.2568", "0.911", "0.2708", "0.5146", "0.2708" ], [ "0.08905", "0.2715", "0.4844", "0.2715", "0.4844", "0.2854", "0.08905", "0.2854" ], [ "0.5146", "0.2708", "0.911", "0.2708", "0.911", "0.2842", "0.5146", "0.2842" ], [ "0.0866", "0.2842", "0.486", "0.2827", "0.486", "0.2974", "0.0866", "0.2986" ], [ "0.5146", "0.2842", "0.911", "0.2842", "0.911", "0.2974", "0.5146", "0.2974" ], [ "0.10455", "0.2986", "0.4844", "0.2986", "0.4844", "0.3132", "0.10455", "0.3132" ], [ "0.513", "0.2974", "0.9087", "0.2986", "0.9087", "0.3132", "0.513", "0.3125" ], [ "0.0866", "0.3113", "0.4844", "0.3125", "0.4844", "0.327", "0.0866", "0.3257" ], [ "0.5156", "0.3132", "0.911", "0.3132", "0.911", "0.327", "0.5156", "0.327" ], [ "0.08826", "0.327", "0.1209", "0.327", "0.1209", "0.339", "0.08826", "0.339" ], [ "0.5146", "0.3257", "0.911", "0.327", "0.911", "0.3416", "0.5146", "0.3403" ], [ "0.5156", "0.3403", "0.911", "0.3403", "0.911", "0.3542", "0.5156", "0.3542" ], [ "0.08826", "0.3547", "0.4697", "0.3547", "0.4697", "0.3655", "0.08826", "0.3655" ], [ "0.513", "0.3542", "0.91", "0.3542", "0.91", "0.3674", "0.513", "0.3674" ], [ "0.513", "0.3674", "0.9067", "0.3687", "0.9067", "0.3833", "0.513", "0.3818" ], [ "0.08905", "0.3738", "0.4844", "0.3738", "0.4844", "0.387", "0.08905", "0.387" ], [ "0.513", "0.3806", "0.911", "0.3818", "0.911", "0.3965", "0.513", "0.3958" ], [ "0.08905", "0.3882", "0.4844", "0.3882", "0.4844", "0.4016", "0.08905", "0.4016" ], [ "0.09235", "0.4016", "0.4844", "0.4016", "0.4844", "0.4153", "0.09235", "0.4153" ], [ "0.5146", "0.3965", "0.911", "0.3965", "0.911", "0.4104", "0.5146", "0.4104" ], [ "0.08826", "0.4153", "0.4836", "0.4153", "0.4836", "0.4287", "0.08826", "0.4287" ], [ "0.511", "0.4104", "0.6763", "0.4104", "0.6763", "0.4236", "0.511", "0.4236" ], [ "0.08826", "0.4287", "0.4844", "0.4287", "0.4844", "0.4426", "0.08826", "0.4426" ], [ "0.513", "0.4382", "0.562", "0.4382", "0.562", "0.4495", "0.513", "0.4495" ], [ "0.556", "0.4382", "0.652", "0.4382", "0.652", "0.4495", "0.556", "0.4495" ], [ "0.08826", "0.457", "0.2051", "0.457", "0.2051", "0.4702", "0.08826", "0.4702" ], [ "0.08826", "0.4426", "0.4844", "0.4426", "0.4844", "0.457", "0.08826", "0.457" ], [ "0.513", "0.457", "0.911", "0.4583", "0.911", "0.473", "0.513", "0.4717" ], [ "0.10297", "0.4717", "0.486", "0.4717", "0.486", "0.486", "0.10297", "0.486" ], [ "0.5156", "0.473", "0.818", "0.473", "0.818", "0.486", "0.5156", "0.486" ], [ "0.8423", "0.473", "0.91", "0.4734", "0.91", "0.4849", "0.8423", "0.4841" ], [ "0.08826", "0.4849", "0.4836", "0.4849", "0.4836", "0.4988", "0.08826", "0.4988" ], [ "0.513", "0.486", "0.911", "0.486", "0.911", "0.5", "0.513", "0.5" ], [ "0.08905", "0.5005", "0.4836", "0.5005", "0.4836", "0.511", "0.08905", "0.511" ], [ "0.513", "0.5", "0.911", "0.5", "0.911", "0.513", "0.513", "0.513" ], [ "0.08905", "0.513", "0.4844", "0.513", "0.4844", "0.5264", "0.08905", "0.5264" ], [ "0.5156", "0.513", "0.911", "0.513", "0.911", "0.5264", "0.5156", "0.5264" ], [ "0.08826", "0.5264", "0.4844", "0.5264", "0.4844", "0.5405", "0.08826", "0.5405" ], [ "0.511", "0.526", "0.911", "0.5264", "0.911", "0.5405", "0.511", "0.539" ], [ "0.08905", "0.5405", "0.4844", "0.5405", "0.4844", "0.5537", "0.08905", "0.5537" ], [ "0.5146", "0.5415", "0.91", "0.5415", "0.91", "0.5547", "0.5146", "0.5547" ], [ "0.08826", "0.5547", "0.486", "0.5547", "0.486", "0.5684", "0.08826", "0.5684" ], [ "0.5156", "0.5566", "0.91", "0.5566", "0.91", "0.5674", "0.5156", "0.5674" ], [ "0.08826", "0.5684", "0.2345", "0.5684", "0.2345", "0.582", "0.08826", "0.582" ], [ "0.5146", "0.5693", "0.911", "0.5693", "0.911", "0.5835", "0.5146", "0.5835" ], [ "0.10455", "0.582", "0.4844", "0.582", "0.4844", "0.595", "0.10455", "0.595" ], [ "0.5146", "0.582", "0.911", "0.582", "0.911", "0.5967", "0.5146", "0.5967" ], [ "0.0866", "0.5977", "0.4844", "0.5977", "0.4844", "0.6113", "0.0866", "0.6113" ], [ "0.513", "0.595", "0.91", "0.595", "0.91", "0.6094", "0.513", "0.6094" ], [ "0.09314", "0.6094", "0.486", "0.61", "0.486", "0.625", "0.09314", "0.6235" ], [ "0.513", "0.61", "0.9126", "0.61", "0.9126", "0.625", "0.513", "0.625" ], [ "0.08905", "0.625", "0.486", "0.625", "0.486", "0.638", "0.08905", "0.638" ], [ "0.5146", "0.6255", "0.669", "0.6255", "0.669", "0.636", "0.5146", "0.636" ], [ "0.08905", "0.638", "0.486", "0.638", "0.486", "0.6514", "0.08905", "0.6514" ], [ "0.5293", "0.6396", "0.911", "0.6396", "0.911", "0.653", "0.5293", "0.653" ], [ "0.0866", "0.651", "0.486", "0.6514", "0.486", "0.6665", "0.0866", "0.6655" ], [ "0.511", "0.653", "0.911", "0.653", "0.911", "0.6665", "0.511", "0.6665" ], [ "0.08905", "0.6665", "0.486", "0.6665", "0.486", "0.6797", "0.08905", "0.6797" ], [ "0.5146", "0.6685", "0.91", "0.6685", "0.91", "0.6787", "0.5146", "0.6787" ], [ "0.08905", "0.6797", "0.4844", "0.6797", "0.4844", "0.6934", "0.08905", "0.6934" ], [ "0.5146", "0.6797", "0.911", "0.6797", "0.911", "0.6943", "0.5146", "0.6943" ], [ "0.08826", "0.6934", "0.486", "0.6934", "0.486", "0.707", "0.08826", "0.707" ], [ "0.513", "0.6934", "0.91", "0.6934", "0.91", "0.707", "0.513", "0.707" ], [ "0.08905", "0.7085", "0.1217", "0.7085", "0.1217", "0.7197", "0.08905", "0.7197" ], [ "0.513", "0.707", "0.911", "0.709", "0.911", "0.7227", "0.513", "0.72" ], [ "0.10455", "0.7217", "0.4836", "0.7217", "0.4836", "0.735", "0.10455", "0.735" ], [ "0.5146", "0.7217", "0.9126", "0.7217", "0.9126", "0.735", "0.5146", "0.735" ], [ "0.08826", "0.7363", "0.4844", "0.7363", "0.4844", "0.75", "0.08826", "0.75" ], [ "0.513", "0.735", "0.6357", "0.735", "0.6357", "0.7485", "0.513", "0.7485" ], [ "0.10297", "0.75", "0.4836", "0.75", "0.4836", "0.763", "0.10297", "0.763" ], [ "0.5312", "0.75", "0.911", "0.75", "0.911", "0.763", "0.5312", "0.763" ], [ "0.08826", "0.7646", "0.4844", "0.7646", "0.4844", "0.778", "0.08826", "0.778" ], [ "0.513", "0.7646", "0.9126", "0.763", "0.9126", "0.778", "0.513", "0.779" ], [ "0.08826", "0.779", "0.1365", "0.779", "0.1365", "0.7905", "0.08826", "0.7905" ], [ "0.5146", "0.779", "0.914", "0.779", "0.914", "0.7925", "0.5146", "0.7925" ], [ "0.10455", "0.7925", "0.486", "0.7925", "0.486", "0.806", "0.10455", "0.806" ], [ "0.517", "0.7925", "0.911", "0.7925", "0.911", "0.803", "0.517", "0.803" ], [ "0.0866", "0.8047", "0.486", "0.8037", "0.486", "0.8184", "0.0866", "0.8193" ], [ "0.513", "0.806", "0.9126", "0.806", "0.9126", "0.8193", "0.513", "0.8193" ], [ "0.08826", "0.8193", "0.4844", "0.8193", "0.4844", "0.8335", "0.08826", "0.8335" ], [ "0.513", "0.8193", "0.9126", "0.8193", "0.9126", "0.8335", "0.513", "0.8335" ], [ "0.08826", "0.834", "0.4844", "0.834", "0.4844", "0.8477", "0.08826", "0.8477" ], [ "0.5146", "0.834", "0.911", "0.834", "0.911", "0.8477", "0.5146", "0.8477" ], [ "0.08905", "0.8477", "0.4844", "0.8477", "0.4844", "0.8613", "0.08905", "0.8613" ], [ "0.513", "0.8477", "0.9126", "0.8477", "0.9126", "0.8613", "0.513", "0.8613" ], [ "0.08826", "0.8623", "0.486", "0.8623", "0.486", "0.8755", "0.08826", "0.8755" ], [ "0.5146", "0.8623", "0.911", "0.8623", "0.911", "0.8755", "0.5146", "0.8755" ], [ "0.08826", "0.8755", "0.1054", "0.8755", "0.1054", "0.8896", "0.08826", "0.8896" ], [ "0.5146", "0.8755", "0.911", "0.8755", "0.911", "0.8896", "0.5146", "0.8896" ] ]

[ "4PracticalAspects", "SOC", "Success Rate", "k", "3", "5", "3", "5", "2", "2", "4", "", "21.2", "上述分析依赖于具有准确的碰撞检测和不安全间隔检测", "20.4", "1.00", "1.00", "0.97", "0.95", "20.3", "1", "25.7", "38.2", "30.5", "30.2", "31.6", "机制。也就是说，如果确实存在冲突，则会检测到冲突", "0.99", "1.00", "0.88", "0.83", "39.0", "0.98", "6", "并且每个给定的操作对都使用最大的不安全间隔。然而", "49.2", "40.7", "39.3", "0.97", "0.74", "0.57", "61.0", "", "50.5", "48.8", "48.4", "0.95", "0.94", "0.54", "0.42", "10", "构建这样准确的机制并不是微不足道的。有多种方法可", "12", "78.0", "64.7", "0.94", "1", "0.86", "以检测连续空间中具有体积的代理之间的冲突，包括闭", "90.8", "75.3", "0.88", "0.64", "", "14", "环几何计算以及基于采样的方法。参见Jiménez等", "16", "102.4", "85.2", "0.76", "0.53", "", "", "0.62", "118.7", "[2001]用于概述，[Tangetal..2014]用于特定碰撞", "18", ".", "131.7", "0.46", "20", "Ao", "检测过程的示例。对于我们在实验中使用的恒速盘形代", "理，存在[GuyandKaramouzas，2015]中描述的闭环精", "Table2:ResultsforCCBSon10×10opengrid", "确碰撞检测机制。计算操作的不安全间隔W。I。另一个", "动作也需要分析代理的运动学和几何形状。然而，与已", "5", "", "Experimental Results", "经研究了多年并且可以在某些设置中使用闭环公式计算", "的碰撞检测不同，据我们所知，没有这样的闭环公式可", "我们在网格上进行了实验，其中代理可以从一个网格单", "用于计算不安全间隔。计算不安全间隔的一般方法是多", "元的中心移动到另一个网格单元的中心。每个单元的大", "次应用碰撞检测机制，从t，开始，递增一些小的△>0", "小为1×1，每个代理的形状都是一个开放的磁盘，其半", "直到碰撞检测机制报告不安全间隔完成。这种方法的", "径等于V2/4。选择此特定值以允许与CBS进行比较", "局限性在于，产生的不安全间隔可能大于真正的不安全", "因为它是允许代理安全执行代理彼此跟随的移动的最大", "间隔", "半径。为了允许非单位边缘成本，我们允许代理在单个", "移动操作中移动到位于其2k邻域中的每个单元，其中", "4.1ConflictDetectionandSelection Heuristics", "k是参数E[Riveraetal.，2017]。仅在代理可以安全地", "移动到目标单元而不与其他代理或障碍物碰撞的情况下", "如上所述，CCBS中的冲突检测比常规CBS中的更复", "才充许从一个单元移动到另一个单元，其中考虑了代理", "杂。事实上，在我们的实验中，我们观察到冲突检测花费", "了很大一部分时间。为了加快冲突检测速度，我们只检", "和障碍物的几何形状。移动的成本对应于网格单元格中", "查了时间上重叠和可能几何重叠的动作之间的冲突。此", "心之间的欧几里得距离", "外，我们实现了两种启发式方法来加快检测过程。我们", "5.1", "Open Grids", "和最优性的保证", "强调，这些启发式方法不会损害我们对健全性，完整性", "对于第一组实验，我们使用10×10开放网格，随机放", "我们使用的第一个试探法，我们称之为past-conflicts", "置代理的起始和目标位置。我们用4，5，", "20试剂", "试探法，跟踪每对代理（i）在代理i和之间发现冲", "进行实验。对于每一个数量的代理，我们创建了250个", "突的次数。然后：它首先检查具有大量过去冲突的一对", "不同的问题。用CCBS与k=2,3,4和5解决了每个", "代理之间的冲突。然后，当发现冲突时，立即停止搜索", "问题。在https://tinyurl.com/ccbs-example中可以看到", "冲突。然后将找到的冲突存储在CT节点中，如果该CT", "CCBS针对具有13个代理和k的不同值的问题找到的", "节点将被扩展，则它将生成旨在解决此冲突的CT节点。", "解决方案的动画。表2显示了这组实验的结果。每行显", "这实现了这样一种直觉，即过去发生冲突的代理对将来", "示不同数量的代理的结果，如最左列所示。最右边的四", "也更有可能发生冲突", "列显示成功率，即。e。，CCBS在超时60秒的情况下解", "我们发现这种启发式方法在实践中对于减少分配给冲", "决问题的比例，共250个问题。标有“_”的数据点表示", "突检测的时间非常有效。然而，使用这种启发式方法有", "成功率低于0的设置。4。接下来的四列显示了SOC的", "一些限制。先前的工作已经确定，在扩展CT节点时智能", "平均值，该平均值是所有成功率大于0的 CCBS实例解", "选择要解决的冲突可能会对CT的大小和整体运行时产", "决的问题的平均值。4", "生巨大影响[Boyarskietal.，2015]。具体来说，Boyarski", "结果表明，如预期的那样，增加K会产生具有较低", "等人。[2015]引入了基本冲突的概念，这些冲突是解决", "SOC的解决方案。从k=2移动到k=3时，SOC中的", "它们的结果会增加SOC的冲突。半基本冲突是指通过", "绝对差异最大，并且随着我们添加更多代理而增长。例", "为其中一个相关代理重新规划来解决冲突会增加解决方", "如，对于14个代理的问题，从 k=2移动到k=3会", "案成本，但为其他相关代理重新规划不会增加解决方案", "产生15的改进。5SOC，对于16个代理的问题，移至", "成本", "k=3的收益为17。2SOC。增加k进一步表现出收益", "对于CBS，选择先解决基数冲突，然后再解决半基", "递减效应，其中从k=4移动到k=5时，SOC的最大", "数冲突，从而产生了显着的加速E[Boyarskietal.,2015]", "平均收益在 0。5", "但是，要检测基数和半基数冲突，需要识别所有冲突", "然而，增加也具有增加分支因子的效果，这反过来", "而启发式的优点是我们可以在识别所有冲突之前停止搜", "意味着路径查找变得更加困难。事实上，与k=4相比，", "索冲突", "k=5的成功率明显较低。一个例外是从k=2到k=3", "为此，我们提出了第二种混合启发式方法。最初，我", "的过渡，我们观察到k三3在少数代理问题上的成功率", "们检测所有冲突并仅选择主要冲突。但是，如果节点N", "略有优势。例如，对于6个代理，k=2的成功率为0。", "不包含任何基数或半基数冲突，则对于其下CT子树中", "99，而它是1。00fork=3。对此的一种解释是，增加k", "的所有节点，我们切换为使用past-conflicts启发式。这", "也意味着每个代理的计划可以更短，这有助于加快搜索", "种混合方法在我们的实验中效果很好，但是充分探索快", "速度。因此，增加k会引人一个权衡w。r。t。解决问题", "速冲突检测和智能冲突选择之间的权衡是未来工作的主", "的难度：低级搜索的搜索空间较浅，但较宽。对于更密", "题", "集的问题，我。e。，使用更多的代理，k=2在成功率方" ]

{ "arxiv_id": "1901.05506", "language": "zh", "page_idx": 3 }

[ [ "0.2303", "0.1465", "0.7754", "0.1465", "0.7754", "0.161", "0.2303", "0.161" ], [ "0.2737", "0.1692", "0.7295", "0.1692", "0.7295", "0.1843", "0.2737", "0.1843" ], [ "0.3774", "0.2153", "0.6274", "0.2153", "0.6274", "0.2291", "0.3774", "0.2291" ], [ "0.4395", "0.2405", "0.5605", "0.2412", "0.5605", "0.2556", "0.4395", "0.2551" ], [ "0.4167", "0.2556", "0.5884", "0.2556", "0.5884", "0.2695", "0.4167", "0.2695" ], [ "0.2656", "0.3093", "0.739", "0.3093", "0.739", "0.3225", "0.2656", "0.3225" ], [ "0.2656", "0.324", "0.7363", "0.324", "0.7363", "0.3372", "0.2656", "0.3372" ], [ "0.2646", "0.3372", "0.74", "0.3372", "0.74", "0.351", "0.2646", "0.351" ], [ "0.2656", "0.3516", "0.739", "0.3516", "0.739", "0.3655", "0.2656", "0.3655" ], [ "0.2656", "0.3655", "0.738", "0.3655", "0.738", "0.379", "0.2656", "0.379" ], [ "0.267", "0.3806", "0.738", "0.3806", "0.738", "0.3914", "0.267", "0.3914" ], [ "0.2646", "0.3926", "0.738", "0.3926", "0.738", "0.406", "0.2646", "0.406" ], [ "0.2632", "0.4072", "0.739", "0.406", "0.739", "0.4204", "0.2632", "0.4216" ], [ "0.267", "0.4224", "0.7363", "0.4224", "0.7363", "0.433", "0.267", "0.433" ], [ "0.267", "0.4363", "0.738", "0.4363", "0.738", "0.4463", "0.267", "0.4463" ], [ "0.2646", "0.4475", "0.74", "0.4475", "0.74", "0.461", "0.2646", "0.461" ], [ "0.2656", "0.4622", "0.74", "0.4622", "0.74", "0.476", "0.2656", "0.476" ], [ "0.2656", "0.476", "0.739", "0.476", "0.739", "0.4893", "0.2656", "0.4893" ], [ "0.2646", "0.4893", "0.739", "0.4893", "0.739", "0.5024", "0.2646", "0.5024" ], [ "0.2632", "0.504", "0.74", "0.5024", "0.74", "0.5166", "0.2632", "0.5176" ], [ "0.2632", "0.5176", "0.738", "0.5166", "0.738", "0.531", "0.2632", "0.532" ], [ "0.2646", "0.532", "0.738", "0.532", "0.738", "0.5454", "0.2646", "0.5454" ], [ "0.2181", "0.587", "0.2524", "0.587", "0.2524", "0.5996", "0.2181", "0.5996" ], [ "0.2443", "0.586", "0.3782", "0.587", "0.3782", "0.6016", "0.2443", "0.601" ], [ "0.2197", "0.619", "0.785", "0.619", "0.785", "0.6333", "0.2197", "0.6333" ], [ "0.2214", "0.6353", "0.7837", "0.6353", "0.7837", "0.6484", "0.2214", "0.6484" ], [ "0.2197", "0.651", "0.7837", "0.651", "0.7837", "0.664", "0.2197", "0.664" ], [ "0.2214", "0.6655", "0.785", "0.6655", "0.785", "0.6787", "0.2214", "0.6787" ], [ "0.2222", "0.6816", "0.7817", "0.6816", "0.7817", "0.6924", "0.2222", "0.6924" ], [ "0.2197", "0.696", "0.7837", "0.696", "0.7837", "0.709", "0.2197", "0.709" ], [ "0.2214", "0.712", "0.7817", "0.712", "0.7817", "0.7227", "0.2214", "0.7227" ], [ "0.2197", "0.726", "0.785", "0.726", "0.785", "0.7393", "0.2197", "0.7393" ], [ "0.2197", "0.741", "0.785", "0.741", "0.785", "0.755", "0.2197", "0.755" ], [ "0.2197", "0.756", "0.6445", "0.756", "0.6445", "0.7695", "0.2197", "0.7695" ], [ "0.2427", "0.771", "0.785", "0.7695", "0.785", "0.784", "0.2427", "0.785" ], [ "0.2197", "0.786", "0.785", "0.786", "0.785", "0.8", "0.2197", "0.8" ], [ "0.2214", "0.802", "0.785", "0.802", "0.785", "0.815", "0.2214", "0.815" ], [ "0.2181", "0.8228", "0.732", "0.8257", "0.732", "0.8433", "0.2181", "0.8403" ] ]

[ "Supervised Attribute Information Removal and", "Reconstruction for Image Manipulation", "Nannan Li and Bryan A.Plummer", "BostonUniversity", "{nnli,bplum}@bu.edu", "Abstract. The goal of attribute manipulation is to control specified at-", "tribute(s) in given images. Prior work approaches this problem by learn-", "ing disentangled representations for each attribute that enables it to", "manipulate the encoded source attributes to the target attributes. How-", "ever, encoded attributes are often correlated with relevant image content.", "Thus, thesourceattributeinformation can often behidden inthe disen-", "tangled features, leading to unwanted image editing effects. In this pa-", "per, we propose an Attribute Information Removal and Reconstruction", "(AIRR)networkthatpreventssuchinformationhidingbylearninghow", "toremovetheattributeinformationentirely,creating attribute excluded", "features, and then learns to directly inject the desired attributes in a", "reconstructed image. We evaluate our approach on four diverse datasets", "with a variety of attributes including DeepFashion Synthesis, DeepFash-", "ion Fine-grained Attribute, CelebA and CelebA-HQ, where our model", "improves attribute manipulation accuracy and top-k retrieval rate by 10", "% on average over prior work. A user study also reports that AIRR ma-", "nipulated images are preferred over prior work in up to 76 % of cases", "1", "Introduction", "属性操作根据所需属性转换图像，该操作可应用于面部编辑[？？,?]，图像检索", "[？,？?]和图像合成[?,?]等。在这些任务中，目标是能够控制指定的属性而不影", "响源图像中的其他信息。虽然生成对抗网络（GAN）在属性操作方面取得了令", "人印象深刻的性能，但一个主要挑战是生成器倾向于通过利用保留的源属性信", "息而不是自标属性来操作？？？：从而导致操作图像中的图像编辑效果不正确", "先前的工作试图通过在重建[?,?]期间添加随机噪声或学习用于操纵图像[?,？]", "的解开属性表示来解决此问题。。但是，低幅度的随机噪声不针对源属性，这可", "能会被模型故意忽略或无意中抑制关键源特征。另一方面，即使使用分离的图", "像表示，属性与相关图像内容之间的相关性也可能导致源属性信息隐藏在其余", "图像特征中。例如，礼服的正式属性通常与礼服的长度相关", "为了解决这些问题，我们提出了一种监督属性信息删除和重建（AIRR）模", "型，该模型学习排除属性的表示并重建具有所需属性的图像。关键挑战在于识", "别保留的源属性信息并将其与图像表示[？,？,？?]解相关。先前关于特征解纠缠", "1 Code and models are available at https://github.com/NannanLi999/AIRR" ]

{ "arxiv_id": "2207.06555", "language": "zh", "page_idx": 0 }

[ [ "0.2133", "0.1187", "0.2279", "0.11426", "0.2336", "0.125", "0.2181", "0.1288" ], [ "0.2695", "0.11554", "0.4182", "0.11554", "0.4182", "0.1288", "0.2695", "0.1288" ], [ "0.3186", "0.1477", "0.4495", "0.1477", "0.4495", "0.1578", "0.3186", "0.1578" ], [ "0.7017", "0.1578", "0.726", "0.1578", "0.726", "0.1622", "0.7017", "0.1622" ], [ "0.486", "0.1736", "0.505", "0.1736", "0.505", "0.1768", "0.486", "0.1768" ], [ "0.539", "0.1724", "0.589", "0.1724", "0.589", "0.1831", "0.539", "0.1831" ], [ "0.482", "0.1812", "0.511", "0.1812", "0.511", "0.1881", "0.482", "0.1881" ], [ "0.6807", "0.1818", "0.739", "0.1818", "0.739", "0.1895", "0.6807", "0.1895" ], [ "0.3096", "0.1901", "0.3594", "0.1925", "0.3594", "0.2039", "0.3096", "0.202" ], [ "0.4102", "0.1901", "0.423", "0.1901", "0.423", "0.1932", "0.4102", "0.1932" ], [ "0.669", "0.1901", "0.751", "0.1901", "0.751", "0.2008", "0.669", "0.2008" ], [ "0.4028", "0.1963", "0.4307", "0.1963", "0.4307", "0.2052", "0.4028", "0.2052" ], [ "0.3987", "0.2058", "0.4143", "0.2058", "0.4143", "0.214", "0.3987", "0.214" ], [ "0.539", "0.2008", "0.591", "0.2008", "0.591", "0.2108", "0.539", "0.2108" ], [ "0.482", "0.207", "0.5063", "0.207", "0.5063", "0.214", "0.482", "0.214" ], [ "0.4795", "0.2197", "0.5", "0.2197", "0.5", "0.2235", "0.4795", "0.2235" ], [ "0.648", "0.2153", "0.7344", "0.2166", "0.7344", "0.2267", "0.648", "0.226" ], [ "0.3186", "0.2317", "0.3513", "0.2317", "0.3513", "0.2393", "0.3186", "0.2393" ], [ "0.406", "0.2343", "0.4575", "0.2343", "0.4575", "0.238", "0.406", "0.238" ], [ "0.2197", "0.254", "0.785", "0.254", "0.785", "0.2644", "0.2197", "0.2644" ], [ "0.2181", "0.2664", "0.5", "0.2664", "0.5", "0.2798", "0.2181", "0.2798" ], [ "0.357", "0.2827", "0.482", "0.2827", "0.482", "0.2935", "0.357", "0.2935" ], [ "0.3064", "0.2998", "0.3882", "0.2998", "0.3882", "0.31", "0.3064", "0.31" ], [ "0.7534", "0.3018", "0.7695", "0.3018", "0.7695", "0.3062", "0.7534", "0.3062" ], [ "0.5884", "0.31", "0.642", "0.31", "0.642", "0.3213", "0.5884", "0.3213" ], [ "0.2754", "0.3225", "0.3342", "0.3225", "0.3342", "0.334", "0.2754", "0.334" ], [ "0.361", "0.3193", "0.4043", "0.3193", "0.4043", "0.3308", "0.361", "0.3308" ], [ "0.3594", "0.33", "0.4028", "0.33", "0.4028", "0.3416", "0.3594", "0.3416" ], [ "0.4436", "0.33", "0.4722", "0.33", "0.4722", "0.3372", "0.4436", "0.3372" ], [ "0.5884", "0.336", "0.6445", "0.336", "0.6445", "0.3467", "0.5884", "0.3467" ], [ "0.451", "0.3655", "0.492", "0.3655", "0.492", "0.3726", "0.451", "0.3726" ], [ "0.3816", "0.385", "0.625", "0.385", "0.625", "0.3984", "0.3816", "0.3984" ], [ "0.2197", "0.4124", "0.785", "0.4124", "0.785", "0.4255", "0.2197", "0.4255" ], [ "0.2181", "0.428", "0.786", "0.428", "0.786", "0.4414", "0.2181", "0.4414" ], [ "0.2181", "0.4426", "0.785", "0.4426", "0.785", "0.4558", "0.2181", "0.4558" ], [ "0.2181", "0.457", "0.786", "0.457", "0.786", "0.4702", "0.2181", "0.4702" ], [ "0.2173", "0.473", "0.785", "0.473", "0.785", "0.486", "0.2173", "0.486" ], [ "0.2181", "0.4873", "0.785", "0.4873", "0.785", "0.5005", "0.2181", "0.5005" ], [ "0.2181", "0.5024", "0.785", "0.5024", "0.785", "0.5166", "0.2181", "0.5166" ], [ "0.2173", "0.5166", "0.5435", "0.518", "0.5435", "0.532", "0.2173", "0.53" ], [ "0.2197", "0.5625", "0.7837", "0.5625", "0.7837", "0.5757", "0.2197", "0.5757" ], [ "0.2214", "0.577", "0.785", "0.577", "0.785", "0.5903", "0.2214", "0.5903" ], [ "0.2181", "0.5923", "0.785", "0.5923", "0.785", "0.606", "0.2181", "0.606" ], [ "0.2222", "0.6094", "0.7837", "0.6094", "0.7837", "0.619", "0.2222", "0.619" ], [ "0.2197", "0.6235", "0.7837", "0.6235", "0.7837", "0.634", "0.2197", "0.634" ], [ "0.2197", "0.637", "0.785", "0.637", "0.785", "0.651", "0.2197", "0.651" ], [ "0.2214", "0.653", "0.3179", "0.653", "0.3179", "0.664", "0.2214", "0.664" ], [ "0.2418", "0.6665", "0.7837", "0.6665", "0.7837", "0.6797", "0.2418", "0.6797" ], [ "0.2214", "0.6816", "0.785", "0.6816", "0.785", "0.6963", "0.2214", "0.6963" ], [ "0.2214", "0.6978", "0.7817", "0.6978", "0.7817", "0.711", "0.2214", "0.711" ], [ "0.2197", "0.7134", "0.7837", "0.7134", "0.7837", "0.7266", "0.2197", "0.7266" ], [ "0.2197", "0.728", "0.786", "0.728", "0.786", "0.741", "0.2197", "0.741" ], [ "0.2181", "0.742", "0.4631", "0.742", "0.4631", "0.756", "0.2181", "0.756" ], [ "0.2443", "0.759", "0.384", "0.759", "0.384", "0.773", "0.2443", "0.773" ], [ "0.232", "0.784", "0.2443", "0.784", "0.2443", "0.7886", "0.232", "0.7886" ], [ "0.246", "0.781", "0.786", "0.781", "0.786", "0.7944", "0.246", "0.7944" ], [ "0.246", "0.797", "0.7837", "0.797", "0.7837", "0.81", "0.246", "0.81" ], [ "0.2467", "0.811", "0.7837", "0.811", "0.7837", "0.8247", "0.2467", "0.8247" ], [ "0.246", "0.827", "0.786", "0.827", "0.786", "0.8403", "0.246", "0.8403" ] ]

[ "2", "N. Li, B. A. Plummer", "Manipulation Input:Sleveeless", "Outpu", "ttrihie", "Decoder", "Sleeveless", "ToLongSleev", "Encoder", "E", "(cycle consistency)", "Attribut", "(long", "Decoder", "ttnte", "ons", "ReconstructedOutput", "Reconst", "onpSipev", "(a)Framework of previous methods??,?,?,?]:Dashed arrowsmean two different", "ways of obtaining the reconstructed image", "Manipulation Input:Sleevelss", "Disentanglement Loss", "utou", "Decoder", "Encoder", "Attribute", "Remover", "Excluded", "Decoder", "ongSleeve", "(b) Pipeline of the proposed method", "Fig. l: In (a), the generator like those used by [?,?,?,?,?] incorrectly utilizes the", "hidden source attribute information instead of the target attribute sleeveless", "for image manipulation. As a result, the manipulated image still contains the", "source attribute long sleeve, causing improper image editing effects. To avoid", "this issue, the proposed method in (b) erases the source attribute information", "in the encoded features through an attribute remover with a disentanglement", "loss, conditioning the manipulated output only on the input target attribute", "sleeveless and the attribute excluded features", "的研究要么不考虑解相关[？？,?】，要么对它可以在前向传递[?,?]中解相关的", "属性数量有限制，这与我们可以解纠缠任意数量的属性的方法不同。此外，如图", "1a所示，这些方法通常依赖于完整的图像信息进行处理和重建。如前所述，这", "可能导致信息隐藏在操纵的图像中。相反，如图1b所示，我们使用移除器擦", "除属性信息以获得属性排除的特征，然后将其用于直接生成重建和操纵的图像。", "由于这应该消除可能隐藏在解开表示中的信息，因此我们避免了先前工作中的", "信息隐藏问题", "我们方法中的一个挑战是我们依赖于能够识别和删除真实图像中的属性信", "息。例如，尽管在图1中输入图像的背景中出现白色，但一个好的属性分类器", "会预测服装是灰色而不是白色。这意味着背景信息可能会误导属性识别。为了", "解决这个问题，我们分割感兴趣的对象（e.9.，使用[??]）将图像编码器分为", "两个分支，分别用于感兴趣的对象和背景。这有助于AIRR将操作集中在感兴", "趣的对象上，而不会影响背景信息", "我们的主要贡献是：", "", "We propose the Attribute Information Removal and Reconstruction method", "(AIRR), a controllable disentangled attribute manipulation framework that", "produces high quality images. The key insight in AIRR is the attribute in-", "formation removal and reconstruction module that produces an attribute" ]

{ "arxiv_id": "2207.06555", "language": "zh", "page_idx": 1 }

[ [ "0.6914", "0.11743", "0.786", "0.11743", "0.786", "0.1288", "0.6914", "0.1288" ], [ "0.2484", "0.1483", "0.7817", "0.1483", "0.7817", "0.1622", "0.2484", "0.1622" ], [ "0.246", "0.1635", "0.4878", "0.1635", "0.4878", "0.1768", "0.246", "0.1768" ], [ "0.2303", "0.1812", "0.2507", "0.1812", "0.2507", "0.1881", "0.2303", "0.1881" ], [ "0.246", "0.1787", "0.7817", "0.1787", "0.7817", "0.1925", "0.246", "0.1925" ], [ "0.246", "0.1945", "0.785", "0.1925", "0.785", "0.207", "0.246", "0.209" ], [ "0.2467", "0.209", "0.7837", "0.2102", "0.7837", "0.2235", "0.2467", "0.2229" ], [ "0.2467", "0.2247", "0.785", "0.2247", "0.785", "0.2393", "0.2467", "0.2393" ], [ "0.246", "0.238", "0.786", "0.2405", "0.786", "0.2551", "0.246", "0.2524" ], [ "0.246", "0.254", "0.562", "0.2551", "0.562", "0.2695", "0.246", "0.2683" ], [ "0.2263", "0.2695", "0.7837", "0.2715", "0.7837", "0.2854", "0.2263", "0.2827" ], [ "0.2467", "0.286", "0.7837", "0.286", "0.7837", "0.2998", "0.2467", "0.2998" ], [ "0.2467", "0.3018", "0.767", "0.3018", "0.767", "0.315", "0.2467", "0.315" ], [ "0.2173", "0.3435", "0.392", "0.3435", "0.392", "0.358", "0.2173", "0.358" ], [ "0.2197", "0.3774", "0.7837", "0.3774", "0.7837", "0.3914", "0.2197", "0.3914" ], [ "0.2197", "0.3933", "0.7837", "0.3933", "0.7837", "0.4072", "0.2197", "0.4072" ], [ "0.2181", "0.4072", "0.7837", "0.4077", "0.7837", "0.4224", "0.2181", "0.4216" ], [ "0.2214", "0.4248", "0.3342", "0.4248", "0.3342", "0.435", "0.2214", "0.435" ], [ "0.2214", "0.4395", "0.785", "0.4395", "0.785", "0.4526", "0.2214", "0.4526" ], [ "0.2214", "0.4539", "0.785", "0.4539", "0.785", "0.4673", "0.2214", "0.4673" ], [ "0.2197", "0.4697", "0.785", "0.4697", "0.785", "0.483", "0.2197", "0.483" ], [ "0.2181", "0.4841", "0.7837", "0.483", "0.7837", "0.4976", "0.2181", "0.4988" ], [ "0.2214", "0.5", "0.785", "0.5", "0.785", "0.513", "0.2214", "0.513" ], [ "0.2214", "0.5146", "0.785", "0.5146", "0.785", "0.528", "0.2214", "0.528" ], [ "0.2214", "0.53", "0.785", "0.53", "0.785", "0.5435", "0.2214", "0.5435" ], [ "0.2181", "0.5435", "0.652", "0.5444", "0.652", "0.558", "0.2181", "0.5566" ], [ "0.2197", "0.5645", "0.7837", "0.565", "0.7837", "0.579", "0.2197", "0.5776" ], [ "0.2197", "0.579", "0.7837", "0.58", "0.7837", "0.5933", "0.2197", "0.5923" ], [ "0.2197", "0.5947", "0.7837", "0.5947", "0.7837", "0.608", "0.2197", "0.608" ], [ "0.2214", "0.61", "0.7837", "0.61", "0.7837", "0.6235", "0.2214", "0.6235" ], [ "0.2214", "0.6255", "0.7793", "0.6255", "0.7793", "0.6396", "0.2214", "0.6396" ], [ "0.2197", "0.6416", "0.7837", "0.6416", "0.7837", "0.655", "0.2197", "0.655" ], [ "0.2181", "0.655", "0.7837", "0.655", "0.7837", "0.6685", "0.2181", "0.6685" ], [ "0.2222", "0.6724", "0.7837", "0.6724", "0.7837", "0.683", "0.2222", "0.683" ], [ "0.2197", "0.686", "0.7837", "0.686", "0.7837", "0.6997", "0.2197", "0.6997" ], [ "0.2197", "0.6997", "0.7817", "0.6997", "0.7817", "0.7134", "0.2197", "0.7134" ], [ "0.2222", "0.7173", "0.7817", "0.7173", "0.7817", "0.728", "0.2222", "0.728" ], [ "0.2214", "0.731", "0.7837", "0.731", "0.7837", "0.744", "0.2214", "0.744" ], [ "0.2197", "0.747", "0.3406", "0.747", "0.3406", "0.7603", "0.2197", "0.7603" ], [ "0.2181", "0.765", "0.7817", "0.765", "0.7817", "0.779", "0.2181", "0.779" ], [ "0.2214", "0.7827", "0.7817", "0.7827", "0.7817", "0.7935", "0.2214", "0.7935" ], [ "0.2214", "0.797", "0.7837", "0.797", "0.7837", "0.81", "0.2214", "0.81" ], [ "0.2214", "0.813", "0.7817", "0.813", "0.7817", "0.8237", "0.2214", "0.8237" ], [ "0.2197", "0.827", "0.7837", "0.827", "0.7837", "0.8403", "0.2197", "0.8403" ] ]

[ "AIRR3", "excluded representation, eliminating sources of information hiding that de", "grades performance in prior work", "二", "Extensive experiments across DeepFashion Synthesis [?] , DeepFashion Fine", "grained Attribute [?], CelebA[?] and CelebA-HQ [?] report that AIRR", "improves the attribute manipulation accuracy and top-k retrieval rate by", "10 % on average over the state-of-the-art. Moreover, we show that AIRR", "can effectively control attribute strength as well as efficiently manipulating", "multiple attributes in a single forward pass.", "-A user study further validates the effectiveness of our approach, where our", "methods are shown to produce high quality images that more accurately", "achieve the target attribute manipulation by up to 76 % over prior work.", "2Related Work", "属性操纵的早期研究[?,?]将目标属性标签直接与图像或图像特征相结合，并将", "其解码为操纵输出。但是，解码器可能会错误地将保留的源属性信息用于图像", "处理。因此，最近的工作（包括本文）集中在学习解开的属性表示，我们将在下", "面更详细地讨论", "无监督解纠缠。几项研究以无监督的方式探索了GAN潜在空间中的解纠缠", "[？,？？,？?,?]。这些方法旨在操纵合成数据上的属性，其中图像内容是随机生成的。", "在[?]中，作者发现预训练的GAN上的特征的主要组成部分代表了高级语义概", "念。在[?]中，作者介绍了StyleGAN[?]的通道式解纠缠。ShoshanEet al.[?]", "利用对比学习来解开潜在空间，从而实现对合成面部图像的显式控制。但是，如", "果不在特征空间上进行手动检查，则很难找到我们要操作的确切属性表示，尤", "其是对于具有高级语义的属性。因此，在我们的工作中，我们专注于我们希望操", "纵的属性已知的情况，使我们能够直接针对我们的特征学习。", "受监督的解纠缠。基于属性注释的监督解纠缠方法编辑真实图像。此任务的先前", "工作可以分为两种类型：空间解纠缠和特征解纠缠。在专注于空间解纠缠[?,？]", "的方法中，属性位于空间上，因此在特征图中被解纠缠。这些方法可以通过注意", "图找到特定于属性的特征，而与属性相关的信息则被隐式保留，从而影响图像", "处理。另一方面，特征解纠缠识别与操纵属性相对应的某些特征。[?]提出了一", "种学习映射StyleGAN潜在代码的线性变换函数的方法。尽管StyleGAN的潜", "在空间是解开的[？]，但没有正交约束，这种线性组合可能会导致不同图像属性", "和内容之间的相关性。为了解决这个问题，Yangetal.？学习了属性相关和不", "相关的特征，但是每个被操纵的属性都需要训练自已的模型，这在计算上是昂", "贵的。相反，ShenEetal.E[?l通过支持向量机（SVM）使用条件子空间投影", "来操纵属性，而操纵精度取决于SVM的功能，并且每个正向传递只能控制单个", "属性。相反，我们提出的方法可以通过利用注入的属性嵌人在单个前向传递中", "操纵多个源属性。", "时尚的属性操作。除了上面提到的主要应用于面部属性编辑的方法之外，时尚", "图像中的属性操纵也得到了很多关注。最近关于该主题的工作主要旨在提高项", "目推荐的图像检索准确性。例如，研究人员在操纵属性E[？,？]时利用了空间信", "息，学习了属性变换？的字典，或者将属性概率分布用作图像检索？的解开", "表示。KwonEetal.E[?]预测了由于更改属性而导致的物品形状的变化，从而" ]

{ "arxiv_id": "2207.06555", "language": "zh", "page_idx": 2 }

[ [ "0.2157", "0.11615", "0.2385", "0.11615", "0.2385", "0.1288", "0.2157", "0.1288" ], [ "0.273", "0.11743", "0.4167", "0.11743", "0.4167", "0.1276", "0.273", "0.1276" ], [ "0.517", "0.154", "0.564", "0.154", "0.564", "0.1642", "0.517", "0.1642" ], [ "0.517", "0.1642", "0.5596", "0.1642", "0.5596", "0.1755", "0.517", "0.1755" ], [ "0.5864", "0.1666", "0.5933", "0.1718", "0.5864", "0.1768", "0.58", "0.1718" ], [ "0.536", "0.1787", "0.5415", "0.1787", "0.5415", "0.1863", "0.536", "0.1863" ], [ "0.5186", "0.2008", "0.5596", "0.2008", "0.5596", "0.2122", "0.5186", "0.2122" ], [ "0.755", "0.1963", "0.7754", "0.1963", "0.7754", "0.209", "0.755", "0.209" ], [ "0.326", "0.207", "0.335", "0.207", "0.335", "0.2108", "0.326", "0.2108" ], [ "0.5986", "0.2026", "0.603", "0.2026", "0.603", "0.2102", "0.5986", "0.2102" ], [ "0.6685", "0.207", "0.701", "0.207", "0.701", "0.214", "0.6685", "0.214" ], [ "0.5186", "0.2108", "0.558", "0.2108", "0.558", "0.2229", "0.5186", "0.2229" ], [ "0.3708", "0.2084", "0.4043", "0.2084", "0.4043", "0.2172", "0.3708", "0.2172" ], [ "0.2385", "0.2153", "0.2443", "0.2153", "0.2443", "0.2203", "0.2385", "0.2203" ], [ "0.597", "0.2153", "0.607", "0.2153", "0.607", "0.2197", "0.597", "0.2197" ], [ "0.539", "0.2235", "0.5728", "0.2235", "0.5728", "0.2311", "0.539", "0.2311" ], [ "0.5933", "0.2229", "0.6445", "0.2229", "0.6445", "0.2343", "0.5933", "0.2343" ], [ "0.7573", "0.2267", "0.7686", "0.2267", "0.7686", "0.2361", "0.7573", "0.2361" ], [ "0.4395", "0.2267", "0.4739", "0.2267", "0.4739", "0.2361", "0.4395", "0.2361" ], [ "0.3228", "0.2299", "0.33", "0.2279", "0.335", "0.238", "0.3284", "0.2405" ], [ "0.3816", "0.2279", "0.4207", "0.2279", "0.4207", "0.238", "0.3816", "0.238" ], [ "0.4756", "0.2393", "0.5596", "0.2393", "0.5596", "0.2468", "0.4756", "0.2468" ], [ "0.689", "0.2349", "0.728", "0.2349", "0.728", "0.2424", "0.689", "0.2424" ], [ "0.3245", "0.2551", "0.33", "0.2551", "0.33", "0.2615", "0.3245", "0.2615" ], [ "0.581", "0.252", "0.619", "0.252", "0.619", "0.2588", "0.581", "0.2588" ], [ "0.3708", "0.2676", "0.407", "0.2695", "0.406", "0.281", "0.3694", "0.279" ], [ "0.755", "0.2727", "0.778", "0.2727", "0.778", "0.286", "0.755", "0.286" ], [ "0.3218", "0.2886", "0.3406", "0.2886", "0.3406", "0.3005", "0.3218", "0.3005" ], [ "0.38", "0.2922", "0.4182", "0.2922", "0.4182", "0.3037", "0.38", "0.3037" ], [ "0.4355", "0.2874", "0.4714", "0.2874", "0.4714", "0.2986", "0.4355", "0.2986" ], [ "0.2181", "0.3193", "0.7817", "0.3193", "0.7817", "0.3333", "0.2181", "0.3333" ], [ "0.2197", "0.3352", "0.7837", "0.3352", "0.7837", "0.3484", "0.2197", "0.3484" ], [ "0.2214", "0.3516", "0.7812", "0.3516", "0.7812", "0.3623", "0.2214", "0.3623" ], [ "0.2197", "0.3655", "0.7837", "0.3655", "0.7837", "0.379", "0.2197", "0.379" ], [ "0.2197", "0.3806", "0.785", "0.3806", "0.785", "0.3945", "0.2197", "0.3945" ], [ "0.2181", "0.3958", "0.785", "0.3945", "0.785", "0.4092", "0.2181", "0.4104" ], [ "0.2197", "0.411", "0.746", "0.411", "0.746", "0.4248", "0.2197", "0.4248" ], [ "0.2214", "0.4526", "0.7837", "0.4526", "0.7837", "0.4666", "0.2214", "0.4666" ], [ "0.2197", "0.4685", "0.774", "0.4685", "0.774", "0.4817", "0.2197", "0.4817" ], [ "0.2157", "0.5024", "0.755", "0.504", "0.755", "0.518", "0.2157", "0.5176" ], [ "0.2197", "0.5327", "0.785", "0.5327", "0.785", "0.547", "0.2197", "0.547" ], [ "0.2197", "0.5474", "0.7837", "0.5474", "0.7837", "0.561", "0.2197", "0.561" ], [ "0.2197", "0.561", "0.7837", "0.5625", "0.7837", "0.577", "0.2197", "0.5757" ], [ "0.2214", "0.5776", "0.7837", "0.5776", "0.7837", "0.5913", "0.2214", "0.5913" ], [ "0.2197", "0.5923", "0.785", "0.5923", "0.785", "0.606", "0.2197", "0.606" ], [ "0.2181", "0.608", "0.642", "0.608", "0.642", "0.622", "0.2181", "0.622" ], [ "0.2181", "0.6426", "0.3896", "0.6426", "0.3896", "0.656", "0.2181", "0.656" ], [ "0.2197", "0.664", "0.785", "0.664", "0.785", "0.678", "0.2197", "0.678" ], [ "0.2214", "0.6797", "0.785", "0.6797", "0.785", "0.6934", "0.2214", "0.6934" ], [ "0.2197", "0.6943", "0.785", "0.6943", "0.785", "0.7085", "0.2197", "0.7085" ], [ "0.2181", "0.709", "0.7817", "0.709", "0.7817", "0.7227", "0.2181", "0.7227" ], [ "0.2197", "0.726", "0.7837", "0.726", "0.7837", "0.7393", "0.2197", "0.7393" ], [ "0.2197", "0.7407", "0.7837", "0.7407", "0.7837", "0.754", "0.2197", "0.754" ], [ "0.2181", "0.754", "0.6313", "0.755", "0.6313", "0.7695", "0.2181", "0.768" ], [ "0.2181", "0.789", "0.423", "0.789", "0.423", "0.803", "0.2181", "0.803" ], [ "0.2214", "0.811", "0.785", "0.811", "0.785", "0.8247", "0.2214", "0.8247" ], [ "0.2197", "0.827", "0.785", "0.827", "0.785", "0.8403", "0.2197", "0.8403" ] ]

[ "4", "N.Li, B. A.Plummer", "Attribute", "Classifier", "7", "", "Attribute", "Irec", "", "", "Decode", "Remover", "Encoder", "A", "te", "(Sec.3.2)", "R(Ei(lob))", "技", "Ei(let)", "", "(Sec. 3.1)", "Disentanglement Loss", "Sec,3.3)", "R", "ASleeve", "Encoder", "Imap", "Ibg", "(5ec.3.1)", "E2(lbg)", "Fig. 2: AIRR framework . In the AIRR generator, a given image is parsed into ar", "object of interest Ici and background Ibg through an ofline parser [?,?] .Ic and", "Ibg are encoded in separate branches(Sec.3.1 .In the Icl branch,the source", "attribute information in the encoded features are erased by an attribute remover", "using a disentanglement loss (Sec. 3.2 ). Subsequently, the source attributes in A", "and the target attributes in A are embedded into the attribute excluded features", "for image reconstruction and image manipulation, respectively (Sec. ??)", "使他们能够对图像中的衣服进行更重大的更改。然而，这些方法中的许多方法", "也遇到了解开属性的问题，通常是由于错误信息隐藏，我们的工作将其最小化", "3AttributeInformation Removal andReconstruction", "给定图像I及其属性A={ai,a2,.,an，其中ai表示第i个属性，我们旨", "在操纵A中的任意数量的属性。为了实现这一目标，生成器首先将真实图像作", "为输入，并使用我们的属性移除器将图像属性与图像特征解相关。然后，将生成", "的属性排除表示与目标属性嵌人相结合，以生成操作输出Imap。下面，我们介", "绍模型的四个组成部分：图像编码器（3.1节），属性移除器（3.2节），解码器", "（？？节）和学习目标（？？节）。图2显示了我们方法的概述", "3.1 Image Encoder", "为了将操作集中在我们要操作的对象上，AIRR中的图像编码器分为两个分支，", "分别用于感兴趣的对象Icl和背景Ibg。先前的工作通常通过学习注意力图", "[?,?,?]来实现Icl的分割，而我们根据经验发现，使用离线解析器[?,?]在分割", "实例时更准确。这种分割对于具有多个对象的图像特别有用，e.9.是穿着上衣", "绑腿和靴子的时装模特的图像。如图2所示，在获得Icl和Ibg之后，图像编码", "器在第一个分支中将Icl编码为Ei(Icl)，在第二个分支中将Ibg编码为E2(Ibg）", "。后来，AIRR仅操作Ei(Ict）而不影响背景信息E2(Ibg)", "3.2AttributeRemover", "虽然先前的工作直接使用图像特征Ei(Icl）来生成图像[?,？?？?,?，但在AIRR", "中，来自我们基本编码器的图像特征Ei（Icl）被馈送到属性移除器中，以学习排" ]

{ "arxiv_id": "2207.06555", "language": "zh", "page_idx": 3 }

[ [ "0.1184", "0.0873", "0.4878", "0.0873", "0.4878", "0.10095", "0.1184", "0.10095" ], [ "0.511", "0.0849", "0.881", "0.0861", "0.881", "0.10095", "0.511", "0.0998" ], [ "0.1184", "0.1027", "0.3208", "0.1027", "0.3208", "0.1164", "0.1184", "0.1164" ], [ "0.5146", "0.10394", "0.5767", "0.10394", "0.5767", "0.1152", "0.5146", "0.1152" ], [ "0.1368", "0.11816", "0.476", "0.11816", "0.476", "0.1318", "0.1368", "0.1318" ], [ "0.531", "0.1247", "0.76", "0.1247", "0.76", "0.1395", "0.531", "0.1395" ], [ "0.1184", "0.1354", "0.4844", "0.1354", "0.4844", "0.1484", "0.1184", "0.1484" ], [ "0.5283", "0.1395", "0.7095", "0.1414", "0.7095", "0.155", "0.5283", "0.1526" ], [ "0.1184", "0.152", "0.486", "0.152", "0.486", "0.165", "0.1184", "0.165" ], [ "0.1184", "0.1675", "0.4844", "0.1675", "0.4844", "0.1812", "0.1184", "0.1812" ], [ "0.5146", "0.1758", "0.5615", "0.1758", "0.5615", "0.1859", "0.5146", "0.1859" ], [ "0.5576", "0.1746", "0.81", "0.1746", "0.81", "0.1882", "0.5576", "0.1882" ], [ "0.1184", "0.1818", "0.4836", "0.1818", "0.4836", "0.1953", "0.1184", "0.1953" ], [ "0.1184", "0.1996", "0.4844", "0.1996", "0.4844", "0.2131", "0.1184", "0.2131" ], [ "0.513", "0.1953", "0.881", "0.1953", "0.881", "0.209", "0.513", "0.209" ], [ "0.1184", "0.2156", "0.486", "0.2156", "0.486", "0.2286", "0.1184", "0.2286" ], [ "0.511", "0.2114", "0.881", "0.2114", "0.881", "0.2245", "0.511", "0.2245" ], [ "0.1184", "0.2328", "0.4836", "0.2328", "0.4836", "0.2434", "0.1184", "0.2434" ], [ "0.513", "0.228", "0.881", "0.228", "0.881", "0.2411", "0.513", "0.2411" ], [ "0.1184", "0.2477", "0.486", "0.2477", "0.486", "0.2612", "0.1184", "0.2612" ], [ "0.513", "0.2434", "0.881", "0.2434", "0.881", "0.257", "0.513", "0.257" ], [ "0.11755", "0.2632", "0.4878", "0.2625", "0.4878", "0.2766", "0.11755", "0.2778" ], [ "0.513", "0.26", "0.8364", "0.26", "0.8364", "0.2737", "0.513", "0.2737" ], [ "0.11755", "0.279", "0.4836", "0.279", "0.4836", "0.2922", "0.11755", "0.2922" ], [ "0.513", "0.2893", "0.55", "0.2893", "0.55", "0.2986", "0.513", "0.2986" ], [ "0.544", "0.2861", "0.7617", "0.288", "0.7617", "0.3027", "0.544", "0.3005" ], [ "0.1184", "0.2957", "0.486", "0.2957", "0.486", "0.3088", "0.1184", "0.3088" ], [ "0.1184", "0.3123", "0.4836", "0.3123", "0.4836", "0.3225", "0.1184", "0.3225" ], [ "0.513", "0.3135", "0.881", "0.3135", "0.881", "0.3267", "0.513", "0.3267" ], [ "0.1184", "0.328", "0.4844", "0.328", "0.4844", "0.3416", "0.1184", "0.3416" ], [ "0.5146", "0.328", "0.879", "0.328", "0.879", "0.3416", "0.5146", "0.3416" ], [ "0.12006", "0.3433", "0.4878", "0.3433", "0.4878", "0.357", "0.12006", "0.357" ], [ "0.511", "0.3433", "0.879", "0.3445", "0.879", "0.3594", "0.511", "0.3582" ], [ "0.1184", "0.3599", "0.4844", "0.3599", "0.4844", "0.3735", "0.1184", "0.3735" ], [ "0.51", "0.3599", "0.879", "0.3599", "0.879", "0.3735", "0.51", "0.3735" ], [ "0.1184", "0.3772", "0.4844", "0.3772", "0.4844", "0.3872", "0.1184", "0.3872" ], [ "0.51", "0.376", "0.575", "0.3777", "0.574", "0.3926", "0.509", "0.3901" ], [ "0.12006", "0.3943", "0.4836", "0.3943", "0.4836", "0.405", "0.12006", "0.405" ], [ "0.5146", "0.405", "0.5557", "0.405", "0.5557", "0.4163", "0.5146", "0.4163" ], [ "0.555", "0.4026", "0.81", "0.4038", "0.81", "0.4187", "0.555", "0.4175" ], [ "0.1184", "0.4092", "0.4844", "0.4092", "0.4844", "0.4229", "0.1184", "0.4229" ], [ "0.552", "0.4187", "0.6406", "0.4204", "0.6406", "0.4353", "0.552", "0.4329" ], [ "0.11755", "0.4246", "0.4844", "0.4246", "0.4844", "0.4382", "0.11755", "0.4382" ], [ "0.11755", "0.4412", "0.3997", "0.4412", "0.3997", "0.4548", "0.11755", "0.4548" ], [ "0.511", "0.4395", "0.881", "0.4407", "0.881", "0.4548", "0.511", "0.4536" ], [ "0.513", "0.4573", "0.881", "0.4573", "0.881", "0.4702", "0.513", "0.4702" ], [ "0.1377", "0.4673", "0.4878", "0.4673", "0.4878", "0.481", "0.1377", "0.481" ], [ "0.5146", "0.4727", "0.882", "0.4727", "0.882", "0.4863", "0.5146", "0.4863" ], [ "0.1545", "0.484", "0.486", "0.484", "0.486", "0.4976", "0.1545", "0.4976" ], [ "0.513", "0.4893", "0.882", "0.4893", "0.882", "0.503", "0.513", "0.503" ], [ "0.1528", "0.4995", "0.4878", "0.4988", "0.4878", "0.513", "0.1528", "0.514" ], [ "0.513", "0.505", "0.6533", "0.505", "0.6533", "0.5186", "0.513", "0.5186" ], [ "0.1511", "0.514", "0.486", "0.5156", "0.486", "0.53", "0.1511", "0.5293" ], [ "0.531", "0.524", "0.8623", "0.524", "0.8623", "0.538", "0.531", "0.538" ], [ "0.1528", "0.533", "0.2854", "0.533", "0.2854", "0.5464", "0.1528", "0.5464" ], [ "0.531", "0.5405", "0.86", "0.5405", "0.86", "0.5547", "0.531", "0.5547" ], [ "0.1377", "0.557", "0.486", "0.557", "0.486", "0.572", "0.1377", "0.572" ], [ "0.531", "0.56", "0.7617", "0.56", "0.7617", "0.5737", "0.531", "0.5737" ], [ "0.1545", "0.5747", "0.4844", "0.5747", "0.4844", "0.589", "0.1545", "0.589" ], [ "0.532", "0.576", "0.574", "0.576", "0.574", "0.5874", "0.532", "0.5874" ], [ "0.6055", "0.5747", "0.692", "0.5747", "0.692", "0.586", "0.6055", "0.586" ], [ "0.771", "0.576", "0.81", "0.576", "0.81", "0.5874", "0.771", "0.5874" ], [ "0.8447", "0.5767", "0.8613", "0.5767", "0.8613", "0.5874", "0.8447", "0.5874" ], [ "0.534", "0.5884", "0.806", "0.589", "0.806", "0.604", "0.534", "0.6025" ], [ "0.7188", "0.5776", "0.748", "0.5776", "0.748", "0.586", "0.7188", "0.586" ], [ "0.1528", "0.5903", "0.3972", "0.5903", "0.3972", "0.605", "0.1528", "0.605" ], [ "0.1377", "0.6177", "0.486", "0.6177", "0.486", "0.6313", "0.1377", "0.6313" ], [ "0.531", "0.616", "0.881", "0.616", "0.881", "0.6313", "0.531", "0.6313" ], [ "0.1528", "0.633", "0.4895", "0.633", "0.4895", "0.6465", "0.1528", "0.6465" ], [ "0.513", "0.633", "0.8774", "0.633", "0.8774", "0.6465", "0.513", "0.6465" ], [ "0.1545", "0.6494", "0.4902", "0.6494", "0.4902", "0.663", "0.1545", "0.663" ], [ "0.51", "0.6484", "0.875", "0.6484", "0.875", "0.662", "0.51", "0.662" ], [ "0.1545", "0.6665", "0.4895", "0.6665", "0.4895", "0.6797", "0.1545", "0.6797" ], [ "0.511", "0.6675", "0.873", "0.6675", "0.873", "0.681", "0.511", "0.681" ], [ "0.1545", "0.6816", "0.4836", "0.6816", "0.4836", "0.6953", "0.1545", "0.6953" ], [ "0.529", "0.684", "0.881", "0.684", "0.881", "0.6978", "0.529", "0.6978" ], [ "0.1545", "0.6987", "0.215", "0.6987", "0.215", "0.7134", "0.1545", "0.7134" ], [ "0.5146", "0.7007", "0.881", "0.7007", "0.881", "0.7144", "0.5146", "0.7144" ], [ "0.5146", "0.7173", "0.54", "0.7173", "0.54", "0.73", "0.5146", "0.73" ], [ "0.1368", "0.7246", "0.4844", "0.7246", "0.4844", "0.7383", "0.1368", "0.7383" ], [ "0.1184", "0.7407", "0.4878", "0.7407", "0.4878", "0.7544", "0.1184", "0.7544" ], [ "0.5356", "0.745", "0.882", "0.745", "0.882", "0.7603", "0.5356", "0.7603" ], [ "0.1142", "0.756", "0.2688", "0.756", "0.2688", "0.7695", "0.1142", "0.7695" ], [ "0.513", "0.775", "0.775", "0.775", "0.775", "0.788", "0.513", "0.788" ], [ "0.11755", "0.7837", "0.1511", "0.7837", "0.1511", "0.7964", "0.11755", "0.7964" ], [ "0.1453", "0.785", "0.2603", "0.785", "0.2603", "0.7974", "0.1453", "0.7974" ], [ "0.513", "0.8003", "0.864", "0.8003", "0.864", "0.814", "0.513", "0.814" ], [ "0.1184", "0.8086", "0.4114", "0.8086", "0.4114", "0.8223", "0.1184", "0.8223" ], [ "0.5146", "0.822", "0.881", "0.822", "0.881", "0.835", "0.5146", "0.835" ], [ "0.11755", "0.836", "0.3997", "0.836", "0.3997", "0.8496", "0.11755", "0.8496" ], [ "0.5146", "0.8374", "0.879", "0.8374", "0.879", "0.851", "0.5146", "0.851" ], [ "0.1184", "0.857", "0.4895", "0.857", "0.4895", "0.8706", "0.1184", "0.8706" ], [ "0.513", "0.854", "0.623", "0.854", "0.623", "0.865", "0.513", "0.865" ], [ "0.11755", "0.8726", "0.4878", "0.8716", "0.4878", "0.8857", "0.11755", "0.887" ], [ "0.5283", "0.868", "0.881", "0.869", "0.881", "0.8843", "0.5283", "0.883" ], [ "0.513", "0.876", "0.523", "0.876", "0.523", "0.879", "0.513", "0.879" ], [ "0.11755", "0.8896", "0.4844", "0.8896", "0.4844", "0.903", "0.11755", "0.903" ], [ "0.5054", "0.887", "0.879", "0.887", "0.879", "0.901", "0.5054", "0.901" ], [ "0.1184", "0.905", "0.4878", "0.905", "0.4878", "0.9185", "0.1184", "0.9185" ], [ "0.513", "0.904", "0.879", "0.904", "0.879", "0.9175", "0.513", "0.9175" ] ]

[ "语CoT，如图1（b）所示。不幸的是，很少有", "们一步一步地思考！”以输出特定的推理路径，", "人关注零射击跨语言CoT", "表示为：", "为了跨语言概括当前的CoT，我们提出了", "Request: [Given sentence X ]", "种新颖的跨语言提示（CLP），旨在有效地弥合", "Let's think step by step!", "不同语言之间的鸿沟。它由两个组件组成：（1）", "跨语言对齐提示和（2）特定于任务的求解器提", "2.2", "Cross-lingual Chain-of-Thought", "示。具体来说，跨语言对齐提示用于对齐不同", "语言之间的表示。在我们的实验中，我们使用", "虽然传统CoT取得了显著的成功，但它仅限于", "“让我们一步一步地理解英语中的任务”，而不", "在单一语言中生成CoT，并且缺乏有效的跨语", "是传统的“让我们一步一步地思考”。”。固有的", "言可移植性。因此，跨语言CoT旨在使模型能", "直觉是，随着模型逐渐理解英语任务，它固有", "够处理任何语言的请求，并专门以目标语言生", "地抓住了源语言和英语之间的关系。在对齐不", "成CoT (i。e。,English) (Shiet al., 2022)", "同语言之间的表示之后，我们进一步利用特定", "3", "Cross-lingual Prompting", "于任务的解决提示，通过设置“让我们逐步解", "决您上面理解的任务！”来完成最终任务。这", "为了引出LLM的跨语言推理能力，我们引入", "种简单而有效的CLP可以大大提高跨语言场", "了跨语言提示（CLP）作为解决方案。具体来", "景的推理能力。此外，受自洽工作的启发，我", "说，CLP由两个组件组成：（1）跨语言对齐提", "们提出了跨语言自洽提示（CLSP），该模型使", "示（$3.1）和（2）特定于任务的求解器提示", "模型能够整合跨语言的推理路径的不同视图。", "($3.2 )。", "实验结果表明：CLP通过以超过1的增益优于", "3.1", "Step 1: Cross-lingual Alignment", "所有基线来实现SOTA性能。8%。此外，CLSP", "Prompting", "可以通过整合不同语言的知识来进一步提高性", "能。这项工作的主要贡献总结如下：", "跨语言对齐是跨语言迁移的核心挑战。因此，", "为了更好地捕获对齐信息，我们首先引入跨语", ": We introduce cross-lingual prompting that", "言对齐提示（请参阅图E2（a)）。具体来说，我", "contains cross-lingual alignment prompting", "们的方法通过对齐信息的特定任务启动LLM。", "and task-specific solver prompting,which", "请求的格式如下：", "jointly improve zero-shot CoT reasoning", "Please act as an expert in multi-lingual", "across languages;", "understanding in [Source Language L, ]", "• We further propose cross-lingual self-", "Request: [Given sentence X ]", "consistent prompting to integrate reasoning", "Let's", "understand", "task", "in", "[Target Language Lt ] step-by-step!", "the", "paths across different languages;", ": Extensive evaluations on several benchmarks", "给定sentenceX，我们首先模拟LLM在多", "reveal that both CLP and CLSP are capable", "语言理解方面的专业知识。此外，我们介绍了", "of improving zero-shot cross-lingual CoT ef-", "从source language Ls到 targetlanguageLt", "fectively and achieving SOTA performance", "的逐步对齐过程。中间语义对齐表示为「ai", "(with over 1.8 % improvement on AVG ac", "其中S表示对齐步骤的数量。总的来说，我", "curacy).", "们的跨语言对齐提示方法的公式可以表示如", "下:", "我们希望这项工作可以激发跨语言CoT的", "进一步研究，并且代码可以在Cross-Lingual-", "A= argmaxp(a1,...,as|X,Ls, Lt), (1)", "Prompting上找到。", "其中A表示步骤1中的对齐响应", "2", "Background", "3.2Step 2: Task-specific Solver Prompting", "本节介绍传统和跨语言思想链的定义", "在实现跨语言对齐之后，我们进一步提出了特", "2.1 Traditional Chain-of-Thought", "定于任务的求解器提示，以促进多语言设置中", "Chan-of-thought是一种强大的技术，可以激发", "的多步骤推理", "大型语言模型（LLM）的强大推理能力，能够", "具体来说，给定targetlanguageLt和从上", "", "完成复杂的任务。对于传统的思想链（CoT）生", "一步获得的对齐文本A，我们提示LLM参与", "成方法，LLM被附加为一个简单的提示“让我", "解析targettastT。并且LLM尝试沿着多步" ]

{ "arxiv_id": "2310.14799", "language": "zh", "page_idx": 2 }

[ [ "0.12177", "0.0831", "0.1721", "0.0831", "0.1721", "0.0974", "0.12177", "0.0974" ], [ "0.571", "0.082", "0.594", "0.0873", "0.5884", "0.10095", "0.565", "0.0956" ], [ "0.831", "0.0831", "0.881", "0.0831", "0.881", "0.0974", "0.831", "0.0974" ], [ "0.4324", "0.08435", "0.4575", "0.08435", "0.4575", "0.0974", "0.4324", "0.0974" ], [ "0.3855", "0.08435", "0.4106", "0.08435", "0.4106", "0.0974", "0.3855", "0.0974" ], [ "0.523", "0.08435", "0.546", "0.08435", "0.546", "0.0974", "0.523", "0.0974" ], [ "0.706", "0.08435", "0.7305", "0.08435", "0.7305", "0.0974", "0.706", "0.0974" ], [ "0.7515", "0.08435", "0.7734", "0.08435", "0.7734", "0.0974", "0.7515", "0.0974" ], [ "0.797", "0.08435", "0.8213", "0.08435", "0.8213", "0.0968", "0.797", "0.0968" ], [ "0.614", "0.0849", "0.6396", "0.0849", "0.6396", "0.0974", "0.614", "0.0974" ], [ "0.4785", "0.0849", "0.5015", "0.0849", "0.5015", "0.0974", "0.4785", "0.0974" ], [ "0.659", "0.0849", "0.687", "0.0849", "0.687", "0.0974", "0.659", "0.0974" ], [ "0.12427", "0.0974", "0.2913", "0.0974", "0.2913", "0.111", "0.12427", "0.111" ], [ "0.4265", "0.111", "0.4626", "0.111", "0.4626", "0.1271", "0.4265", "0.1271" ], [ "0.471", "0.111", "0.509", "0.111", "0.509", "0.1271", "0.471", "0.1271" ], [ "0.559", "0.111", "0.5996", "0.111", "0.5996", "0.1271", "0.559", "0.1271" ], [ "0.7925", "0.111", "0.825", "0.1134", "0.8237", "0.1259", "0.791", "0.12354" ], [ "0.839", "0.111", "0.875", "0.111", "0.875", "0.1271", "0.839", "0.1271" ], [ "0.3855", "0.11224", "0.4114", "0.11224", "0.4114", "0.1247", "0.3855", "0.1247" ], [ "0.126", "0.11224", "0.294", "0.111", "0.294", "0.1247", "0.126", "0.1259" ], [ "0.519", "0.11224", "0.553", "0.11224", "0.553", "0.1247", "0.519", "0.1247" ], [ "0.608", "0.11224", "0.644", "0.11224", "0.644", "0.1247", "0.608", "0.1247" ], [ "0.657", "0.11224", "0.686", "0.11224", "0.686", "0.1259", "0.657", "0.1259" ], [ "0.7036", "0.11224", "0.7314", "0.11224", "0.7314", "0.1259", "0.7036", "0.1259" ], [ "0.75", "0.11224", "0.777", "0.11224", "0.777", "0.1259", "0.75", "0.1259" ], [ "0.6543", "0.1259", "0.6885", "0.1259", "0.6885", "0.1414", "0.6543", "0.1414" ], [ "0.744", "0.12354", "0.7827", "0.1271", "0.7793", "0.1414", "0.7407", "0.1383" ], [ "0.79", "0.1259", "0.827", "0.1259", "0.827", "0.1414", "0.79", "0.1414" ], [ "0.837", "0.1259", "0.873", "0.1259", "0.873", "0.1414", "0.837", "0.1414" ], [ "0.3855", "0.1271", "0.4114", "0.1271", "0.4114", "0.1401", "0.3855", "0.1401" ], [ "0.12427", "0.1271", "0.3494", "0.1271", "0.3494", "0.1401", "0.12427", "0.1401" ], [ "0.4265", "0.1271", "0.4626", "0.1271", "0.4626", "0.1395", "0.4265", "0.1395" ], [ "0.471", "0.1271", "0.5073", "0.1271", "0.5073", "0.1395", "0.471", "0.1395" ], [ "0.519", "0.1271", "0.553", "0.1271", "0.553", "0.1395", "0.519", "0.1395" ], [ "0.5615", "0.1271", "0.5996", "0.1271", "0.5996", "0.1395", "0.5615", "0.1395" ], [ "0.6094", "0.1271", "0.644", "0.1271", "0.644", "0.1395", "0.6094", "0.1395" ], [ "0.702", "0.1271", "0.733", "0.1271", "0.733", "0.1414", "0.702", "0.1414" ], [ "0.4248", "0.1395", "0.4626", "0.1395", "0.4626", "0.155", "0.4248", "0.155" ], [ "0.471", "0.1395", "0.509", "0.1395", "0.509", "0.155", "0.471", "0.155" ], [ "0.517", "0.1395", "0.555", "0.1395", "0.555", "0.155", "0.517", "0.155" ], [ "0.5615", "0.1395", "0.601", "0.1395", "0.601", "0.155", "0.5615", "0.155" ], [ "0.608", "0.1395", "0.6455", "0.1395", "0.6455", "0.155", "0.608", "0.155" ], [ "0.744", "0.1395", "0.7827", "0.1395", "0.7827", "0.155", "0.744", "0.155" ], [ "0.79", "0.1395", "0.825", "0.1395", "0.825", "0.155", "0.79", "0.155" ], [ "0.8364", "0.1395", "0.873", "0.1395", "0.873", "0.155", "0.8364", "0.155" ], [ "0.3855", "0.1401", "0.4114", "0.1401", "0.4114", "0.1538", "0.3855", "0.1538" ], [ "0.126", "0.1401", "0.3123", "0.1401", "0.3123", "0.1538", "0.126", "0.1538" ], [ "0.6543", "0.1401", "0.687", "0.1401", "0.687", "0.1526", "0.6543", "0.1526" ], [ "0.7036", "0.1401", "0.7305", "0.1401", "0.7305", "0.1538", "0.7036", "0.1538" ], [ "0.12427", "0.155", "0.3687", "0.155", "0.3687", "0.1687", "0.12427", "0.1687" ], [ "0.3804", "0.155", "0.4165", "0.155", "0.4165", "0.1675", "0.3804", "0.1675" ], [ "0.4282", "0.155", "0.461", "0.155", "0.461", "0.1675", "0.4282", "0.1675" ], [ "0.471", "0.155", "0.5073", "0.155", "0.5073", "0.1675", "0.471", "0.1675" ], [ "0.519", "0.155", "0.553", "0.155", "0.553", "0.1675", "0.519", "0.1675" ], [ "0.5635", "0.155", "0.5977", "0.155", "0.5977", "0.1675", "0.5635", "0.1675" ], [ "0.608", "0.155", "0.644", "0.155", "0.644", "0.1675", "0.608", "0.1675" ], [ "0.6543", "0.155", "0.69", "0.155", "0.69", "0.1675", "0.6543", "0.1675" ], [ "0.7", "0.155", "0.735", "0.155", "0.735", "0.1675", "0.7", "0.1675" ], [ "0.7446", "0.155", "0.7793", "0.155", "0.7793", "0.1675", "0.7446", "0.1675" ], [ "0.791", "0.155", "0.8237", "0.155", "0.8237", "0.1675", "0.791", "0.1675" ], [ "0.8364", "0.1538", "0.876", "0.1538", "0.876", "0.1692", "0.8364", "0.1692" ], [ "0.1234", "0.1687", "0.2065", "0.1687", "0.2065", "0.1799", "0.1234", "0.1799" ], [ "0.5205", "0.1812", "0.553", "0.1829", "0.552", "0.1965", "0.519", "0.1948" ], [ "0.7466", "0.1818", "0.7793", "0.1841", "0.7773", "0.1965", "0.7446", "0.1948" ], [ "0.837", "0.1812", "0.876", "0.1841", "0.873", "0.1978", "0.8345", "0.1948" ], [ "0.4265", "0.1829", "0.461", "0.1829", "0.461", "0.1953", "0.4265", "0.1953" ], [ "0.126", "0.1829", "0.2957", "0.1829", "0.2957", "0.1965", "0.126", "0.1965" ], [ "0.4727", "0.1829", "0.5073", "0.1829", "0.5073", "0.1953", "0.4727", "0.1953" ], [ "0.5635", "0.1829", "0.5996", "0.1829", "0.5996", "0.1953", "0.5635", "0.1953" ], [ "0.6094", "0.1829", "0.644", "0.1829", "0.644", "0.1953", "0.6094", "0.1953" ], [ "0.656", "0.1829", "0.6885", "0.1829", "0.6885", "0.1953", "0.656", "0.1953" ], [ "0.6987", "0.1829", "0.733", "0.1829", "0.733", "0.1953", "0.6987", "0.1953" ], [ "0.7925", "0.1829", "0.825", "0.1829", "0.825", "0.1953", "0.7925", "0.1953" ], [ "0.382", "0.1936", "0.419", "0.1965", "0.4165", "0.2114", "0.3787", "0.2079" ], [ "0.3838", "0.1841", "0.4165", "0.1841", "0.4165", "0.1965", "0.3838", "0.1965" ], [ "0.561", "0.1936", "0.601", "0.1965", "0.5977", "0.2114", "0.559", "0.209" ], [ "0.608", "0.1948", "0.6465", "0.1978", "0.644", "0.212", "0.6055", "0.209" ], [ "0.6533", "0.1948", "0.692", "0.1978", "0.6885", "0.2114", "0.65", "0.2079" ], [ "0.7446", "0.1936", "0.7793", "0.1953", "0.7773", "0.2114", "0.742", "0.209" ], [ "0.7866", "0.1948", "0.827", "0.1965", "0.825", "0.2114", "0.785", "0.209" ], [ "0.6978", "0.1953", "0.735", "0.1953", "0.735", "0.2114", "0.6978", "0.2114" ], [ "0.4248", "0.1953", "0.4626", "0.1953", "0.4626", "0.2114", "0.4248", "0.2114" ], [ "0.837", "0.1953", "0.873", "0.1953", "0.873", "0.2114", "0.837", "0.2114" ], [ "0.4727", "0.1965", "0.5073", "0.1965", "0.5073", "0.209", "0.4727", "0.209" ], [ "0.12427", "0.1965", "0.3494", "0.1965", "0.3494", "0.2102", "0.12427", "0.2102" ], [ "0.519", "0.1965", "0.553", "0.1965", "0.553", "0.209", "0.519", "0.209" ], [ "0.608", "0.2079", "0.6465", "0.2114", "0.644", "0.2257", "0.6055", "0.2227" ], [ "0.8364", "0.2079", "0.875", "0.2114", "0.8716", "0.2257", "0.833", "0.2227" ], [ "0.4248", "0.2102", "0.4634", "0.2102", "0.4634", "0.2257", "0.4248", "0.2257" ], [ "0.3804", "0.2102", "0.4182", "0.2102", "0.4182", "0.2257", "0.3804", "0.2257" ], [ "0.5615", "0.2102", "0.601", "0.2102", "0.601", "0.2257", "0.5615", "0.2257" ], [ "0.6533", "0.2102", "0.692", "0.2102", "0.692", "0.2257", "0.6533", "0.2257" ], [ "0.6978", "0.2102", "0.735", "0.2102", "0.735", "0.2257", "0.6978", "0.2257" ], [ "0.744", "0.2102", "0.781", "0.2102", "0.781", "0.2257", "0.744", "0.2257" ], [ "0.4727", "0.2114", "0.5073", "0.2114", "0.5073", "0.2239", "0.4727", "0.2239" ], [ "0.126", "0.2114", "0.3105", "0.2114", "0.3105", "0.2245", "0.126", "0.2245" ], [ "0.519", "0.2114", "0.553", "0.2114", "0.553", "0.2239", "0.519", "0.2239" ], [ "0.79", "0.2114", "0.825", "0.2114", "0.825", "0.2239", "0.79", "0.2239" ], [ "0.424", "0.2239", "0.4634", "0.2239", "0.4634", "0.2393", "0.424", "0.2393" ], [ "0.471", "0.2239", "0.509", "0.2239", "0.509", "0.2393", "0.471", "0.2393" ], [ "0.517", "0.2239", "0.555", "0.2239", "0.555", "0.2393", "0.517", "0.2393" ], [ "0.561", "0.2239", "0.601", "0.2239", "0.601", "0.2393", "0.561", "0.2393" ], [ "0.607", "0.2239", "0.6455", "0.2239", "0.6455", "0.2393", "0.607", "0.2393" ], [ "0.6533", "0.2239", "0.69", "0.2239", "0.69", "0.2393", "0.6533", "0.2393" ], [ "0.6978", "0.2239", "0.7363", "0.2239", "0.7363", "0.2393", "0.6978", "0.2393" ], [ "0.744", "0.2239", "0.7827", "0.2239", "0.7827", "0.2393", "0.744", "0.2393" ], [ "0.7866", "0.2227", "0.8286", "0.2245", "0.827", "0.2393", "0.785", "0.2369" ], [ "0.8364", "0.2239", "0.8774", "0.2239", "0.8774", "0.2393", "0.8364", "0.2393" ], [ "0.382", "0.2257", "0.4165", "0.2257", "0.4165", "0.2382", "0.382", "0.2382" ], [ "0.126", "0.2245", "0.3687", "0.2245", "0.3687", "0.2382", "0.126", "0.2382" ], [ "0.1234", "0.2382", "0.2778", "0.2393", "0.2778", "0.2542", "0.1234", "0.253" ], [ "0.382", "0.2542", "0.4165", "0.2542", "0.4165", "0.2666", "0.382", "0.2666" ], [ "0.4265", "0.2542", "0.461", "0.2542", "0.461", "0.2666", "0.4265", "0.2666" ], [ "0.519", "0.2542", "0.553", "0.2542", "0.553", "0.2666", "0.519", "0.2666" ], [ "0.5615", "0.2542", "0.5977", "0.2542", "0.5977", "0.2666", "0.5615", "0.2666" ], [ "0.6094", "0.2542", "0.644", "0.2542", "0.644", "0.2666", "0.6094", "0.2666" ], [ "0.6533", "0.2505", "0.69", "0.2542", "0.687", "0.2683", "0.65", "0.2654" ], [ "0.7036", "0.2542", "0.7314", "0.2542", "0.7314", "0.2666", "0.7036", "0.2666" ], [ "0.7466", "0.2542", "0.7793", "0.2542", "0.7793", "0.2666", "0.7466", "0.2666" ], [ "0.471", "0.2546", "0.5054", "0.253", "0.5073", "0.2654", "0.4727", "0.2673" ], [ "0.1276", "0.2546", "0.1838", "0.2546", "0.1838", "0.2666", "0.1276", "0.2666" ], [ "0.791", "0.2546", "0.825", "0.2546", "0.825", "0.2666", "0.791", "0.2666" ], [ "0.839", "0.2546", "0.873", "0.2546", "0.873", "0.2673", "0.839", "0.2673" ], [ "0.3804", "0.2666", "0.4182", "0.2666", "0.4182", "0.282", "0.3804", "0.282" ], [ "0.4224", "0.2666", "0.4626", "0.2666", "0.4626", "0.282", "0.4224", "0.282" ], [ "0.517", "0.2666", "0.555", "0.2666", "0.555", "0.282", "0.517", "0.282" ], [ "0.8364", "0.2673", "0.875", "0.2673", "0.875", "0.2832", "0.8364", "0.2832" ], [ "0.5615", "0.2683", "0.5977", "0.2683", "0.5977", "0.2808", "0.5615", "0.2808" ], [ "0.4727", "0.2683", "0.5073", "0.2683", "0.5073", "0.2808", "0.4727", "0.2808" ], [ "0.6094", "0.2683", "0.644", "0.2683", "0.644", "0.2808", "0.6094", "0.2808" ], [ "0.6514", "0.2683", "0.6885", "0.2683", "0.6885", "0.2808", "0.6514", "0.2808" ], [ "0.6987", "0.2683", "0.735", "0.2683", "0.735", "0.2808", "0.6987", "0.2808" ], [ "0.7446", "0.2683", "0.7793", "0.2683", "0.7793", "0.2808", "0.7446", "0.2808" ], [ "0.79", "0.2683", "0.825", "0.2683", "0.825", "0.2808", "0.79", "0.2808" ], [ "0.3804", "0.2798", "0.4182", "0.2798", "0.4182", "0.2957", "0.3804", "0.2957" ], [ "0.126", "0.2695", "0.2225", "0.2695", "0.2225", "0.2808", "0.126", "0.2808" ], [ "0.4265", "0.2808", "0.4626", "0.2808", "0.4626", "0.2964", "0.4265", "0.2964" ], [ "0.79", "0.2808", "0.825", "0.2808", "0.825", "0.2964", "0.79", "0.2964" ], [ "0.8364", "0.2808", "0.875", "0.2808", "0.875", "0.2964", "0.8364", "0.2964" ], [ "0.519", "0.282", "0.553", "0.282", "0.553", "0.2944", "0.519", "0.2944" ], [ "0.4727", "0.282", "0.5073", "0.282", "0.5073", "0.2944", "0.4727", "0.2944" ], [ "0.5635", "0.282", "0.5977", "0.282", "0.5977", "0.2944", "0.5635", "0.2944" ], [ "0.6094", "0.282", "0.644", "0.282", "0.644", "0.2944", "0.6094", "0.2944" ], [ "0.6543", "0.282", "0.6885", "0.282", "0.6885", "0.2944", "0.6543", "0.2944" ], [ "0.6987", "0.282", "0.735", "0.282", "0.735", "0.2944", "0.6987", "0.2944" ], [ "0.7446", "0.282", "0.781", "0.282", "0.781", "0.2944", "0.7446", "0.2944" ], [ "0.8364", "0.2935", "0.876", "0.2964", "0.873", "0.3113", "0.8345", "0.3088" ], [ "0.126", "0.2832", "0.1855", "0.2832", "0.1855", "0.2944", "0.126", "0.2944" ], [ "0.3804", "0.2944", "0.4182", "0.2944", "0.4182", "0.31", "0.3804", "0.31" ], [ "0.4265", "0.2944", "0.4626", "0.2944", "0.4626", "0.31", "0.4265", "0.31" ], [ "0.471", "0.2944", "0.509", "0.2944", "0.509", "0.31", "0.471", "0.31" ], [ "0.517", "0.2944", "0.555", "0.2944", "0.555", "0.31", "0.517", "0.31" ], [ "0.5615", "0.2944", "0.5996", "0.2944", "0.5996", "0.31", "0.5615", "0.31" ], [ "0.608", "0.2944", "0.6455", "0.2944", "0.6455", "0.31", "0.608", "0.31" ], [ "0.6978", "0.2944", "0.735", "0.2944", "0.735", "0.31", "0.6978", "0.31" ], [ "0.7446", "0.2944", "0.781", "0.2944", "0.781", "0.31", "0.7446", "0.31" ], [ "0.79", "0.2944", "0.825", "0.2944", "0.825", "0.31", "0.79", "0.31" ], [ "0.6543", "0.2964", "0.6885", "0.2964", "0.6885", "0.3088", "0.6543", "0.3088" ], [ "0.12427", "0.2964", "0.2418", "0.2986", "0.2418", "0.31", "0.12427", "0.308" ], [ "0.4165", "0.3088", "0.465", "0.3088", "0.465", "0.3237", "0.4165", "0.3237" ], [ "0.4685", "0.3088", "0.513", "0.3088", "0.513", "0.3237", "0.4685", "0.3237" ], [ "0.511", "0.3088", "0.5557", "0.3088", "0.5557", "0.3237", "0.511", "0.3237" ], [ "0.5576", "0.3088", "0.602", "0.3088", "0.602", "0.3237", "0.5576", "0.3237" ], [ "0.6484", "0.3088", "0.693", "0.3088", "0.693", "0.3247", "0.6484", "0.3247" ], [ "0.742", "0.3088", "0.781", "0.3088", "0.781", "0.3247", "0.742", "0.3247" ], [ "0.785", "0.3088", "0.827", "0.3088", "0.827", "0.3247", "0.785", "0.3247" ], [ "0.8345", "0.3088", "0.876", "0.3088", "0.876", "0.3247", "0.8345", "0.3247" ], [ "0.3804", "0.31", "0.4165", "0.31", "0.4165", "0.3225", "0.3804", "0.3225" ], [ "0.12177", "0.31", "0.1587", "0.31", "0.1587", "0.3225", "0.12177", "0.3225" ], [ "0.607", "0.31", "0.6455", "0.31", "0.6455", "0.3225", "0.607", "0.3225" ], [ "0.6978", "0.31", "0.739", "0.31", "0.739", "0.3237", "0.6978", "0.3237" ], [ "0.1234", "0.3237", "0.168", "0.3237", "0.168", "0.3384", "0.1234", "0.3384" ], [ "0.744", "0.3237", "0.781", "0.3237", "0.781", "0.339", "0.744", "0.339" ], [ "0.8364", "0.3218", "0.876", "0.3247", "0.873", "0.339", "0.8345", "0.3374" ], [ "0.4282", "0.3247", "0.461", "0.3247", "0.461", "0.3374", "0.4282", "0.3374" ], [ "0.382", "0.3247", "0.4165", "0.3247", "0.4165", "0.3374", "0.382", "0.3374" ], [ "0.4727", "0.3247", "0.5073", "0.3247", "0.5073", "0.3374", "0.4727", "0.3374" ], [ "0.517", "0.3247", "0.552", "0.3247", "0.552", "0.3374", "0.517", "0.3374" ], [ "0.5615", "0.3247", "0.596", "0.3247", "0.596", "0.3374", "0.5615", "0.3374" ], [ "0.6094", "0.3247", "0.644", "0.3247", "0.644", "0.3374", "0.6094", "0.3374" ], [ "0.656", "0.3247", "0.69", "0.3247", "0.69", "0.3374", "0.656", "0.3374" ], [ "0.7", "0.3247", "0.735", "0.3247", "0.735", "0.3374", "0.7", "0.3374" ], [ "0.7925", "0.3247", "0.827", "0.3247", "0.827", "0.3374", "0.7925", "0.3374" ], [ "0.11584", "0.3499", "0.881", "0.3499", "0.881", "0.3635", "0.11584", "0.3635" ], [ "0.11584", "0.3652", "0.881", "0.3652", "0.881", "0.379", "0.11584", "0.379" ], [ "0.11755", "0.3777", "0.869", "0.379", "0.869", "0.3926", "0.11755", "0.3914" ], [ "0.1184", "0.3938", "0.882", "0.3938", "0.882", "0.4067", "0.1184", "0.4067" ], [ "0.1184", "0.4092", "0.5015", "0.4092", "0.5015", "0.4192", "0.1184", "0.4192" ], [ "0.11755", "0.449", "0.4878", "0.449", "0.4878", "0.462", "0.11755", "0.462" ], [ "0.513", "0.4478", "0.882", "0.4478", "0.882", "0.462", "0.513", "0.462" ], [ "0.1184", "0.4644", "0.4878", "0.4644", "0.4878", "0.478", "0.1184", "0.478" ], [ "0.513", "0.4644", "0.771", "0.4644", "0.771", "0.478", "0.513", "0.478" ], [ "0.1184", "0.4797", "0.1948", "0.4797", "0.1948", "0.4946", "0.1184", "0.4946" ], [ "0.511", "0.4922", "0.6675", "0.4922", "0.6675", "0.506", "0.511", "0.506" ], [ "0.11584", "0.5083", "0.2688", "0.5083", "0.2688", "0.5225", "0.11584", "0.5225" ], [ "0.51", "0.5127", "0.8564", "0.514", "0.8564", "0.531", "0.51", "0.5293" ], [ "0.11755", "0.531", "0.486", "0.531", "0.486", "0.5444", "0.11755", "0.5444" ], [ "0.513", "0.535", "0.879", "0.535", "0.879", "0.549", "0.513", "0.549" ], [ "0.12006", "0.549", "0.2183", "0.549", "0.2183", "0.56", "0.12006", "0.56" ], [ "0.513", "0.5503", "0.881", "0.5503", "0.881", "0.564", "0.513", "0.564" ], [ "0.5146", "0.568", "0.881", "0.568", "0.881", "0.5806", "0.5146", "0.5806" ], [ "0.1234", "0.576", "0.4895", "0.576", "0.4895", "0.589", "0.1234", "0.589" ], [ "0.513", "0.583", "0.58", "0.583", "0.58", "0.597", "0.513", "0.597" ], [ "0.11584", "0.5913", "0.486", "0.5903", "0.486", "0.605", "0.11584", "0.6055" ], [ "0.531", "0.5996", "0.881", "0.5996", "0.881", "0.6143", "0.531", "0.6143" ], [ "0.12006", "0.608", "0.4878", "0.608", "0.4878", "0.6216", "0.12006", "0.6216" ], [ "0.511", "0.615", "0.879", "0.615", "0.879", "0.629", "0.511", "0.629" ], [ "0.11755", "0.6235", "0.4878", "0.6235", "0.4878", "0.637", "0.11755", "0.637" ], [ "0.511", "0.6313", "0.879", "0.6313", "0.879", "0.6445", "0.511", "0.6445" ], [ "0.1184", "0.64", "0.4895", "0.64", "0.4895", "0.654", "0.1184", "0.654" ], [ "0.511", "0.648", "0.879", "0.648", "0.879", "0.661", "0.511", "0.661" ], [ "0.11755", "0.6567", "0.4895", "0.6567", "0.4895", "0.6704", "0.11755", "0.6704" ], [ "0.511", "0.663", "0.879", "0.663", "0.879", "0.677", "0.511", "0.677" ], [ "0.11755", "0.6714", "0.4844", "0.673", "0.4844", "0.687", "0.11755", "0.686" ], [ "0.513", "0.6797", "0.881", "0.6787", "0.881", "0.6934", "0.513", "0.695" ], [ "0.11584", "0.687", "0.2024", "0.6895", "0.2024", "0.7036", "0.11584", "0.7017" ], [ "0.513", "0.697", "0.881", "0.697", "0.881", "0.7104", "0.513", "0.7104" ], [ "0.12177", "0.7163", "0.4844", "0.7163", "0.4844", "0.73", "0.12177", "0.73" ], [ "0.511", "0.712", "0.881", "0.712", "0.881", "0.7256", "0.511", "0.7256" ], [ "0.12006", "0.7324", "0.486", "0.7324", "0.486", "0.745", "0.12006", "0.745" ], [ "0.513", "0.7285", "0.882", "0.7285", "0.882", "0.742", "0.513", "0.742" ], [ "0.12006", "0.749", "0.486", "0.749", "0.486", "0.7627", "0.12006", "0.7627" ], [ "0.51", "0.7437", "0.876", "0.7437", "0.876", "0.7573", "0.51", "0.7573" ], [ "0.11755", "0.764", "0.4895", "0.763", "0.4895", "0.778", "0.11755", "0.7793" ], [ "0.513", "0.761", "0.881", "0.761", "0.881", "0.775", "0.513", "0.775" ], [ "0.11755", "0.781", "0.486", "0.781", "0.486", "0.7944", "0.11755", "0.7944" ], [ "0.513", "0.777", "0.644", "0.777", "0.644", "0.7905", "0.513", "0.7905" ], [ "0.11755", "0.796", "0.4844", "0.7964", "0.4844", "0.811", "0.11755", "0.81" ], [ "0.513", "0.804", "0.553", "0.804", "0.553", "0.8154", "0.513", "0.8154" ], [ "0.5693", "0.802", "0.8447", "0.804", "0.8447", "0.8184", "0.5693", "0.8164" ], [ "0.11755", "0.811", "0.482", "0.8125", "0.482", "0.8267", "0.11755", "0.826" ], [ "0.565", "0.8193", "0.833", "0.822", "0.833", "0.835", "0.565", "0.833" ], [ "0.1184", "0.829", "0.4114", "0.829", "0.4114", "0.843", "0.1184", "0.843" ], [ "0.513", "0.839", "0.881", "0.8403", "0.881", "0.855", "0.513", "0.854" ], [ "0.12177", "0.857", "0.4895", "0.857", "0.4895", "0.8706", "0.12177", "0.8706" ], [ "0.5146", "0.857", "0.881", "0.857", "0.881", "0.8706", "0.5146", "0.8706" ], [ "0.11584", "0.8716", "0.4878", "0.8726", "0.4878", "0.887", "0.11584", "0.8857" ], [ "0.513", "0.8726", "0.879", "0.8726", "0.879", "0.8857", "0.513", "0.8857" ], [ "0.1184", "0.888", "0.4878", "0.888", "0.4878", "0.903", "0.1184", "0.903" ], [ "0.513", "0.888", "0.881", "0.888", "0.881", "0.903", "0.513", "0.903" ], [ "0.11755", "0.905", "0.4844", "0.905", "0.4844", "0.9185", "0.11755", "0.9185" ], [ "0.513", "0.905", "0.742", "0.905", "0.742", "0.9185", "0.513", "0.9185" ] ]

[ "Model", "ja", "AVG", "de", "bn", "fr", "te", "th", "zh ", "ru", "es", "sw", "GPT-3 (text-davinci-002)", "14.8", "17.2", "11.2", "18.0", "11.3", "4.4", "Direct (Shietal.,2022)", "16.8", "12.4", "8.8", "0.8", "8.8", "11.2", "10.8", "40.0", "23.7", "6.4", "Native-CoT t (Shiet al., 2022)", "36.0", "40.4", "37.6", "26.0", "28.4", "0.4", "44.0", "44.8", "46.0", "32.4", "28.4", "19.6", "40.8", "29.2", "9.6", "En-CoT t (Shi et al., 2022)", "20.8", "5.6", "Translate-En t (Shi et al., 2022)", "41.2", "46.4", "51.6", "46.4", "44.8", "48.8", "37.6", "42.8", "41.2", "47.2", "44.8", "PaLM-540B", "19.6", "16.8", "18.3", "18.8", "Direct (Shi et al., 2022)", "20.0", "16.0", "22.0", "15.6", "17.6", "19.2", "46.0", "17.2", "40.0", "48.4", "35.2", "52.8", "46.8", "45.6", "49.2", "48.7", "56.8", "Native-CoT t (Shiet al., 2022)", "46.4", "55.6", "50.1", "53.6", "46.4", "49.6", "44.4", "46.8", "49.6", "58.0", "En-CoT t (Shi et al., 2022)", "51.2", "46.0", "57.2", "60.0", "55.2", "50.0", "59.6", "51.2", "49.6", "50.8", "55.6", "54.2", "53.2", "Translate-En t (Shi et al., 2022)", "GPT3.5 (gpt-3.5-turbo)", "33.6", "56.0", "62.0", "52.8", "62.0", "48.0", "7.6", "42.4", "61.2", "Direct", "60.0", "48.6", "26.4", "70.0", "64.4", "51.0", "52.8", "70.4", "62.4", "54.0", "10.4", "40.0", "59.6", "50.0", "Native-CoT", "73.6", "63.2", "57.8", "70.0", "69.6", "60.4", "65.6", "55.2", "22.0", "48.0", "68.4", "En-CoT", "66.4", "75.6", "74.4", "72.4", "66.0", "72.8", "58.0", "57.6", "71.6", "69.6", "Translate-En", "80.0", "82.4", "79.2", "69.2", "74.8", "62.0", "73.6", "70.6", "64.8", "CLP", "81.6", "38.8", "CLSP", "68.0", "76.7", "86.8", "75.2", "84.8", "82.0", "77.2", "87.6", "76.0", "52.0", "77.2", "Table 1: Main results on MGSM. \" Direct \" denotes the original input request will be given to model di-", "rectly. \" Nat ive-CoT *\" signifies that the model generates inference steps in the same language as the input.", "\" En-CoT \" indicates the given non-English input request and returned with English chain-of-thought result.", "Translate-En *\" denotes we translate non-English input requests into English by Google translation API. t", "denotesthe6-shotresultssourcedfromShietal.(2022).", "1。我们从[0,2]中选择跨语言对齐提示中的温", "知识可以有效地提高跨语言CoT的推理性能，", "度，从[0,1]中选择特定于任务的求解器提示", "验证了跨语言自洽提示的有效性", "中的温度", "5.3CLP Analysis", "5.2Main Results", "5.3.1 CLP results better reasoning quality", "主要结果如表1所示。从结果来看，我们有以", "为了进一步研究CLP的工作原理，我们使用", "下观察结果：", "RoscoEE(Golovnevaetal.,2022）的框架来评", "估模型思想链中推理路径的质量。实现细节见", "(1) GPT-3.5 exhibits notable cross-lingual rea-", "附录?？", "soningsuperiority在表1，GPT-3中提到的", "如图E4所示，我们发现CLP的推理路径表", "所有方案中进行评估时。5明显超过了PaLM-", "现出更高的忠实性，在推理过程中与关键步骤", "540B和GPT-3的少量拍摄结果（实现了30", "表现出更好的一致性。具体来说，忠诚得分增", "的改进。3%,2。3%7。7%和14。2%overPaLM-", "加了1。6%，表明该模型更好地理解了问题陈", "540B,respectively)。如Wangetal.（2023a）所示，", "述，并确保了清晰的推理链，而不会产生不相", "多语言SFT和RLHF技术可大大提高跨语言", "关或滥用的信息。此外，我们观察到2。8%和", "推理性能。", "2。5%分别对“步骤”和“链”的信息度量进", "(2) CLP achieves state-of-the-art performance.", "行了改进。这表明，在跨语言对齐之后，该模", "如表E1所示，CLP超越了以前的所有基线，特", "型的推理可以提供更有根据的推理步骤。此外，", "别是优于PALM-540B（Translate-En），提", "CLP 显示了一个 2。8%在失步度量中的增强", "高了16。4%。这种改进不能完全归因于GPT-", "表明模型的推理可以包含完整的逻辑链，从而", "3。5(中电甚至达到了2。2%的平均精度高于", "带来更好的性能", "Translate-En）。这些发现表明，跨语言对", "5.3.2", "Two-stage interactive prompting is", "齐提示（CLP）超越了简单的文本翻译，进一", "betterthan single turn prompting", "步增强了模型固有的跨语言理解能力", "本节探讨两阶段交互式提示的有效性。我们不", "(3) CLSP further significantly improves perfor-", "使用两圈跨语言对齐提示和特定于任务的求解", "mance.如表国1所示，CLSP在所有语言中都", "器提示来分别执行对齐和任务求解，而是使用", "比CLP具有显着优势（6。1%在平均精度上", "换行字符“n”直接连接跨语言对齐提示和特", "的改进)。这一观察结果表明，整合不同语言的", "定于任务的求解器提示LLM" ]

{ "arxiv_id": "2310.14799", "language": "zh", "page_idx": 4 }

[ [ "0.336", "0.0861", "0.6665", "0.0861", "0.6665", "0.10095", "0.336", "0.10095" ], [ "0.07806", "0.1247", "0.9355", "0.1247", "0.9355", "0.1395", "0.07806", "0.1395" ], [ "0.2957", "0.155", "0.7095", "0.1567", "0.7095", "0.1746", "0.2957", "0.1729" ], [ "0.2896", "0.1758", "0.727", "0.1758", "0.727", "0.1907", "0.2896", "0.1907" ], [ "0.251", "0.1895", "0.756", "0.1907", "0.756", "0.2079", "0.251", "0.2073" ], [ "0.367", "0.2073", "0.6484", "0.2079", "0.6484", "0.2257", "0.367", "0.2245" ], [ "0.2627", "0.2632", "0.3435", "0.2654", "0.3435", "0.2808", "0.2627", "0.279" ], [ "0.5557", "0.2725", "0.567", "0.2766", "0.5615", "0.282", "0.552", "0.279" ], [ "0.678", "0.2737", "0.727", "0.2737", "0.727", "0.2808", "0.678", "0.2808" ], [ "0.738", "0.2756", "0.8037", "0.2756", "0.8037", "0.2798", "0.738", "0.2798" ], [ "0.8096", "0.2725", "0.837", "0.275", "0.8364", "0.282", "0.8076", "0.2798" ], [ "0.592", "0.2832", "0.7793", "0.2832", "0.7793", "0.2935", "0.592", "0.2935" ], [ "0.1486", "0.2957", "0.4592", "0.2957", "0.4592", "0.3088", "0.1486", "0.3088" ], [ "0.529", "0.2998", "0.748", "0.2998", "0.748", "0.3071", "0.529", "0.3071" ], [ "0.837", "0.2976", "0.873", "0.2976", "0.873", "0.3088", "0.837", "0.3088" ], [ "0.7886", "0.3047", "0.827", "0.3047", "0.827", "0.3123", "0.7886", "0.3123" ], [ "0.8364", "0.3047", "0.873", "0.3071", "0.8716", "0.3184", "0.8345", "0.3164" ], [ "0.5283", "0.306", "0.7715", "0.3071", "0.7715", "0.3176", "0.5283", "0.3164" ], [ "0.147", "0.3088", "0.4575", "0.3088", "0.4575", "0.3225", "0.147", "0.3225" ], [ "0.5283", "0.3142", "0.775", "0.3154", "0.775", "0.326", "0.5283", "0.3247" ], [ "0.802", "0.3142", "0.812", "0.3142", "0.812", "0.3196", "0.802", "0.3196" ], [ "0.837", "0.3164", "0.8716", "0.3164", "0.8716", "0.3247", "0.837", "0.3247" ], [ "0.5264", "0.3225", "0.723", "0.3237", "0.723", "0.3342", "0.5264", "0.3333" ], [ "0.8403", "0.3247", "0.8696", "0.3247", "0.8696", "0.3333", "0.8403", "0.3333" ], [ "0.147", "0.3374", "0.4575", "0.3374", "0.4575", "0.3508", "0.147", "0.3508" ], [ "0.1486", "0.326", "0.4575", "0.326", "0.4575", "0.3362", "0.1486", "0.3362" ], [ "0.1436", "0.3508", "0.461", "0.3516", "0.461", "0.3665", "0.1436", "0.3652" ], [ "0.5283", "0.3486", "0.7197", "0.3486", "0.7197", "0.3557", "0.5283", "0.3557" ], [ "0.7197", "0.3516", "0.738", "0.3516", "0.738", "0.3552", "0.7197", "0.3552" ], [ "0.79", "0.354", "0.8076", "0.354", "0.8076", "0.3582", "0.79", "0.3582" ], [ "0.8135", "0.357", "0.8228", "0.354", "0.8237", "0.357", "0.817", "0.3599" ], [ "0.7886", "0.361", "0.8213", "0.361", "0.8213", "0.3682", "0.7886", "0.3682" ], [ "0.147", "0.3665", "0.4592", "0.3665", "0.4592", "0.3801", "0.147", "0.3801" ], [ "0.6978", "0.3552", "0.7515", "0.357", "0.7515", "0.3677", "0.6978", "0.3652" ], [ "0.531", "0.3665", "0.764", "0.3665", "0.764", "0.3735", "0.531", "0.3735" ], [ "0.534", "0.3772", "0.6533", "0.3772", "0.6533", "0.3813", "0.534", "0.3813" ], [ "0.7886", "0.3735", "0.797", "0.3735", "0.797", "0.3772", "0.7886", "0.3772" ], [ "0.147", "0.3813", "0.4592", "0.3813", "0.4592", "0.3943", "0.147", "0.3943" ], [ "0.147", "0.3955", "0.4592", "0.3955", "0.4592", "0.4092", "0.147", "0.4092" ], [ "0.147", "0.4092", "0.461", "0.4092", "0.461", "0.4229", "0.147", "0.4229" ], [ "0.5146", "0.4038", "0.879", "0.4038", "0.879", "0.4146", "0.5146", "0.4146" ], [ "0.3257", "0.4216", "0.461", "0.4233", "0.461", "0.437", "0.3257", "0.4353" ], [ "0.511", "0.4163", "0.881", "0.415", "0.881", "0.43", "0.511", "0.4312" ], [ "0.1486", "0.4246", "0.2032", "0.4246", "0.2032", "0.4358", "0.1486", "0.4358" ], [ "0.2007", "0.4275", "0.2083", "0.4275", "0.2083", "0.4316", "0.2007", "0.4316" ], [ "0.2065", "0.4246", "0.3267", "0.4246", "0.3267", "0.4353", "0.2065", "0.4353" ], [ "0.511", "0.4312", "0.881", "0.4312", "0.881", "0.4448", "0.511", "0.4448" ], [ "0.1436", "0.437", "0.4592", "0.4358", "0.4592", "0.4507", "0.1436", "0.452" ], [ "0.513", "0.4448", "0.784", "0.4448", "0.784", "0.459", "0.513", "0.459" ], [ "0.147", "0.452", "0.4592", "0.452", "0.4592", "0.4656", "0.147", "0.4656" ], [ "0.1486", "0.4673", "0.1948", "0.4673", "0.1948", "0.4785", "0.1486", "0.4785" ], [ "0.2284", "0.4644", "0.4592", "0.4656", "0.4592", "0.4797", "0.2284", "0.4785" ], [ "0.534", "0.4685", "0.882", "0.4685", "0.882", "0.4822", "0.534", "0.4822" ], [ "0.1453", "0.4785", "0.4592", "0.4797", "0.4592", "0.4946", "0.1453", "0.4934" ], [ "0.5156", "0.4827", "0.882", "0.484", "0.882", "0.4988", "0.5156", "0.4976" ], [ "0.147", "0.4946", "0.4592", "0.4946", "0.4592", "0.5083", "0.147", "0.5083" ], [ "0.5146", "0.4995", "0.882", "0.5005", "0.882", "0.5156", "0.5146", "0.514" ], [ "0.147", "0.5083", "0.4592", "0.5083", "0.4592", "0.5215", "0.147", "0.5215" ], [ "0.1453", "0.5225", "0.4592", "0.5215", "0.4592", "0.536", "0.1453", "0.5376" ], [ "0.5156", "0.517", "0.8833", "0.517", "0.8833", "0.531", "0.5156", "0.531" ], [ "0.147", "0.5376", "0.4575", "0.5376", "0.4575", "0.5503", "0.147", "0.5503" ], [ "0.5146", "0.533", "0.5366", "0.533", "0.5366", "0.5474", "0.5146", "0.5474" ], [ "0.147", "0.5503", "0.4592", "0.5503", "0.4592", "0.564", "0.147", "0.564" ], [ "0.534", "0.5527", "0.8833", "0.5527", "0.8833", "0.5664", "0.534", "0.5664" ], [ "0.1486", "0.5654", "0.4592", "0.5654", "0.4592", "0.579", "0.1486", "0.579" ], [ "0.5146", "0.568", "0.882", "0.5684", "0.882", "0.583", "0.5146", "0.582" ], [ "0.147", "0.579", "0.4568", "0.579", "0.4568", "0.593", "0.147", "0.593" ], [ "0.513", "0.5845", "0.882", "0.585", "0.882", "0.5996", "0.513", "0.5986" ], [ "0.147", "0.5933", "0.4592", "0.5933", "0.4592", "0.607", "0.147", "0.607" ], [ "0.517", "0.6016", "0.882", "0.6016", "0.882", "0.615", "0.517", "0.615" ], [ "0.147", "0.608", "0.378", "0.608", "0.378", "0.6216", "0.147", "0.6216" ], [ "0.3728", "0.6123", "0.3804", "0.6123", "0.3804", "0.616", "0.3728", "0.616" ], [ "0.3787", "0.6094", "0.4568", "0.6094", "0.4568", "0.619", "0.3787", "0.619" ], [ "0.5156", "0.616", "0.882", "0.6177", "0.882", "0.632", "0.5156", "0.6313" ], [ "0.147", "0.6226", "0.4592", "0.6226", "0.4592", "0.6357", "0.147", "0.6357" ], [ "0.147", "0.6353", "0.2703", "0.637", "0.2703", "0.651", "0.147", "0.6484" ], [ "0.2678", "0.6353", "0.4575", "0.6353", "0.4575", "0.6484", "0.2678", "0.6484" ], [ "0.5156", "0.6343", "0.882", "0.6343", "0.882", "0.648", "0.5156", "0.648" ], [ "0.2678", "0.64", "0.2737", "0.64", "0.2737", "0.6455", "0.2678", "0.6455" ], [ "0.1486", "0.651", "0.4592", "0.651", "0.4592", "0.6646", "0.1486", "0.6646" ], [ "0.5156", "0.651", "0.882", "0.651", "0.882", "0.6646", "0.5156", "0.6646" ], [ "0.1453", "0.6646", "0.4592", "0.665", "0.4592", "0.6787", "0.1453", "0.6777" ], [ "0.517", "0.6665", "0.882", "0.6665", "0.882", "0.6797", "0.517", "0.6797" ], [ "0.1486", "0.6787", "0.4592", "0.6787", "0.4592", "0.6924", "0.1486", "0.6924" ], [ "0.517", "0.6816", "0.882", "0.6816", "0.882", "0.6953", "0.517", "0.6953" ], [ "0.1486", "0.6924", "0.444", "0.6924", "0.444", "0.706", "0.1486", "0.706" ], [ "0.517", "0.6978", "0.882", "0.6978", "0.882", "0.712", "0.517", "0.712" ], [ "0.5146", "0.712", "0.8833", "0.7134", "0.8833", "0.728", "0.5146", "0.727" ], [ "0.1184", "0.7246", "0.1318", "0.7246", "0.1318", "0.7354", "0.1184", "0.7354" ], [ "0.1528", "0.7227", "0.266", "0.7236", "0.266", "0.7383", "0.1528", "0.737" ], [ "0.5146", "0.7285", "0.8833", "0.731", "0.8833", "0.7446", "0.5146", "0.742" ], [ "0.12006", "0.7505", "0.4878", "0.7505", "0.4878", "0.764", "0.12006", "0.764" ], [ "0.5156", "0.7466", "0.76", "0.7466", "0.76", "0.761", "0.5156", "0.761" ], [ "0.12177", "0.7666", "0.4878", "0.7666", "0.4878", "0.78", "0.12177", "0.78" ], [ "0.532", "0.7666", "0.882", "0.7666", "0.882", "0.78", "0.532", "0.78" ], [ "0.12177", "0.782", "0.4895", "0.782", "0.4895", "0.796", "0.12177", "0.796" ], [ "0.5156", "0.782", "0.882", "0.782", "0.882", "0.7964", "0.5156", "0.7964" ], [ "0.12006", "0.799", "0.4895", "0.799", "0.4895", "0.8125", "0.12006", "0.8125" ], [ "0.5156", "0.799", "0.8833", "0.799", "0.8833", "0.8125", "0.5156", "0.8125" ], [ "0.12006", "0.814", "0.4895", "0.814", "0.4895", "0.8276", "0.12006", "0.8276" ], [ "0.5146", "0.814", "0.882", "0.814", "0.882", "0.8276", "0.5146", "0.8276" ], [ "0.12006", "0.8306", "0.4895", "0.8306", "0.4895", "0.844", "0.12006", "0.844" ], [ "0.5146", "0.8306", "0.882", "0.8306", "0.882", "0.844", "0.5146", "0.844" ], [ "0.12006", "0.8467", "0.4895", "0.8467", "0.4895", "0.861", "0.12006", "0.861" ], [ "0.5156", "0.847", "0.8833", "0.847", "0.8833", "0.861", "0.5156", "0.861" ], [ "0.1184", "0.8633", "0.486", "0.8633", "0.486", "0.877", "0.1184", "0.877" ], [ "0.5146", "0.8633", "0.882", "0.8633", "0.882", "0.877", "0.5146", "0.877" ], [ "0.12006", "0.879", "0.4878", "0.879", "0.4878", "0.8926", "0.12006", "0.8926" ], [ "0.5156", "0.879", "0.882", "0.879", "0.882", "0.8926", "0.5156", "0.8926" ], [ "0.12006", "0.8955", "0.3074", "0.8955", "0.3074", "0.909", "0.12006", "0.909" ], [ "0.5156", "0.8955", "0.827", "0.8955", "0.827", "0.909", "0.5156", "0.909" ] ]

[ "为多跳问答任务重新添加附属品", "Harsh Trivedi*, Heeyoung Kwon*, Tushar Khot*, Ashish Sabharwal, Niranjan Balasubramanian*", "+ Stony Brook University, Stony Brook, U.S.A.", "[ hjtrivedi,heekwon,niranjan @cs.stonybrook.edu", " Allen Institute for Artificial Intelligence, Seattle, U.S.A.", "[ tushark,ashishs @allenai.org", "Abstract", "Nn", "ed? (A) Ca", "rdaetBlsiic", "Valler", "H:Facebook was launched in Cambridge", "Question Answering (QA) naturally re-", "P1:FacebookwaslaunchedatHarvardUniversit", "P1:0.4", "Relevance", "P2:0.1", "P2:Facebook headquarters was set up in Silicon Valley", "duces to an entailment problem, namely.", "P3:Harvard University is at Cambridge,Massachusetts.", "od", "P3:0.4", "P4:Harvard isonlya fewmilesfrom Boston", "P4:0.1", "swer to a question. However, for multi-", "verifying whether sometext entailsthean-", "hop QA tasks, which require reasoning", "Facehookwas launchedatHarvardUnivers", "eit", "te", "", "Aggregati", "with multiple sentences, it remains un-", "loy.Harvard", "UniversityisatCambridge,Massachusetts.Harvardi", "nvalew.milesromBoso", "tara", "clear how best to utilize entailment mod-", "els pre-trained on large scale datasets such", "as SNLI, which are based on sentence", "Figure l:Anexample illustrating the challenges", "Multee,a", "in using sentence-level entailment model for multi", "pairs.", "", "Weintroduce", " sentence reasoning needed for QA, and the high-", "general architecture that can effectively", "level approach used in Multee.", "use entailment models for multi-hop QA", "tasks.", "Multee uses (i)a local", "将蕴涵模型用于QA的一个主要困难是两", "module that helps locate important sen-", "个任务的输入之间的不匹配：大规模蕴涵数据", "tences, thereby avoiding distracting infor-", "集通常是在句子级别构建的，而问题回答则需", "mation, and (i) a global module that", "aggregates information by effectively in-", "要验证多个句子是否在一起作为前提，需要假", "corporating importance weights.Impor-", "设", "tantly, we show that both modules can", "有两种直接的方法可以解决这种不匹配：（1）", "use entailment functions pre-trained on a", "在每个前提句子上聚合独立的蕴涵决策，或者", "large scale NLI datasets. We evaluate per", "（2）在连接所有前提句子后做出单个蕴涵决策。", "formance on MultiRC and OpenBookQA,", "两种方法都不完全令人满意。要理解原因，请", "two multihop QA datasets.", "", "When us-", "考虑图E1中的前提集，其中包含假设Hc。具", "ing an entailment function pre-trained on", "NLIdatasets,", "Multee outperforms", "体来说，P1和P3中的组合信息需要Hc，它", "8", "QA models trained only on the target", "对应于正确答案Cambridge。一方面，聚合独", "QA datasets and the OpenAI transformer", "立决策将失败，因为没有单独的前提需要Hc", "models.The code is available at https:", "。另一方面，简单地将前提连接成一个段落将", "//github.com/StonyBrookNLP/multee.", "失败，因为P2和P4中的分散信息会混淆P1", "和P3中的有用信息。因此，一种有效的方法", "1", "Introduction", "必须认识到相关的句子（i。e。，避免分散注意", "我们如何有效地使用文本蕴涵模型来回答", "力），并撰写他们的句子级信息", "问题？以前在这方面的尝试取得了有限的", "我们对这一挑战的解决方案是基于这样的观", "成功 (Harabagiu and Hickl, 2006; Sacaleanu", "察，即句子级蕴涵函数可以重新用于识别相关", "etal.,2008;Clark etal.,2012)。随着最近", "句子和计算句子级表示。这两个任务都需要比", "的大规模蕴涵数据集E(Bowmanetal.，2015;", "较一对文本中的信息，但是比较的目标是不同", "Williams et al.,2018;Khot et al.,2018)将蕴", "的。这意味着我们可以采用为基本蕴涵训练的", "涵模型推向高精度E(Chenetal.，2017;Parikh", "蕴涵函数(i。e。，比较文本中的信息)，并使其", "etal.,2016;Wangetal.,2017），我们重新审", "适用于识别相关性和计算表示。因此，这种架", "视了这一挑战，并提出了一种将神经蕴涵模型", "构使我们能够整合蕴涵架构的进步，并利用使", "重新用于QA的新方法", "用大规模蕴涵数据集获得的预训练模型" ]

{ "arxiv_id": "1904.09380", "language": "zh", "page_idx": 0 }

[ [ "0.1394", "0.0766", "0.4902", "0.0766", "0.4902", "0.0903", "0.1394", "0.0903" ], [ "0.486", "0.08075", "0.492", "0.08075", "0.492", "0.0861", "0.486", "0.0861" ], [ "0.4878", "0.0766", "0.7305", "0.0766", "0.7305", "0.0903", "0.4878", "0.0903" ], [ "0.12177", "0.0932", "0.4878", "0.0932", "0.4878", "0.1069", "0.12177", "0.1069" ], [ "0.12006", "0.1081", "0.486", "0.10925", "0.486", "0.12354", "0.12006", "0.1223" ], [ "0.561", "0.1057", "0.8833", "0.1069", "0.8833", "0.1247", "0.561", "0.12354" ], [ "0.12177", "0.1259", "0.4878", "0.1259", "0.4878", "0.1395", "0.12177", "0.1395" ], [ "0.1184", "0.1401", "0.486", "0.1401", "0.486", "0.1538", "0.1184", "0.1538" ], [ "0.513", "0.1372", "0.881", "0.1372", "0.881", "0.1509", "0.513", "0.1509" ], [ "0.12177", "0.1567", "0.4878", "0.1567", "0.4878", "0.1704", "0.12177", "0.1704" ], [ "0.5146", "0.1526", "0.882", "0.1526", "0.882", "0.1663", "0.5146", "0.1663" ], [ "0.1184", "0.1729", "0.4895", "0.1716", "0.4895", "0.1859", "0.1184", "0.187" ], [ "0.513", "0.1687", "0.7188", "0.1692", "0.7188", "0.1841", "0.513", "0.1829" ], [ "0.12006", "0.1895", "0.4878", "0.1895", "0.4878", "0.2031", "0.12006", "0.2031" ], [ "0.534", "0.1859", "0.8833", "0.1859", "0.8833", "0.1996", "0.534", "0.1996" ], [ "0.11755", "0.2037", "0.2124", "0.206", "0.2124", "0.2203", "0.11755", "0.2185" ], [ "0.5146", "0.2007", "0.8833", "0.2031", "0.8833", "0.2162", "0.5146", "0.2144" ], [ "0.1377", "0.2227", "0.4895", "0.2227", "0.4895", "0.2363", "0.1377", "0.2363" ], [ "0.5156", "0.2185", "0.881", "0.2185", "0.881", "0.2322", "0.5156", "0.2322" ], [ "0.12006", "0.2382", "0.4895", "0.2382", "0.4895", "0.253", "0.12006", "0.253" ], [ "0.5146", "0.234", "0.5786", "0.234", "0.5786", "0.2488", "0.5146", "0.2488" ], [ "0.12006", "0.2546", "0.4878", "0.2546", "0.4878", "0.2695", "0.12006", "0.2695" ], [ "0.596", "0.257", "0.665", "0.257", "0.665", "0.2715", "0.596", "0.2715" ], [ "0.777", "0.257", "0.8447", "0.257", "0.8447", "0.2715", "0.777", "0.2715" ], [ "0.12177", "0.2715", "0.4878", "0.2715", "0.4878", "0.2852", "0.12177", "0.2852" ], [ "0.6265", "0.2737", "0.638", "0.2737", "0.638", "0.284", "0.6265", "0.284" ], [ "0.12006", "0.2874", "0.4895", "0.2874", "0.4895", "0.3005", "0.12006", "0.3005" ], [ "0.6245", "0.288", "0.6406", "0.288", "0.6406", "0.2964", "0.6245", "0.2964" ], [ "0.1184", "0.3027", "0.4878", "0.3018", "0.4878", "0.3164", "0.1184", "0.3176" ], [ "0.7944", "0.304", "0.8237", "0.304", "0.8237", "0.3237", "0.7944", "0.3237" ], [ "0.12006", "0.3196", "0.4878", "0.3196", "0.4878", "0.3333", "0.12006", "0.3333" ], [ "0.552", "0.3218", "0.581", "0.3218", "0.581", "0.3403", "0.552", "0.3403" ], [ "0.615", "0.3225", "0.6533", "0.3225", "0.6533", "0.3384", "0.615", "0.3384" ], [ "0.6724", "0.3218", "0.706", "0.3218", "0.706", "0.339", "0.6724", "0.339" ], [ "0.1184", "0.3342", "0.486", "0.3342", "0.486", "0.3474", "0.1184", "0.3474" ], [ "0.1184", "0.3508", "0.1957", "0.3508", "0.1957", "0.3652", "0.1184", "0.3652" ], [ "0.7886", "0.3508", "0.8286", "0.3499", "0.8286", "0.3582", "0.7886", "0.3594" ], [ "0.5474", "0.361", "0.575", "0.3594", "0.5767", "0.3677", "0.5483", "0.3694" ], [ "0.6206", "0.3623", "0.6455", "0.3623", "0.6455", "0.3694", "0.6206", "0.3694" ], [ "0.678", "0.3635", "0.693", "0.3635", "0.693", "0.3665", "0.678", "0.3665" ], [ "0.544", "0.3718", "0.5884", "0.3718", "0.5884", "0.386", "0.544", "0.386" ], [ "0.6094", "0.3723", "0.6753", "0.3723", "0.6753", "0.3843", "0.6094", "0.3843" ], [ "0.1377", "0.3665", "0.4878", "0.3665", "0.4878", "0.3813", "0.1377", "0.3813" ], [ "0.669", "0.3735", "0.7188", "0.3735", "0.7188", "0.3855", "0.669", "0.3855" ], [ "0.7676", "0.3735", "0.867", "0.3735", "0.867", "0.3855", "0.7676", "0.3855" ], [ "0.1184", "0.3855", "0.4878", "0.3855", "0.4878", "0.3984", "0.1184", "0.3984" ], [ "0.12177", "0.401", "0.486", "0.401", "0.486", "0.4146", "0.12177", "0.4146" ], [ "0.5146", "0.3997", "0.881", "0.3997", "0.881", "0.4133", "0.5146", "0.4133" ], [ "0.1184", "0.4163", "0.486", "0.4163", "0.486", "0.43", "0.1184", "0.43" ], [ "0.513", "0.4146", "0.876", "0.4146", "0.876", "0.4275", "0.513", "0.4275" ], [ "0.12177", "0.4329", "0.4878", "0.4329", "0.4878", "0.4465", "0.12177", "0.4465" ], [ "0.513", "0.4329", "0.882", "0.434", "0.882", "0.449", "0.513", "0.4478" ], [ "0.12006", "0.449", "0.486", "0.449", "0.486", "0.462", "0.12006", "0.462" ], [ "0.5146", "0.4495", "0.882", "0.4495", "0.882", "0.4644", "0.5146", "0.4644" ], [ "0.12006", "0.4656", "0.4878", "0.4656", "0.4878", "0.4785", "0.12006", "0.4785" ], [ "0.5146", "0.466", "0.8623", "0.466", "0.8623", "0.4797", "0.5146", "0.4797" ], [ "0.1184", "0.481", "0.486", "0.481", "0.486", "0.4946", "0.1184", "0.4946" ], [ "0.12177", "0.4963", "0.486", "0.4963", "0.486", "0.511", "0.12177", "0.511" ], [ "0.571", "0.4954", "0.882", "0.4954", "0.882", "0.5103", "0.571", "0.5103" ], [ "0.1184", "0.513", "0.486", "0.513", "0.486", "0.527", "0.1184", "0.527" ], [ "0.12006", "0.5293", "0.444", "0.5293", "0.444", "0.5435", "0.12006", "0.5435" ], [ "0.5156", "0.527", "0.882", "0.528", "0.882", "0.542", "0.5156", "0.5415" ], [ "0.1368", "0.5464", "0.4878", "0.546", "0.4878", "0.56", "0.1368", "0.561" ], [ "0.517", "0.5444", "0.882", "0.5444", "0.882", "0.558", "0.517", "0.558" ], [ "0.12006", "0.5635", "0.4878", "0.5635", "0.4878", "0.5767", "0.12006", "0.5767" ], [ "0.5146", "0.5586", "0.882", "0.56", "0.882", "0.5747", "0.5146", "0.5737" ], [ "0.12006", "0.579", "0.3728", "0.579", "0.3728", "0.593", "0.12006", "0.593" ], [ "0.517", "0.5767", "0.804", "0.5767", "0.804", "0.5913", "0.517", "0.5913" ], [ "0.1377", "0.5957", "0.4878", "0.5957", "0.4878", "0.6094", "0.1377", "0.6094" ], [ "0.5903", "0.6055", "0.8096", "0.605", "0.8096", "0.619", "0.5903", "0.6206" ], [ "0.8564", "0.6055", "0.882", "0.6055", "0.882", "0.619", "0.8564", "0.619" ], [ "0.1184", "0.627", "0.4878", "0.6274", "0.4878", "0.6426", "0.1184", "0.641" ], [ "0.12177", "0.611", "0.4878", "0.611", "0.4878", "0.626", "0.12177", "0.626" ], [ "0.532", "0.6343", "0.8833", "0.6343", "0.8833", "0.648", "0.532", "0.648" ], [ "0.12006", "0.6445", "0.486", "0.6445", "0.486", "0.6577", "0.12006", "0.6577" ], [ "0.5146", "0.6484", "0.882", "0.6494", "0.882", "0.6646", "0.5146", "0.663" ], [ "0.1234", "0.662", "0.486", "0.662", "0.486", "0.673", "0.1234", "0.673" ], [ "0.5146", "0.6665", "0.881", "0.6665", "0.881", "0.6797", "0.5146", "0.6797" ], [ "0.12006", "0.676", "0.486", "0.676", "0.486", "0.6895", "0.12006", "0.6895" ], [ "0.5146", "0.6816", "0.881", "0.6816", "0.881", "0.6953", "0.5146", "0.6953" ], [ "0.12006", "0.6924", "0.4895", "0.6914", "0.4895", "0.706", "0.12006", "0.707" ], [ "0.5146", "0.6987", "0.882", "0.6987", "0.882", "0.712", "0.5146", "0.712" ], [ "0.12006", "0.709", "0.1411", "0.709", "0.1411", "0.7227", "0.12006", "0.7227" ], [ "0.5146", "0.7144", "0.881", "0.7144", "0.881", "0.728", "0.5146", "0.728" ], [ "0.5146", "0.731", "0.8833", "0.731", "0.8833", "0.7446", "0.5146", "0.7446" ], [ "0.12006", "0.741", "0.1453", "0.741", "0.1453", "0.753", "0.12006", "0.753" ], [ "0.1561", "0.7393", "0.4802", "0.7393", "0.4802", "0.753", "0.1561", "0.753" ], [ "0.5146", "0.745", "0.882", "0.7476", "0.882", "0.761", "0.5146", "0.759" ], [ "0.12177", "0.767", "0.4895", "0.767", "0.4895", "0.781", "0.12177", "0.781" ], [ "0.5156", "0.763", "0.882", "0.763", "0.882", "0.777", "0.5156", "0.777" ], [ "0.12177", "0.7837", "0.4895", "0.7837", "0.4895", "0.7974", "0.12177", "0.7974" ], [ "0.5156", "0.7793", "0.882", "0.7793", "0.882", "0.792", "0.5156", "0.792" ], [ "0.12177", "0.8", "0.4878", "0.8", "0.4878", "0.8135", "0.12177", "0.8135" ], [ "0.5146", "0.7944", "0.882", "0.7944", "0.882", "0.808", "0.5146", "0.808" ], [ "0.1301", "0.8154", "0.4878", "0.8154", "0.4878", "0.829", "0.1301", "0.829" ], [ "0.5146", "0.811", "0.882", "0.811", "0.882", "0.8247", "0.5146", "0.8247" ], [ "0.12177", "0.832", "0.4878", "0.832", "0.4878", "0.8457", "0.12177", "0.8457" ], [ "0.5146", "0.8276", "0.553", "0.8276", "0.553", "0.839", "0.5146", "0.839" ], [ "0.1394", "0.858", "0.4902", "0.8594", "0.4902", "0.8735", "0.1394", "0.8726" ], [ "0.5156", "0.858", "0.529", "0.858", "0.529", "0.866", "0.5156", "0.866" ], [ "0.5474", "0.8525", "0.806", "0.8516", "0.806", "0.869", "0.5474", "0.8706" ], [ "0.529", "0.8594", "0.546", "0.858", "0.546", "0.8623", "0.529", "0.8633" ], [ "0.1184", "0.8706", "0.4844", "0.8716", "0.4844", "0.8857", "0.1184", "0.8853" ], [ "0.12006", "0.8843", "0.4895", "0.8843", "0.4895", "0.898", "0.12006", "0.898" ], [ "0.5156", "0.879", "0.8833", "0.879", "0.8833", "0.8936", "0.5156", "0.8936" ], [ "0.12006", "0.898", "0.2317", "0.898", "0.2317", "0.9077", "0.12006", "0.9077" ], [ "0.517", "0.8936", "0.885", "0.8945", "0.885", "0.909", "0.517", "0.9077" ] ]

[ "为此，我们提出了一种通用体系结构，该体", "", "造的假设语句Hga的概率：", "系结构将（预训练的）蕴涵函数fe用于多句子", "QA。给定一个表示候选答案的假设语句Hqa", "Pr[A | Q,{P,}] ≥Pr[{P} Hqa] (1)", "和一组前提句子P}，我们提出的体系结构", "将相同的函数fe用于两个组件：（a）句子相", "在这里，我们使用表示文本含义。给定QA", "关性模块，该模块根据每个P，的潜在相关性", "训练数据，然后我们可以学习一个近似蕴涵概", "对其进行评分Hga；目的是清除干扰物；（b)", "率Pr[(P}Hqa】的模型", "相关性加权聚合器，它结合了来自XMATHX", "能否使用已在大规模蕴涵数据集上预先训练", "的多个信息", "的现有蕴涵模型fe构建有效的QA模型9e？", "因此，我们构建了更大上下文的有效包含感", "图E2说明了使用fe作为黑盒函数的两种简", "知表示(i。e。，多个句子)来自小上下文(i。", "单方法：", "e。，个别句子)。我们方法的主要优势在于，与", "Pr(P) = Hga)", "Pr(P)E Ha)", "标准注意力机制不同，聚合器模块在多个抽象", "十", "级别上使用来自相关性模块的注意力分数（e。", "max", "g。，神经网络的多个层)在fe内，使用在每个", "fe", "级别上组成表示的连接操作。我们将这一实际", "fe", "fe", "fe", "信息表示的水平聚合称为参Multee（发音", "为multi)", "concat", "", "", "", "(P,Hga)", "(PaH)", "我们的参Multee实现使用ESIM（Chen", "(Pa,Ha)", "(Pi.Pa...Pa)Haa", "etal.，2017），这是一个最近的句子级蕴涵模", "型，在SNLI和MultiNLI数据集上进行了", "Figure 2: Black Box Applications of Textual En-", "预训练。我们证明了它在两个具有挑战性的", "tailment Model for QA: Max and Concat models", "多句子推理数据集上的有效性：MultiRCE", "(i)Aggregate Local Decisions(Max)：使用 fe", "(Khashabi et al., 2018)和 OpenBookQAE", "检查每个句子P,本身需要多少Hga，并聚合", "(Mihaylovetal.,2018）。使用ELMo上下文", "这些局部蕴涵决策，例如，使用max操作。", "嵌人的Multee（Petersetal.,2018）与", "大型基于转染子的模型（Radfordetal.，2018)", "ge({Pi), Hqa) = max fe(Pi, Hqa)(2)", "获得的最新结果相匹配，这些模型是在一系列", "大规模任务（Sunetal.,2019）上训练的", "(i）ConcatenatePremises（Concat)：将前提句", "消融研究表明，相关性评分和多级聚合都是", "子按顺序组合在一起，以形成单个大段落P", "有价值的，并且在大型蕴涵语料库上进行预训", "并使用fe检查该段落作为一个整体是否包含", "练对OpenBookQA特别有帮助", "假设Hga，从而做出单个包含决定：", "这项工作有三个主要贡献：（i）一种使用预", "ge([Pi), Hqa) = fe(P, Hqa)", "(3)", "将本地（句子级别）蕴涵决策与全局（文档级", "先训练的蕴涵模型进行问答的新颖方法。(ii)", "然而，我们的实验表明，这两种方法都不", "别）蕴涵决策相结合以有效聚合多跳QA任务", "是使用预先训练的包含模型进行QA的有效", "信息的模型。（ii）经验评估表明，基于蕴涵的", "方法（请参见表E1）。FortheexampleinFig-", "QA可以在两个具有挑战性的多跳QA数据集", "ure 1 ,Maxmodelwouldnotbeabletoconsiderin-", "OpenBookQA和MultiRC上实现最先进的性", "formationfromPlandP3together。相反，它将", "能", "把硅谷作为答案，因为P2接近H，“Face-", "book是在硅谷推出的”。同样，Concat也会被", "2", "Question Answering using Entailment", "P2中的分散信息所迷惑，这将削弱其对答案", "非提取性问答可以看作是一个文本蕴涵问题，", "剑桥的信心。因此，如果没有仔细的指导，简", "我们验证由一个问题和一个候选答案构建的假", "单的聚合很容易将分散注意力的信息添加到前", "设是否被知识所蕴涵一源文本中的句子集合", "提表示中，从而导致蕴涵失败。这促使人们需", "。在给定问题Q的情况下，答案A的概率", "要新的、有效的机制来对一系列前提进行全局", "可以建模为一组前提{P}包含由Q和A构", "推理", "IThis collection can be a sequence in the case of", "3", "Our Approach: Multee", "", "passage comprehension or a list of sentences, poten-", "tially from varied sources, in the case of QA over mul-", "我们提出了一个新的基于蕴涵的QA模型", "tiple documents.", "参Multee，该模型具有两个组成部分：（i)" ]

{ "arxiv_id": "1904.09380", "language": "zh", "page_idx": 1 }

[ [ "0.434", "0.07245", "0.482", "0.07245", "0.482", "0.08435", "0.434", "0.08435" ], [ "0.7573", "0.0968", "0.804", "0.0968", "0.804", "0.111", "0.7573", "0.111" ], [ "0.852", "0.0944", "0.888", "0.0944", "0.888", "0.114", "0.852", "0.114" ], [ "0.4375", "0.10394", "0.486", "0.10394", "0.486", "0.114", "0.4375", "0.114" ], [ "0.678", "0.10156", "0.6826", "0.10156", "0.6826", "0.1069", "0.678", "0.1069" ], [ "0.678", "0.1081", "0.6826", "0.1081", "0.6826", "0.1152", "0.678", "0.1152" ], [ "0.678", "0.11755", "0.6826", "0.11755", "0.6826", "0.1223", "0.678", "0.1223" ], [ "0.754", "0.11816", "0.806", "0.12054", "0.804", "0.136", "0.7524", "0.1342" ], [ "0.1276", "0.1289", "0.1663", "0.1289", "0.1663", "0.1414", "0.1276", "0.1414" ], [ "0.44", "0.1223", "0.4592", "0.1223", "0.4592", "0.13", "0.44", "0.13" ], [ "0.635", "0.1395", "0.681", "0.1395", "0.681", "0.1497", "0.635", "0.1497" ], [ "0.7583", "0.1383", "0.791", "0.1383", "0.791", "0.1509", "0.7583", "0.1509" ], [ "0.1276", "0.1443", "0.1663", "0.1443", "0.1663", "0.1567", "0.1276", "0.1567" ], [ "0.455", "0.1478", "0.4727", "0.1509", "0.4668", "0.1663", "0.4492", "0.1633" ], [ "0.2141", "0.155", "0.2351", "0.155", "0.2351", "0.1692", "0.2141", "0.1692" ], [ "0.523", "0.1509", "0.5425", "0.1509", "0.5425", "0.165", "0.523", "0.165" ], [ "0.6987", "0.155", "0.742", "0.158", "0.7407", "0.1704", "0.6978", "0.1675" ], [ "0.7573", "0.1561", "0.804", "0.1561", "0.804", "0.1704", "0.7573", "0.1704" ], [ "0.818", "0.1561", "0.867", "0.1561", "0.867", "0.1704", "0.818", "0.1704" ], [ "0.1394", "0.1663", "0.1528", "0.1663", "0.1528", "0.1799", "0.1394", "0.1799" ], [ "0.3862", "0.1567", "0.3972", "0.1567", "0.3972", "0.1621", "0.3862", "0.1621" ], [ "0.6987", "0.1787", "0.7407", "0.1812", "0.739", "0.1924", "0.6978", "0.1907" ], [ "0.76", "0.1812", "0.8003", "0.1812", "0.8003", "0.1924", "0.76", "0.1924" ], [ "0.8228", "0.1812", "0.864", "0.1812", "0.864", "0.1924", "0.8228", "0.1924" ], [ "0.12427", "0.187", "0.168", "0.187", "0.168", "0.2031", "0.12427", "0.2031" ], [ "0.2107", "0.1912", "0.287", "0.1912", "0.287", "0.2019", "0.2107", "0.2019" ], [ "0.534", "0.187", "0.575", "0.1895", "0.574", "0.2007", "0.532", "0.199" ], [ "0.4954", "0.1965", "0.5225", "0.1965", "0.5225", "0.1996", "0.4954", "0.1996" ], [ "0.771", "0.1965", "0.7925", "0.199", "0.791", "0.206", "0.769", "0.2037" ], [ "0.6978", "0.2114", "0.7407", "0.2114", "0.7407", "0.2227", "0.6978", "0.2227" ], [ "0.76", "0.212", "0.8003", "0.212", "0.8003", "0.2227", "0.76", "0.2227" ], [ "0.817", "0.212", "0.8657", "0.212", "0.8657", "0.2227", "0.817", "0.2227" ], [ "0.513", "0.2131", "0.55", "0.2131", "0.55", "0.2245", "0.513", "0.2245" ], [ "0.372", "0.2131", "0.4568", "0.2131", "0.4568", "0.2245", "0.372", "0.2245" ], [ "0.4802", "0.2197", "0.4902", "0.2197", "0.4902", "0.2227", "0.4802", "0.2227" ], [ "0.12006", "0.2446", "0.2258", "0.2446", "0.2258", "0.2559", "0.12006", "0.2559" ], [ "0.2301", "0.2477", "0.2377", "0.2477", "0.2377", "0.253", "0.2301", "0.253" ], [ "0.2393", "0.2434", "0.885", "0.2434", "0.885", "0.257", "0.2393", "0.257" ], [ "0.1184", "0.2583", "0.8833", "0.2583", "0.8833", "0.2725", "0.1184", "0.2725" ], [ "0.1184", "0.2725", "0.8833", "0.2725", "0.8833", "0.2861", "0.1184", "0.2861" ], [ "0.1184", "0.2852", "0.8833", "0.2874", "0.8833", "0.3018", "0.1184", "0.2998" ], [ "0.1184", "0.3005", "0.8833", "0.3005", "0.8833", "0.3154", "0.1184", "0.3154" ], [ "0.1184", "0.3154", "0.3687", "0.3154", "0.3687", "0.3289", "0.1184", "0.3289" ], [ "0.12006", "0.3425", "0.4978", "0.3425", "0.4978", "0.3557", "0.12006", "0.3557" ], [ "0.511", "0.3433", "0.6885", "0.3425", "0.6885", "0.357", "0.511", "0.3582" ], [ "0.12006", "0.3594", "0.2024", "0.3594", "0.2024", "0.3735", "0.12006", "0.3735" ], [ "0.58", "0.3801", "0.8037", "0.3818", "0.8037", "0.3984", "0.58", "0.3967" ], [ "0.1763", "0.4026", "0.3525", "0.4062", "0.3525", "0.4233", "0.1763", "0.4204" ], [ "0.5615", "0.401", "0.831", "0.402", "0.831", "0.4204", "0.5615", "0.4192" ], [ "0.1755", "0.4258", "0.4324", "0.4258", "0.4324", "0.4407", "0.1755", "0.4407" ], [ "0.6406", "0.4233", "0.6826", "0.4233", "0.6826", "0.4448", "0.6406", "0.4448" ], [ "0.565", "0.434", "0.615", "0.434", "0.615", "0.4536", "0.565", "0.4536" ], [ "0.1797", "0.4453", "0.3123", "0.4478", "0.3123", "0.462", "0.1797", "0.4602" ], [ "0.6265", "0.4453", "0.6978", "0.4453", "0.6978", "0.4656", "0.6265", "0.4656" ], [ "0.5767", "0.4697", "0.784", "0.4697", "0.784", "0.484", "0.5767", "0.484" ], [ "0.5825", "0.4863", "0.739", "0.4893", "0.739", "0.5073", "0.5825", "0.5044" ], [ "0.1377", "0.5005", "0.4844", "0.5005", "0.4844", "0.514", "0.1377", "0.514" ], [ "0.12177", "0.517", "0.444", "0.517", "0.444", "0.531", "0.12177", "0.531" ], [ "0.5146", "0.532", "0.81", "0.532", "0.81", "0.55", "0.5146", "0.55" ], [ "0.12006", "0.539", "0.486", "0.539", "0.486", "0.5527", "0.12006", "0.5527" ], [ "0.1318", "0.5547", "0.4895", "0.5547", "0.4895", "0.5693", "0.1318", "0.5693" ], [ "0.5356", "0.552", "0.8833", "0.554", "0.8833", "0.572", "0.5356", "0.5693" ], [ "0.12177", "0.572", "0.4878", "0.572", "0.4878", "0.585", "0.12177", "0.585" ], [ "0.517", "0.572", "0.882", "0.572", "0.882", "0.586", "0.517", "0.586" ], [ "0.12006", "0.586", "0.4878", "0.586", "0.4878", "0.5996", "0.12006", "0.5996" ], [ "0.511", "0.586", "0.882", "0.5874", "0.882", "0.6016", "0.511", "0.601" ], [ "0.11584", "0.5996", "0.2241", "0.6016", "0.2241", "0.6206", "0.11584", "0.618" ], [ "0.5156", "0.604", "0.882", "0.604", "0.882", "0.6177", "0.5156", "0.6177" ], [ "0.5156", "0.619", "0.882", "0.619", "0.882", "0.633", "0.5156", "0.633" ], [ "0.1377", "0.626", "0.4878", "0.626", "0.4878", "0.6396", "0.1377", "0.6396" ], [ "0.5146", "0.6343", "0.882", "0.6353", "0.882", "0.6494", "0.5146", "0.6484" ], [ "0.12177", "0.641", "0.4878", "0.641", "0.4878", "0.655", "0.12177", "0.655" ], [ "0.517", "0.652", "0.882", "0.652", "0.882", "0.665", "0.517", "0.665" ], [ "0.1234", "0.6577", "0.4895", "0.6577", "0.4895", "0.6714", "0.1234", "0.6714" ], [ "0.5156", "0.6685", "0.882", "0.6685", "0.882", "0.6816", "0.5156", "0.6816" ], [ "0.12006", "0.675", "0.3955", "0.675", "0.3955", "0.6885", "0.12006", "0.6885" ], [ "0.517", "0.684", "0.881", "0.684", "0.881", "0.6987", "0.517", "0.6987" ], [ "0.12006", "0.695", "0.486", "0.6953", "0.486", "0.7104", "0.12006", "0.709" ], [ "0.5156", "0.7007", "0.592", "0.7007", "0.592", "0.7153", "0.5156", "0.7153" ], [ "0.12177", "0.712", "0.486", "0.712", "0.486", "0.7256", "0.12177", "0.7256" ], [ "0.12177", "0.728", "0.4878", "0.728", "0.4878", "0.741", "0.12177", "0.741" ], [ "0.513", "0.7393", "0.546", "0.7393", "0.546", "0.7505", "0.513", "0.7505" ], [ "0.561", "0.7383", "0.754", "0.7383", "0.754", "0.752", "0.561", "0.752" ], [ "0.12006", "0.7583", "0.4878", "0.7583", "0.4878", "0.7725", "0.12006", "0.7725" ], [ "0.12006", "0.7437", "0.4878", "0.742", "0.4878", "0.7573", "0.12006", "0.7583" ], [ "0.517", "0.763", "0.8833", "0.7656", "0.8833", "0.78", "0.517", "0.778" ], [ "0.1184", "0.7754", "0.4878", "0.775", "0.4878", "0.789", "0.1184", "0.7905" ], [ "0.513", "0.781", "0.879", "0.781", "0.879", "0.7944", "0.513", "0.7944" ], [ "0.1184", "0.792", "0.4878", "0.7915", "0.4878", "0.8057", "0.1184", "0.807" ], [ "0.5156", "0.799", "0.882", "0.799", "0.882", "0.8125", "0.5156", "0.8125" ], [ "0.12006", "0.808", "0.1763", "0.808", "0.1763", "0.8223", "0.12006", "0.8223" ], [ "0.5156", "0.814", "0.882", "0.814", "0.882", "0.829", "0.5156", "0.829" ], [ "0.1394", "0.8306", "0.4878", "0.8306", "0.4878", "0.844", "0.1394", "0.844" ], [ "0.513", "0.83", "0.881", "0.83", "0.881", "0.844", "0.513", "0.844" ], [ "0.12006", "0.847", "0.4895", "0.847", "0.4895", "0.861", "0.12006", "0.861" ], [ "0.5146", "0.847", "0.881", "0.847", "0.881", "0.861", "0.5146", "0.861" ], [ "0.12006", "0.8633", "0.4895", "0.8633", "0.4895", "0.877", "0.12006", "0.877" ], [ "0.513", "0.8623", "0.881", "0.8623", "0.881", "0.877", "0.513", "0.877" ], [ "0.12177", "0.879", "0.4878", "0.879", "0.4878", "0.8926", "0.12177", "0.8926" ], [ "0.5156", "0.879", "0.8833", "0.879", "0.8833", "0.8926", "0.5156", "0.8926" ], [ "0.1184", "0.8936", "0.4895", "0.8945", "0.4895", "0.909", "0.1184", "0.9077" ], [ "0.5146", "0.8936", "0.827", "0.8945", "0.827", "0.909", "0.5146", "0.9077" ] ]

[ "Entails?", "layera", "fea:", "Feed Forward", "-", "-", "-", "layen+", "(P.H)", "Cong", "fan-", "joine", "(P2.H)", "fea", "fu", "fen", "layere", "layera", "layeri", "目", "", "layer2", "layer2", "layera", "(P.H)", "Relevance Module", "Muliayer", "AAnrae", "auer", "(P,Hga)", "(Pa,Hga)", "(Pn,Hea)", "(Pa,H)", "(P.H)(P.H)", ".", "Figure 3:", "", "Multee overview: Multee includes two main components, a relevance module, and a", "multi-layer aggregator module. Both modules use pre-trained entailment functions ( fe, and fe ). fe, is", "the full entailment model that gives entailment probability, and fe, is part of it excluding last projection", "to logits and softmax. The multi-level aggregator uses multiple copies of entailment function fe , one", "for each sub-aggregator performing a join at a different layer. Right part of figure zooms in on one such", "sub-aggregator joining at layer l .", "缩放，并将它们连接在一起以形成[PHg", "重新归一化最终矩阵：", "通过不变。", "X, = (Mhpi,[P],[Hqa]Vi", "X,= ([P],[Hqal) Vi", "Mhp- [aiMhpl; ....; QiMhpn]", "[P] = [αi[Pi]; Q2[P2]; ...; αn[Pn]]", "Mtp", "Mhp", "Y= ((P],[Hqal)", "ZAMp", "[P] =[Pi];[P2]; ..; [Pn]]", "Y =(Mhp, P,Hqa)", "对于非上下文单词嵌入，当{α}=1时，这", "会简化为ConcatPremises（等式E3）", "其中 Mhp是ith行和Mhp的jth列", "最终层（FL)：包含堆栈中的最后一层通常输出", "个向量h，然后将其用于线性层和softmax", "参Multee的多层聚合器模块在两个级别", "中以产生标签概率。这里的连接运算是前提级", "上使用连接操作：交叉关注层（CA）和最终层", "向量的加权和。所以我们有X；=hiVi和", "(FL)。两个相应的聚合器共享参数，直到两个", "Y=E,aihi", "连接层中的较低层（在本例中为CA），它们", "都在句子级别运行。在这一层之上，一个聚合", "这类似于标准的注意力机制，其中通过对缩", "器切换到在段落级别运行，在那里它有自己的", "放表示进行求和来计算出席表示。然而，当来", "非共享参数。通常，如果参Multee要在层", "自较低层的输出不是相同的形状时，这种缩放", "l，l2，..,l上聚合，则分别在层l和!上具", "加法是不可能的，如在以下情况下", "有连接的聚合器可以共享层1,·,min[e,e\"]", "交叉注意层（CA）：交叉注意是许多蕴涵和阅", "上的参数", "读理解模型的标准组成部分。该层产生三个输", "出：（i）对于每个前提Pi，我们得到一个假", "3.3", "Implementation Details", "阵 Mhpi，其中 h 是前提 P;中的令牌数；（i)", "设，以其形状为（h×pi）的前提交叉关注矩", "参Multee使用ESIM堆栈作为在SNLI和", "对于每个前提XMATHX序列Pi，。Mhpi注", "MultiNLI上预先训练的关联模块和多层聚合", "意力矩阵是通过从[H.a]到[P]的交叉注意力", "器模块的蕴涵函数。它在两个级别上使用聚", "生成的", "合，一个在交叉注意力级别（CA），一个在最", "此层中的连接操作会产生一个跨整个通道的", "后一层（FL）。参Multee中蕴涵函数的所", "交叉关注矩阵，即。e。，具有形状（h×p）", "有用法都使用相同的预训练蕴涵模型权重进行", "其中p是所有场所的令牌总数。该操作首先按", "初始化。嵌人层和BiLSTM层处理段落级上", "句子相关性权重「α缩放交叉注意矩阵，以", "下文，但更高层的处理是在前提级或段落级完", "“降低”分散注意力/不相关句子的影响，然后", "成的，具体取决于执行连接操作的位置" ]

{ "arxiv_id": "1904.09380", "language": "zh", "page_idx": 3 }

[ [ "0.11755", "0.0766", "0.266", "0.0766", "0.266", "0.09143", "0.11755", "0.09143" ], [ "0.5146", "0.0766", "0.881", "0.0766", "0.881", "0.0903", "0.5146", "0.0903" ], [ "0.513", "0.09265", "0.881", "0.09265", "0.881", "0.1057", "0.513", "0.1057" ], [ "0.1184", "0.1027", "0.4836", "0.10156", "0.4836", "0.1164", "0.1184", "0.11755" ], [ "0.5156", "0.10986", "0.882", "0.10986", "0.882", "0.12354", "0.5156", "0.12354" ], [ "0.12006", "0.12054", "0.4895", "0.12054", "0.4895", "0.1342", "0.12006", "0.1342" ], [ "0.5146", "0.1259", "0.79", "0.1259", "0.79", "0.1395", "0.5146", "0.1395" ], [ "0.12006", "0.136", "0.4895", "0.136", "0.4895", "0.1497", "0.12006", "0.1497" ], [ "0.5146", "0.1426", "0.882", "0.1426", "0.882", "0.1567", "0.5146", "0.1567" ], [ "0.12006", "0.152", "0.4878", "0.152", "0.4878", "0.165", "0.12006", "0.165" ], [ "0.5156", "0.1592", "0.882", "0.1592", "0.882", "0.1729", "0.5156", "0.1729" ], [ "0.12006", "0.1687", "0.4878", "0.1687", "0.4878", "0.1818", "0.12006", "0.1818" ], [ "0.5146", "0.1733", "0.881", "0.1733", "0.881", "0.187", "0.5146", "0.187" ], [ "0.12177", "0.1853", "0.4878", "0.1853", "0.4878", "0.199", "0.12177", "0.199" ], [ "0.5146", "0.1895", "0.7446", "0.1907", "0.7446", "0.2048", "0.5146", "0.2037" ], [ "0.12177", "0.2007", "0.4878", "0.2007", "0.4878", "0.2144", "0.12177", "0.2144" ], [ "0.511", "0.206", "0.882", "0.2073", "0.882", "0.2214", "0.511", "0.2203" ], [ "0.12177", "0.2162", "0.4878", "0.2162", "0.4878", "0.2299", "0.12177", "0.2299" ], [ "0.513", "0.2227", "0.882", "0.2239", "0.882", "0.2382", "0.513", "0.2369" ], [ "0.12006", "0.2328", "0.492", "0.2328", "0.492", "0.2465", "0.12006", "0.2465" ], [ "0.513", "0.2382", "0.882", "0.2393", "0.882", "0.2542", "0.513", "0.253" ], [ "0.12006", "0.2488", "0.4785", "0.2488", "0.4785", "0.2625", "0.12006", "0.2625" ], [ "0.5146", "0.2559", "0.8833", "0.2559", "0.8833", "0.2695", "0.5146", "0.2695" ], [ "0.12006", "0.2654", "0.4895", "0.2654", "0.4895", "0.279", "0.12006", "0.279" ], [ "0.5156", "0.2725", "0.882", "0.2725", "0.882", "0.2861", "0.5156", "0.2861" ], [ "0.12177", "0.2808", "0.4895", "0.2808", "0.4895", "0.2944", "0.12177", "0.2944" ], [ "0.513", "0.2874", "0.777", "0.2874", "0.777", "0.3005", "0.513", "0.3005" ], [ "0.12177", "0.2976", "0.4878", "0.2976", "0.4878", "0.3113", "0.12177", "0.3113" ], [ "0.513", "0.304", "0.882", "0.3047", "0.882", "0.3196", "0.513", "0.3184" ], [ "0.12177", "0.3135", "0.4878", "0.3135", "0.4878", "0.3267", "0.12177", "0.3267" ], [ "0.5156", "0.3218", "0.882", "0.3218", "0.882", "0.335", "0.5156", "0.335" ], [ "0.12006", "0.3289", "0.4878", "0.3289", "0.4878", "0.3425", "0.12006", "0.3425" ], [ "0.5146", "0.3374", "0.882", "0.3374", "0.882", "0.3508", "0.5146", "0.3508" ], [ "0.12006", "0.3433", "0.4878", "0.3433", "0.4878", "0.3582", "0.12006", "0.3582" ], [ "0.5146", "0.3516", "0.882", "0.3528", "0.882", "0.3677", "0.5146", "0.3665" ], [ "0.1184", "0.3599", "0.3787", "0.3599", "0.3787", "0.3735", "0.1184", "0.3735" ], [ "0.5146", "0.3694", "0.8833", "0.3682", "0.8833", "0.383", "0.5146", "0.3843" ], [ "0.12006", "0.3772", "0.486", "0.3772", "0.486", "0.3901", "0.12006", "0.3901" ], [ "0.513", "0.3843", "0.881", "0.3843", "0.881", "0.398", "0.513", "0.398" ], [ "0.12006", "0.3943", "0.4878", "0.3943", "0.4878", "0.408", "0.12006", "0.408" ], [ "0.5146", "0.402", "0.882", "0.402", "0.882", "0.415", "0.5146", "0.415" ], [ "0.12006", "0.4104", "0.4895", "0.4104", "0.4895", "0.4233", "0.12006", "0.4233" ], [ "0.5156", "0.4175", "0.8657", "0.4175", "0.8657", "0.4312", "0.5156", "0.4312" ], [ "0.12006", "0.427", "0.4878", "0.427", "0.4878", "0.4407", "0.12006", "0.4407" ], [ "0.532", "0.4353", "0.882", "0.4353", "0.882", "0.449", "0.532", "0.449" ], [ "0.12006", "0.4436", "0.4878", "0.4436", "0.4878", "0.4573", "0.12006", "0.4573" ], [ "0.513", "0.4495", "0.882", "0.4507", "0.882", "0.4656", "0.513", "0.4644" ], [ "0.12177", "0.459", "0.4878", "0.459", "0.4878", "0.4727", "0.12177", "0.4727" ], [ "0.1184", "0.4739", "0.4878", "0.4744", "0.4878", "0.4893", "0.1184", "0.488" ], [ "0.5156", "0.4685", "0.881", "0.4685", "0.881", "0.4785", "0.5156", "0.4785" ], [ "0.5146", "0.4827", "0.881", "0.4827", "0.881", "0.4963", "0.5146", "0.4963" ], [ "0.12177", "0.4912", "0.4878", "0.4912", "0.4878", "0.505", "0.12177", "0.505" ], [ "0.5146", "0.4988", "0.882", "0.4988", "0.882", "0.5127", "0.5146", "0.5127" ], [ "0.12006", "0.5073", "0.4878", "0.5073", "0.4878", "0.521", "0.12006", "0.521" ], [ "0.5146", "0.5156", "0.882", "0.5156", "0.882", "0.5293", "0.5146", "0.5293" ], [ "0.12177", "0.525", "0.4878", "0.525", "0.4878", "0.535", "0.12177", "0.535" ], [ "0.513", "0.532", "0.8833", "0.532", "0.8833", "0.546", "0.513", "0.546" ], [ "0.12006", "0.539", "0.4895", "0.539", "0.4895", "0.5527", "0.12006", "0.5527" ], [ "0.513", "0.5474", "0.663", "0.5474", "0.663", "0.561", "0.513", "0.561" ], [ "0.12006", "0.5547", "0.4895", "0.5547", "0.4895", "0.5693", "0.12006", "0.5693" ], [ "0.1184", "0.571", "0.1763", "0.571", "0.1763", "0.585", "0.1184", "0.585" ], [ "0.5146", "0.576", "0.6245", "0.576", "0.6245", "0.5903", "0.5146", "0.5903" ], [ "0.1184", "0.585", "0.4895", "0.585", "0.4895", "0.6025", "0.1184", "0.6025" ], [ "0.5146", "0.5986", "0.8833", "0.5986", "0.8833", "0.6133", "0.5146", "0.6133" ], [ "0.12006", "0.605", "0.4895", "0.605", "0.4895", "0.618", "0.12006", "0.618" ], [ "0.12177", "0.6206", "0.4895", "0.6206", "0.4895", "0.6343", "0.12177", "0.6343" ], [ "0.5156", "0.615", "0.882", "0.615", "0.882", "0.629", "0.5156", "0.629" ], [ "0.12177", "0.6353", "0.4895", "0.6353", "0.4895", "0.6484", "0.12177", "0.6484" ], [ "0.5156", "0.6313", "0.8833", "0.6313", "0.8833", "0.6445", "0.5156", "0.6445" ], [ "0.513", "0.6465", "0.8833", "0.6455", "0.8833", "0.66", "0.513", "0.661" ], [ "0.12006", "0.653", "0.4878", "0.653", "0.4878", "0.6665", "0.12006", "0.6665" ], [ "0.12006", "0.6685", "0.4878", "0.6685", "0.4878", "0.6816", "0.12006", "0.6816" ], [ "0.5146", "0.663", "0.8833", "0.663", "0.8833", "0.677", "0.5146", "0.677" ], [ "0.12006", "0.685", "0.4878", "0.685", "0.4878", "0.6987", "0.12006", "0.6987" ], [ "0.5146", "0.681", "0.621", "0.681", "0.621", "0.6924", "0.5146", "0.6924" ], [ "0.12006", "0.7017", "0.4878", "0.7017", "0.4878", "0.7153", "0.12006", "0.7153" ], [ "0.531", "0.697", "0.882", "0.697", "0.882", "0.7104", "0.531", "0.7104" ], [ "0.519", "0.712", "0.882", "0.712", "0.882", "0.7256", "0.519", "0.7256" ], [ "0.12006", "0.7188", "0.4878", "0.7188", "0.4878", "0.7324", "0.12006", "0.7324" ], [ "0.12177", "0.734", "0.4844", "0.734", "0.4844", "0.7446", "0.12177", "0.7446" ], [ "0.5156", "0.7285", "0.882", "0.7285", "0.882", "0.742", "0.5156", "0.742" ], [ "0.1184", "0.749", "0.4878", "0.749", "0.4878", "0.7627", "0.1184", "0.7627" ], [ "0.5146", "0.7446", "0.882", "0.7446", "0.882", "0.7583", "0.5146", "0.7583" ], [ "0.12006", "0.7656", "0.4902", "0.7656", "0.4902", "0.7793", "0.12006", "0.7793" ], [ "0.5156", "0.761", "0.882", "0.761", "0.882", "0.775", "0.5156", "0.775" ], [ "0.519", "0.777", "0.882", "0.777", "0.882", "0.7905", "0.519", "0.7905" ], [ "0.12177", "0.783", "0.486", "0.783", "0.486", "0.7935", "0.12177", "0.7935" ], [ "0.5156", "0.7935", "0.882", "0.7935", "0.882", "0.807", "0.5156", "0.807" ], [ "0.1368", "0.807", "0.4878", "0.8086", "0.4878", "0.8247", "0.1368", "0.8223" ], [ "0.5156", "0.81", "0.882", "0.81", "0.882", "0.8237", "0.5156", "0.8237" ], [ "0.1184", "0.822", "0.4458", "0.822", "0.4458", "0.835", "0.1184", "0.835" ], [ "0.517", "0.826", "0.882", "0.826", "0.882", "0.839", "0.517", "0.839" ], [ "0.1411", "0.8345", "0.4475", "0.8345", "0.4475", "0.847", "0.1411", "0.847" ], [ "0.1394", "0.8457", "0.4844", "0.847", "0.4844", "0.861", "0.1394", "0.8594" ], [ "0.5156", "0.8413", "0.882", "0.8413", "0.882", "0.855", "0.5156", "0.855" ], [ "0.12006", "0.8594", "0.4878", "0.8594", "0.4878", "0.8726", "0.12006", "0.8726" ], [ "0.513", "0.858", "0.881", "0.858", "0.881", "0.8716", "0.513", "0.8716" ], [ "0.12177", "0.8726", "0.1696", "0.8726", "0.1696", "0.883", "0.12177", "0.883" ], [ "0.1411", "0.883", "0.4307", "0.883", "0.4307", "0.8965", "0.1411", "0.8965" ], [ "0.532", "0.882", "0.882", "0.8843", "0.882", "0.8984", "0.532", "0.8965" ], [ "0.1411", "0.8965", "0.2795", "0.8965", "0.2795", "0.91", "0.1411", "0.91" ], [ "0.5146", "0.898", "0.5576", "0.898", "0.5576", "0.909", "0.5146", "0.909" ] ]

[ "4Experiments", "etal.，2017）实现中相同的ESIM（Chenetal.", "2017）超参数，并微调模型参数。我们不对我", "Datasets：我们在两个数据集上评估参", "们的任何模型执行任何超参数调整。我们微调", "Multee , OpenBookQAE(Mihaylov et al.,", "ESIM中除了嵌人层之外的所有层", "2018)和 MultiRCE(Khashabi et al., 2018)，", "ModelsCompared：我们对参Multee的", "这两个数据集都是专门为测试多个句子的推", "Glove (Pennington et al., 2014)和 ELMo (Pe-", "理而设计的。MultiRC是基于段落的多项选择", "ters etal.2018）嵌人进行了实验，并与以下", "QA数据集，来自不同的主题，其中问题可根", "三种类型的系统进行了比较：", "据段落中的信息回答。在MultiRC中，每个问", "（A）使用蕴涵作为黑箱的基线我们以两种方", "题可以有多个正确答案选择，因此可以将其视", "式将预先训练的蕴涵模型用作黑箱：连接前提", "为二进制分类任务（每个答案选择一个预测），", "（Concat）和具有最大操作（Max）的聚合句子", "在开发/测试集中有4,8484,583个示例。另", "级别决策。这两个模型还在SNLI和MultiNLI", "方面，OpenBookQA有多项选择的科学问题，", "数据集上进行了预训练，并在具有相同预处理", "只有一个正确的答案选择，没有相关的段落。", "的目标QA数据集上进行了微调", "因此，此数据集需要从辅助资源中检索相关", "(B）以前发布的结果：对于MultiRC，有两个", "事实，包括与论文一起发布的公开事实书以及", "发布的基线：IR（信息检索）和LR（逻辑回归）。", "其他来源，例如WordNetE（Miller，1995）和", "在这个相对较小的数据集上，这些简单的模型", "ConceptNet(Speer and Havasi, 2012)。它在", "被证明是强大的基线。对于OpenBookQA，我", "开发和测试集中包含500个问题", "们报告了（Mihaylovetal.，2018）发布的基", "预处理：对于每个问题和答案选择，我们使用", "线：与ELMo的问题匹配（QM+ELMo），与", "SciTailE(Khotetal.，2018）构造中使用的脚", "ELMo的问答ESIM（ESIM+ELMo）以及他", "本的修改版本创建一个答案假设语句。我们", "们在知识增强阅读器（KER）中的最佳结果", "编写了一些规则，以更好地将问题和答案转", "基于（C）大型Transformer的模型：我们与", "换为假设。我们还使用特殊的开始和结束标记", "OpenAI-Transformer(OFT）进行了比较，对", "@@@answer和answer@@@分别在假设中标", "记答案的范围3。对于MultiRC，我们还应", "大规模语言建模任务进行了预训练，并对相", "应的数据集进行了微调。发表了这些转换器结", "用了现成的协指解析模型4，并在它们解析", "果的同期作品8也在大规模阅读理解数据集", "为不同句子5中出现的代词时替换提及。对于", "RACEE(Laietal.，2017)上进一步微调了该", "OpenBookQA，我们使用与OpenBookQA6作", "转换器，然后使用其方法对目标QA数据集进", "者发布的完全相同的检索，并使用OpenBook", "行微调，阅读策略", "和WordNet作为知识源，每个查询检索前5", "个句子", "4.1Results", "培训Multee：对于OpenBookQA，我们", "表1总结了所有型号的性能。参Multee优", "对与4个答案选择相对应的标签使用交叉熵损", "失。对于MultiRC，我们分别为每个答案选择", "于在相同数据，以前发布的基线，OpenAI变压", "使用二进制交叉损失，因为在MultiRC中，", "器模型上预先训练的黑盒包含基线（Concat和", "Max)。我们注意到，OpenBookQA和MultiRC", "每个问题可以有多个正确的答案选择。蕴涵组", "件在句子级蕴涵任务上进行预培训，然后作为", "的基线精度周围的95%置信区间为4。3%和", "端到端QA培训的一部分进行微调。MultiRC", "1。3%，分别", "数据集包括句子级相关性标签。我们使用二", "在OpenBookQA测试集上，带有GloVe的", "参Multee的性能优于OpenAI转换器的集", "进制交叉摘损失来监督句子相关性模块，以", "便在可用时预测这些相关性标签。我们使用", "成版本3。精度0分。它还优于阅读策略系", "PyTorch (Paszke et al., 2017)和 AllenNLP 来", "统的单一模型版本，可与它们的集成版本相", "实现我们的模型，并在Beaker？上运行它们。", "媲美。在MultiRC开发集上，带有ELMo的", "参Multee比OpenAI转换器的集成版本高", "对于预训练，我们使用与AllenNLP（Gardner", "出1。F1a中9点,2。7在F1m和6。3在EM。", "3Answer span marking gave substantial gains for all", "它还优于阅读策略系统的单一模型版本，可", "entailment based models including the baselines", "与它们的集成版本相媲美。回想一下，除了我", "4https://github.com/huggingface/neuralcoref", "5It is hard to learn co-reference, as these target", "们在这里使用的目标QA数据集之外，阅读", "datasets are too small to learn this in an end-to-end", "策略结果是通过对另一个更大的QA数据集", "fashion", "6https://github.com/allenai/OpenBookQA", "8Published on arXiv on Oct 31, 2018 (Sun et al.,", "https://beaker.org/", "2019)" ]

{ "arxiv_id": "1904.09380", "language": "zh", "page_idx": 4 }

[ [ "0.293", "0.0956", "0.706", "0.0956", "0.706", "0.10986", "0.293", "0.10986" ], [ "0.1318", "0.1426", "0.869", "0.1426", "0.869", "0.1567", "0.1318", "0.1567" ], [ "0.168", "0.1561", "0.8345", "0.1567", "0.8345", "0.1746", "0.168", "0.1733" ], [ "0.157", "0.177", "0.8564", "0.177", "0.8564", "0.1907", "0.157", "0.1907" ], [ "0.3032", "0.1936", "0.7", "0.1936", "0.7", "0.2073", "0.3032", "0.2073" ], [ "0.2627", "0.2625", "0.3435", "0.2625", "0.3435", "0.2766", "0.2627", "0.2766" ], [ "0.1436", "0.2861", "0.4575", "0.2874", "0.4575", "0.3005", "0.1436", "0.2998" ], [ "0.147", "0.3005", "0.4592", "0.3005", "0.4592", "0.3142", "0.147", "0.3142" ], [ "0.1453", "0.3154", "0.4592", "0.3154", "0.4592", "0.3289", "0.1453", "0.3289" ], [ "0.1453", "0.3289", "0.4592", "0.3289", "0.4592", "0.3425", "0.1453", "0.3425" ], [ "0.1453", "0.3425", "0.4568", "0.3425", "0.4568", "0.3557", "0.1453", "0.3557" ], [ "0.147", "0.3594", "0.4575", "0.3594", "0.4575", "0.3694", "0.147", "0.3694" ], [ "0.1453", "0.3718", "0.4575", "0.3718", "0.4575", "0.386", "0.1453", "0.386" ], [ "0.1453", "0.386", "0.4592", "0.386", "0.4592", "0.3997", "0.1453", "0.3997" ], [ "0.147", "0.402", "0.4568", "0.402", "0.4568", "0.412", "0.147", "0.412" ], [ "0.1453", "0.4146", "0.4575", "0.4146", "0.4575", "0.4275", "0.1453", "0.4275" ], [ "0.147", "0.4287", "0.4592", "0.4287", "0.4592", "0.4424", "0.147", "0.4424" ], [ "0.1453", "0.4436", "0.4592", "0.4436", "0.4592", "0.4573", "0.1453", "0.4573" ], [ "0.1486", "0.459", "0.4575", "0.459", "0.4575", "0.4697", "0.1486", "0.4697" ], [ "0.1453", "0.4714", "0.4568", "0.4714", "0.4568", "0.485", "0.1453", "0.485" ], [ "0.513", "0.4785", "0.882", "0.4785", "0.882", "0.4934", "0.513", "0.4934" ], [ "0.1453", "0.485", "0.4575", "0.485", "0.4575", "0.4988", "0.1453", "0.4988" ], [ "0.511", "0.4934", "0.881", "0.4934", "0.881", "0.5073", "0.511", "0.5073" ], [ "0.1453", "0.4995", "0.4568", "0.4995", "0.4568", "0.513", "0.1453", "0.513" ], [ "0.511", "0.5073", "0.882", "0.5083", "0.882", "0.5215", "0.511", "0.521" ], [ "0.1436", "0.513", "0.4575", "0.513", "0.4575", "0.527", "0.1436", "0.527" ], [ "0.1453", "0.528", "0.4575", "0.528", "0.4575", "0.5415", "0.1453", "0.5415" ], [ "0.511", "0.5225", "0.6943", "0.5225", "0.6943", "0.536", "0.511", "0.536" ], [ "0.1453", "0.542", "0.4592", "0.542", "0.4592", "0.5557", "0.1453", "0.5557" ], [ "0.1453", "0.5557", "0.4568", "0.557", "0.4568", "0.571", "0.1453", "0.5693" ], [ "0.1453", "0.571", "0.4592", "0.571", "0.4592", "0.5845", "0.1453", "0.5845" ], [ "0.5146", "0.5654", "0.881", "0.5654", "0.881", "0.579", "0.5146", "0.579" ], [ "0.1453", "0.585", "0.4575", "0.585", "0.4575", "0.5986", "0.1453", "0.5986" ], [ "0.511", "0.58", "0.879", "0.58", "0.879", "0.5933", "0.511", "0.5933" ], [ "0.1453", "0.5986", "0.4575", "0.5986", "0.4575", "0.6123", "0.1453", "0.6123" ], [ "0.511", "0.5967", "0.8774", "0.5967", "0.8774", "0.61", "0.511", "0.61" ], [ "0.147", "0.6133", "0.4592", "0.6133", "0.4592", "0.627", "0.147", "0.627" ], [ "0.5146", "0.6133", "0.881", "0.6133", "0.881", "0.627", "0.5146", "0.627" ], [ "0.1453", "0.627", "0.4575", "0.627", "0.4575", "0.64", "0.1453", "0.64" ], [ "0.513", "0.63", "0.881", "0.63", "0.881", "0.6436", "0.513", "0.6436" ], [ "0.147", "0.641", "0.4592", "0.641", "0.4592", "0.655", "0.147", "0.655" ], [ "0.511", "0.6455", "0.559", "0.6455", "0.559", "0.66", "0.511", "0.66" ], [ "0.1453", "0.6562", "0.4575", "0.6562", "0.4575", "0.6694", "0.1453", "0.6694" ], [ "0.531", "0.662", "0.881", "0.661", "0.881", "0.676", "0.531", "0.677" ], [ "0.1436", "0.6704", "0.4568", "0.6694", "0.4568", "0.683", "0.1436", "0.684" ], [ "0.1453", "0.684", "0.4575", "0.684", "0.4575", "0.6978", "0.1453", "0.6978" ], [ "0.5146", "0.6787", "0.881", "0.6787", "0.881", "0.6924", "0.5146", "0.6924" ], [ "0.511", "0.6934", "0.879", "0.6934", "0.879", "0.707", "0.511", "0.707" ], [ "0.147", "0.6997", "0.4568", "0.6997", "0.4568", "0.7104", "0.147", "0.7104" ], [ "0.1453", "0.7134", "0.3047", "0.712", "0.3047", "0.7256", "0.1453", "0.727" ], [ "0.5146", "0.7114", "0.881", "0.7114", "0.881", "0.7246", "0.5146", "0.7246" ], [ "0.511", "0.7256", "0.881", "0.7256", "0.881", "0.7393", "0.511", "0.7393" ], [ "0.1184", "0.7393", "0.1335", "0.7393", "0.1335", "0.7476", "0.1184", "0.7476" ], [ "0.1318", "0.741", "0.1427", "0.741", "0.1427", "0.745", "0.1318", "0.745" ], [ "0.1511", "0.737", "0.2612", "0.737", "0.2612", "0.749", "0.1511", "0.749" ], [ "0.513", "0.742", "0.879", "0.742", "0.879", "0.756", "0.513", "0.756" ], [ "0.1184", "0.761", "0.486", "0.7603", "0.486", "0.775", "0.1184", "0.7754" ], [ "0.513", "0.7603", "0.642", "0.7603", "0.642", "0.774", "0.513", "0.774" ], [ "0.11755", "0.777", "0.4836", "0.777", "0.4836", "0.7905", "0.11755", "0.7905" ], [ "0.534", "0.777", "0.882", "0.777", "0.882", "0.7905", "0.534", "0.7905" ], [ "0.1184", "0.7935", "0.486", "0.7935", "0.486", "0.807", "0.1184", "0.807" ], [ "0.5146", "0.7935", "0.881", "0.7935", "0.881", "0.807", "0.5146", "0.807" ], [ "0.11755", "0.81", "0.486", "0.81", "0.486", "0.8237", "0.11755", "0.8237" ], [ "0.5146", "0.8086", "0.882", "0.8086", "0.882", "0.8223", "0.5146", "0.8223" ], [ "0.1184", "0.826", "0.4668", "0.826", "0.4668", "0.8403", "0.1184", "0.8403" ], [ "0.511", "0.8247", "0.881", "0.8247", "0.881", "0.8384", "0.511", "0.8384" ], [ "0.4626", "0.8306", "0.4668", "0.8306", "0.4668", "0.836", "0.4626", "0.836" ], [ "0.1184", "0.843", "0.413", "0.843", "0.413", "0.8555", "0.1184", "0.8555" ], [ "0.5146", "0.8403", "0.881", "0.8403", "0.881", "0.855", "0.5146", "0.855" ], [ "0.1368", "0.8594", "0.4836", "0.8594", "0.4836", "0.8726", "0.1368", "0.8726" ], [ "0.511", "0.857", "0.879", "0.857", "0.879", "0.8706", "0.511", "0.8706" ], [ "0.1184", "0.8745", "0.4844", "0.8745", "0.4844", "0.888", "0.1184", "0.888" ], [ "0.5146", "0.8726", "0.684", "0.8726", "0.684", "0.8857", "0.5146", "0.8857" ], [ "0.12006", "0.89", "0.4844", "0.89", "0.4844", "0.905", "0.12006", "0.905" ], [ "0.531", "0.891", "0.882", "0.891", "0.882", "0.905", "0.531", "0.905" ], [ "0.1184", "0.9067", "0.4844", "0.9067", "0.4844", "0.9204", "0.1184", "0.9204" ], [ "0.511", "0.9067", "0.881", "0.9067", "0.881", "0.9204", "0.511", "0.9204" ] ]

[ "OSCaR：对象状态标题和状态更改表示", "Nguyen Nguyen', Jing Bi', Ali Vosoughi', Yapeng Tian?, Pooyan Fazlr, Chenliang Xu", "'University of Rochester, 2University of Texas at Dallas, 3Arizona State University", "( nguyen.nguyen, jing.bi, ali.vosoughi, chenliang.xu } @rochester.edu,", "[email protected], [email protected]", "Abstract", "The capability of intelligent models to extrapo", "late and comprehend changes in object states is", "a crucial yet demanding aspect of AI research,", "particularly through the lens of human interac-", "tion in real-world settings. This task involves", "describingcomplexvisualenvironments,iden-", "tifying active objects, and interpreting their", "changes as conveyed through language. Tra-", "ditionalmethods,whichisolateobjectcaption", "ing and state change detection, offer a limited", "view of dynamic environments. Moreover, re-", "lying on a small set of symbolic words to rep-", "resent changes has restricted the expressive-", "ness of language. To address these challenges,", "Figure 1: Surpassing prior models in aligning with hu", "in this paper, we introduce the Object State", "man judgements. Our method achieves near parity with", "Captioning and State Change Representation", "GPT-4V ratings across helpfulness, accuracy, reason-", "(OSCaR) dataset and benchmark. OSCaR", "consists of 14,084 annotated video segments", "ing, and other key metrics.", "with nearly 1,o00 unique objects from vari-", "ous egocentric video collections. It sets a new", "testbed for evaluating Multimodal Large Lan-", "是没有问题的。此外，考虑到“切面包”的说", "guage Models (MLLMs). Our experiments", "法，我们自然会想象会发生什么状态变化。然", "demonstrate that while MLLMs show some", "而，尽管在知识表示，自动推理和机器学习方", "skill, they lack a full understanding of ob-", "面取得了巨大进步，但人工代理仍然缺乏对物", "ject state changes. The benchmark includes a", "理世界（Gaoetal.,2018）的天真的因果关系的", "fine-tuned model that, despite initial capabili-", "理解。", "ties, requires significant improvements in ac-", "想象一个人工代理在物理世界中与人类合作", "curacy and generalization ability for effective", "understanding of these changes. Our code and", "的场景，他们将需要了解物理动作的效果来推", "理，学习和帮助人类（Bietal.2023）。为了使", "datasetareavailableathttps://github.com/", "nguyennm1024/0sCaR", "机器具有这种功能，本文介绍了一种新颖的基", "准测试，该基准测试侧重于从自我中心的视觉", "1", "", "Introduction", "输入中理解对象状态的变化，该基准测试具有", "自然语言处理（NLP）领域已经超越了单纯", "人眼镜头的优势", "的文本解释和生成，进入了必须理解和与物", "了解对象状态变化不仅是一项复杂的任务，", "理世界互动的领域。从研究因果推理E(Gao", "而且是许多其他任务的实用和基础，例如帮", "etal,2018）到建立因果预测E(Gaoetal.,2016;", "助智能代理了解环境动态并完成任务E(Pad-", "Alayracetal.，2017）的世界模型，研究人员", "makumar et al., 2023; Sarch et al., 2023; Merullo", "", "直在研究物理世界中的因果关系问题", "etal.,2022），跟踪对话框（Leetal.,2022）的状", "在本文中，我们研究了具体动作与该动作引", "态，为复杂问题的知识表示创建因果图并回", "起的对象状态变化之间的非常基本的因果关", "答E(Ates et al.,2020)", "系。例如，给定一个如图2所示的图像，我们", "对对象状态变化进行建模需要两种能力：1)", "作为人类，理解哪个对象正在积极地与之交互", "场景理解，这涉及通过以对象为中心的镜头解" ]

{ "arxiv_id": "2402.17128", "language": "zh", "page_idx": 1 }

[ [ "0.2007", "0.0968", "0.287", "0.0968", "0.287", "0.10394", "0.2007", "0.10394" ], [ "0.3418", "0.0944", "0.3745", "0.0944", "0.3745", "0.1027", "0.3418", "0.1027" ], [ "0.6055", "0.0974", "0.6304", "0.0974", "0.6304", "0.10156", "0.6055", "0.10156" ], [ "0.1696", "0.1057", "0.1797", "0.11224", "0.1755", "0.1164", "0.1646", "0.10986" ], [ "0.79", "0.1069", "0.797", "0.1069", "0.797", "0.10986", "0.79", "0.10986" ], [ "0.1621", "0.2239", "0.1797", "0.2239", "0.1797", "0.2465", "0.1621", "0.2465" ], [ "0.11755", "0.284", "0.881", "0.284", "0.881", "0.2976", "0.11755", "0.2976" ], [ "0.1184", "0.2998", "0.879", "0.2998", "0.879", "0.31", "0.1184", "0.31" ], [ "0.1184", "0.3123", "0.882", "0.3123", "0.882", "0.326", "0.1184", "0.326" ], [ "0.11584", "0.3267", "0.8506", "0.3267", "0.8506", "0.3403", "0.11584", "0.3403" ], [ "0.1184", "0.3677", "0.4844", "0.3677", "0.4844", "0.3813", "0.1184", "0.3813" ], [ "0.5483", "0.3682", "0.586", "0.3682", "0.586", "0.3813", "0.5483", "0.3813" ], [ "0.12006", "0.383", "0.486", "0.383", "0.486", "0.3967", "0.12006", "0.3967" ], [ "0.12006", "0.3997", "0.1914", "0.3997", "0.1914", "0.411", "0.12006", "0.411" ], [ "0.5356", "0.3967", "0.882", "0.3967", "0.882", "0.411", "0.5356", "0.411" ], [ "0.1368", "0.415", "0.4844", "0.415", "0.4844", "0.4287", "0.1368", "0.4287" ], [ "0.5474", "0.412", "0.882", "0.4133", "0.882", "0.4275", "0.5474", "0.427" ], [ "0.1184", "0.4312", "0.4844", "0.4312", "0.4844", "0.4448", "0.1184", "0.4448" ], [ "0.55", "0.4287", "0.882", "0.4287", "0.882", "0.4424", "0.55", "0.4424" ], [ "0.11584", "0.4478", "0.4836", "0.4478", "0.4836", "0.4614", "0.11584", "0.4614" ], [ "0.5483", "0.4453", "0.882", "0.4453", "0.882", "0.4602", "0.5483", "0.4602" ], [ "0.12006", "0.4644", "0.4844", "0.4644", "0.4844", "0.478", "0.12006", "0.478" ], [ "0.55", "0.462", "0.8833", "0.462", "0.8833", "0.4756", "0.55", "0.4756" ], [ "0.1184", "0.4797", "0.4844", "0.4797", "0.4844", "0.4934", "0.1184", "0.4934" ], [ "0.5474", "0.4768", "0.69", "0.478", "0.69", "0.4922", "0.5474", "0.4912" ], [ "0.1184", "0.4946", "0.4844", "0.4954", "0.4844", "0.5103", "0.1184", "0.509" ], [ "0.1184", "0.511", "0.4844", "0.511", "0.4844", "0.5254", "0.1184", "0.5254" ], [ "0.5146", "0.5127", "0.5225", "0.5127", "0.5225", "0.5186", "0.5146", "0.5186" ], [ "0.544", "0.5073", "0.678", "0.5083", "0.678", "0.5225", "0.544", "0.5215" ], [ "0.11755", "0.528", "0.4836", "0.528", "0.4836", "0.5415", "0.11755", "0.5415" ], [ "0.513", "0.536", "0.876", "0.536", "0.876", "0.55", "0.513", "0.55" ], [ "0.1184", "0.5444", "0.4844", "0.5444", "0.4844", "0.558", "0.1184", "0.558" ], [ "0.5146", "0.552", "0.881", "0.552", "0.881", "0.5654", "0.5146", "0.5654" ], [ "0.12177", "0.561", "0.1545", "0.561", "0.1545", "0.5737", "0.12177", "0.5737" ], [ "0.513", "0.5664", "0.881", "0.568", "0.881", "0.582", "0.513", "0.5806" ], [ "0.1368", "0.5767", "0.4844", "0.5767", "0.4844", "0.5903", "0.1368", "0.5903" ], [ "0.513", "0.5845", "0.881", "0.5845", "0.881", "0.597", "0.513", "0.597" ], [ "0.1184", "0.5933", "0.4844", "0.5933", "0.4844", "0.607", "0.1184", "0.607" ], [ "0.513", "0.5986", "0.8774", "0.5986", "0.8774", "0.6123", "0.513", "0.6123" ], [ "0.12006", "0.6094", "0.4836", "0.6094", "0.4836", "0.6226", "0.12006", "0.6226" ], [ "0.5146", "0.616", "0.881", "0.616", "0.881", "0.63", "0.5146", "0.63" ], [ "0.12006", "0.6245", "0.4844", "0.6245", "0.4844", "0.638", "0.12006", "0.638" ], [ "0.511", "0.632", "0.881", "0.6313", "0.881", "0.6455", "0.511", "0.6465" ], [ "0.1184", "0.641", "0.4844", "0.641", "0.4844", "0.655", "0.1184", "0.655" ], [ "0.511", "0.6484", "0.881", "0.6484", "0.881", "0.662", "0.511", "0.662" ], [ "0.1184", "0.6567", "0.4844", "0.6567", "0.4844", "0.6704", "0.1184", "0.6704" ], [ "0.511", "0.663", "0.881", "0.6646", "0.881", "0.6787", "0.511", "0.6777" ], [ "0.12006", "0.6733", "0.4844", "0.6733", "0.4844", "0.687", "0.12006", "0.687" ], [ "0.513", "0.681", "0.882", "0.681", "0.882", "0.695", "0.513", "0.695" ], [ "0.11755", "0.6885", "0.4836", "0.6885", "0.4836", "0.7017", "0.11755", "0.7017" ], [ "0.513", "0.697", "0.881", "0.697", "0.881", "0.7104", "0.513", "0.7104" ], [ "0.1184", "0.706", "0.4844", "0.706", "0.4844", "0.7197", "0.1184", "0.7197" ], [ "0.5146", "0.712", "0.881", "0.712", "0.881", "0.727", "0.5146", "0.727" ], [ "0.11755", "0.7227", "0.4844", "0.72", "0.4844", "0.734", "0.11755", "0.7363" ], [ "0.5146", "0.7285", "0.879", "0.7285", "0.879", "0.742", "0.5146", "0.742" ], [ "0.1368", "0.7383", "0.3838", "0.7383", "0.3838", "0.752", "0.1368", "0.752" ], [ "0.513", "0.7437", "0.879", "0.7446", "0.879", "0.759", "0.513", "0.7583" ], [ "0.513", "0.759", "0.879", "0.761", "0.879", "0.775", "0.513", "0.7725" ], [ "0.1377", "0.7666", "0.486", "0.7666", "0.486", "0.78", "0.1377", "0.78" ], [ "0.513", "0.7754", "0.879", "0.7754", "0.879", "0.789", "0.513", "0.789" ], [ "0.1511", "0.782", "0.4902", "0.781", "0.4902", "0.796", "0.1511", "0.7964" ], [ "0.1511", "0.7964", "0.2585", "0.799", "0.2585", "0.8135", "0.1511", "0.811" ], [ "0.5146", "0.7935", "0.881", "0.7935", "0.881", "0.807", "0.5146", "0.807" ], [ "0.5146", "0.81", "0.881", "0.81", "0.881", "0.8237", "0.5146", "0.8237" ], [ "0.1368", "0.8247", "0.4895", "0.826", "0.4895", "0.8403", "0.1368", "0.839" ], [ "0.5146", "0.826", "0.754", "0.826", "0.754", "0.839", "0.5146", "0.839" ], [ "0.1528", "0.843", "0.4878", "0.843", "0.4878", "0.8555", "0.1528", "0.8555" ], [ "0.5146", "0.843", "0.882", "0.843", "0.882", "0.8555", "0.5146", "0.8555" ], [ "0.1528", "0.8594", "0.486", "0.8594", "0.486", "0.8726", "0.1528", "0.8726" ], [ "0.5146", "0.86", "0.8774", "0.86", "0.8774", "0.8706", "0.5146", "0.8706" ], [ "0.1545", "0.8745", "0.486", "0.8745", "0.486", "0.888", "0.1545", "0.888" ], [ "0.513", "0.8735", "0.879", "0.8735", "0.879", "0.888", "0.513", "0.888" ], [ "0.1528", "0.891", "0.4895", "0.891", "0.4895", "0.905", "0.1528", "0.905" ], [ "0.5146", "0.891", "0.864", "0.891", "0.864", "0.905", "0.5146", "0.905" ], [ "0.1528", "0.9067", "0.4902", "0.9067", "0.4902", "0.9214", "0.1528", "0.9214" ], [ "0.511", "0.9062", "0.881", "0.905", "0.881", "0.9194", "0.511", "0.9204" ] ]

[ "eatance.giveinsightino.th", "detail way", "hecut", "", "", "", "Figure 2: OsCaR's description of state, state change, and illustration of reasoning. State description involves the", "characterization of a specific region of interest within the video and the associated activity.State changeentails the", "description of the evolution of a system over a defined temporal sequence. Furthermore, the analysis of the state", "of an object is centered on comprehending and elucidating the mechanisms underlying the object's evolution.", "析世界；2）因果效应理解，这需要通过观察之", "cess.", "前，期间和之后的图像来识别可能的动作及其", "效果动作", ": Our paper introduces OSCaR, a novel dataset", "以前的研究工作集中在为地面变化和状态构", "and a benchmark leveraged by the power of", "建符号表示（Wuetal.,2023;Zellers et al.,2021;", "GPT-4V that contains different tasks for ob-", "NagarajanandGrauman，2018）。但是，鉴于对", "ject state understanding, including visual cap-", "象及其状态的多样性和复杂性，受上下文和时", "tioning, visual question answering visual dia-", "间因素的影响，仅符号表示就不足。本文建议", "log, and reasoning.", "使用自然语言作为此任务的更具表现力和直观", "的媒介。这种方法不仅使人类和AI系统之间", "2", "Related Works", "对视觉内容的理解保持一致，而且还增强了它", "对象状态变化：定位和识别对象状态的变化", "们之间的交流，提供了比单峰模型更丰富的上", "在程序规划E(Bietal.2021），机器人和视频", "下文", "动作理解E(Duet al.,2023;Zhong et al.,2023;", "从本质上讲，我们将场景理解形成为以对象", "Tang et al., 2023b; Wang et al., 2023; Song et al.)", "为中心的视觉字幕问题。我们可以利用自然语", "等应用中发挥关键作用。识别对象状态变化需", "言来描述对象和可能发生的任何变化。另一方", "要通过理解它们的因果关系来共同发现状态", "面，理解因果效应的能力形成为基于3个图像", "和动作，如先前的作品E(Alayracetal.，2017;", "的视觉问答问题：动作之前，期间和之后。我", "Liu et al., 2017; Soucek et al., 2022; Naeem et al.,", "们的数据集和实验在未来的研究中展示了跨各", "2021）中所述。最近，已经提出了一种自监督", "个领域的可扩展应用的巨大潜力。在进行这项", "方法来联合定位来自噜杂的未修剪的长视频E", "研究的同时，还进行了另一项研究，以了解使", "(Souceketal.2022）的时间上的动作和状态变", "用不同方法的对象状态变化E(Xueetal.2024)", "化。此外，(Saini etal.,2023）为对象状态的", "。这表明了这一研究方向的重要性和巨大潜力", "生成引人了一个新的基准，但它们的重点非常", "总之，我们的贡献有三个方面：", "限于切割动作和一个小的数据集。但是，以前", "的研究通常将场景理解与对象状态变化识别分", ": We introduce a new problem to understand", "开，并且倾向于在封闭世界假设下运行，这限", "states and state changes of object through nat-", "ural language.", "制了它们在现实世界场景中的适用性。我们的", "研究旨在通过整合自我中心的观点和语言来弥", ". We present a method to generate good-", "合人类和机器感知之间的差距", "quality visual instructions guided by simple", "多模态大型语言模型：大型语言模型（LLM）", "annotations, applicable to both images and", "的最新进展（Ouyangetal.,2022：Touvronetal.", "videos, advancing future research in visual", "2023; Chiang et al., 2023; Chung et al., 2022) 在", "instruction tuning. Our pipeline provides a", "语言理解和生成方面取得了重大成就。这", "good starting point for the data collection pro-", "进展引发了人们对创建MLLM的兴趣，这些" ]

{ "arxiv_id": "2402.17128", "language": "zh", "page_idx": 2 }

[ [ "0.1184", "0.0873", "0.4878", "0.0873", "0.4878", "0.10095", "0.1184", "0.10095" ], [ "0.513", "0.0873", "0.881", "0.0873", "0.881", "0.10095", "0.513", "0.10095" ], [ "0.1184", "0.1027", "0.4895", "0.1027", "0.4895", "0.1164", "0.1184", "0.1164" ], [ "0.513", "0.1027", "0.881", "0.1027", "0.881", "0.1164", "0.513", "0.1164" ], [ "0.1184", "0.1194", "0.4878", "0.1194", "0.4878", "0.133", "0.1184", "0.133" ], [ "0.5156", "0.12054", "0.879", "0.12054", "0.879", "0.1312", "0.5156", "0.1312" ], [ "0.1184", "0.1354", "0.486", "0.1354", "0.486", "0.1484", "0.1184", "0.1484" ], [ "0.513", "0.1354", "0.638", "0.1354", "0.638", "0.1484", "0.513", "0.1484" ], [ "0.11755", "0.1509", "0.4836", "0.152", "0.4836", "0.165", "0.11755", "0.1644" ], [ "0.1184", "0.1675", "0.4844", "0.1675", "0.4844", "0.1812", "0.1184", "0.1812" ], [ "0.531", "0.1633", "0.881", "0.1633", "0.881", "0.177", "0.531", "0.177" ], [ "0.1184", "0.1829", "0.4844", "0.1829", "0.4844", "0.1965", "0.1184", "0.1965" ], [ "0.513", "0.1799", "0.881", "0.1799", "0.881", "0.1936", "0.513", "0.1936" ], [ "0.11755", "0.199", "0.486", "0.199", "0.486", "0.212", "0.11755", "0.212" ], [ "0.5146", "0.1953", "0.881", "0.1953", "0.881", "0.209", "0.5146", "0.209" ], [ "0.1184", "0.2156", "0.4878", "0.2156", "0.4878", "0.2286", "0.1184", "0.2286" ], [ "0.513", "0.212", "0.881", "0.212", "0.881", "0.2257", "0.513", "0.2257" ], [ "0.1184", "0.2322", "0.4492", "0.2322", "0.4492", "0.2452", "0.1184", "0.2452" ], [ "0.5146", "0.2299", "0.879", "0.2299", "0.879", "0.2405", "0.5146", "0.2405" ], [ "0.513", "0.2434", "0.879", "0.2446", "0.879", "0.2583", "0.513", "0.257" ], [ "0.11584", "0.2625", "0.3284", "0.2625", "0.3284", "0.2756", "0.11584", "0.2756" ], [ "0.513", "0.26", "0.881", "0.26", "0.881", "0.2737", "0.513", "0.2737" ], [ "0.513", "0.2766", "0.8774", "0.2766", "0.8774", "0.2903", "0.513", "0.2903" ], [ "0.1184", "0.2903", "0.4844", "0.2903", "0.4844", "0.304", "0.1184", "0.304" ], [ "0.12006", "0.306", "0.4844", "0.306", "0.4844", "0.3196", "0.12006", "0.3196" ], [ "0.1184", "0.3225", "0.486", "0.3225", "0.486", "0.3362", "0.1184", "0.3362" ], [ "0.1184", "0.3384", "0.4844", "0.3384", "0.4844", "0.3516", "0.1184", "0.3516" ], [ "0.513", "0.3467", "0.797", "0.3467", "0.797", "0.3599", "0.513", "0.3599" ], [ "0.1184", "0.354", "0.4844", "0.354", "0.4844", "0.3677", "0.1184", "0.3677" ], [ "0.11755", "0.3694", "0.4844", "0.3694", "0.4844", "0.383", "0.11755", "0.383" ], [ "0.1184", "0.3872", "0.346", "0.3872", "0.346", "0.401", "0.1184", "0.401" ], [ "0.513", "0.3938", "0.881", "0.3938", "0.881", "0.4067", "0.513", "0.4067" ], [ "0.5146", "0.4092", "0.879", "0.4092", "0.879", "0.4229", "0.5146", "0.4229" ], [ "0.11755", "0.4163", "0.3015", "0.4163", "0.3015", "0.43", "0.11755", "0.43" ], [ "0.5146", "0.427", "0.876", "0.427", "0.876", "0.437", "0.5146", "0.437" ], [ "0.1184", "0.4395", "0.4878", "0.4395", "0.4878", "0.4531", "0.1184", "0.4531" ], [ "0.513", "0.4424", "0.8774", "0.4424", "0.8774", "0.4531", "0.513", "0.4531" ], [ "0.11755", "0.4548", "0.486", "0.4548", "0.486", "0.4685", "0.11755", "0.4685" ], [ "0.511", "0.4573", "0.881", "0.4573", "0.881", "0.4702", "0.511", "0.4702" ], [ "0.11755", "0.4702", "0.4844", "0.4702", "0.4844", "0.484", "0.11755", "0.484" ], [ "0.5156", "0.4744", "0.879", "0.4744", "0.879", "0.485", "0.5156", "0.485" ], [ "0.1184", "0.4868", "0.4844", "0.4868", "0.4844", "0.5005", "0.1184", "0.5005" ], [ "0.513", "0.4893", "0.881", "0.4893", "0.881", "0.503", "0.513", "0.503" ], [ "0.11755", "0.503", "0.486", "0.502", "0.486", "0.5156", "0.11755", "0.5166" ], [ "0.513", "0.505", "0.881", "0.505", "0.881", "0.5186", "0.513", "0.5186" ], [ "0.1184", "0.5195", "0.4844", "0.5195", "0.4844", "0.533", "0.1184", "0.533" ], [ "0.511", "0.5195", "0.881", "0.521", "0.881", "0.535", "0.511", "0.5337" ], [ "0.11755", "0.5337", "0.4844", "0.535", "0.4844", "0.549", "0.11755", "0.5474" ], [ "0.511", "0.5376", "0.8774", "0.5376", "0.8774", "0.5474", "0.511", "0.5474" ], [ "0.1184", "0.552", "0.4844", "0.552", "0.4844", "0.5654", "0.1184", "0.5654" ], [ "0.5146", "0.554", "0.881", "0.554", "0.881", "0.568", "0.5146", "0.568" ], [ "0.1184", "0.5684", "0.4844", "0.5684", "0.4844", "0.582", "0.1184", "0.582" ], [ "0.511", "0.5684", "0.879", "0.5684", "0.879", "0.582", "0.511", "0.582" ], [ "0.12006", "0.586", "0.3728", "0.586", "0.3728", "0.5967", "0.12006", "0.5967" ], [ "0.5205", "0.5874", "0.879", "0.5874", "0.879", "0.597", "0.5205", "0.597" ], [ "0.1368", "0.601", "0.486", "0.601", "0.486", "0.6143", "0.1368", "0.6143" ], [ "0.513", "0.6016", "0.882", "0.6016", "0.882", "0.615", "0.513", "0.615" ], [ "0.11755", "0.616", "0.486", "0.6177", "0.486", "0.6313", "0.11755", "0.63" ], [ "0.5146", "0.619", "0.756", "0.619", "0.756", "0.63", "0.5146", "0.63" ], [ "0.1184", "0.633", "0.4878", "0.633", "0.4878", "0.6465", "0.1184", "0.6465" ], [ "0.513", "0.6343", "0.881", "0.6343", "0.881", "0.648", "0.513", "0.648" ], [ "0.1184", "0.6484", "0.4836", "0.6484", "0.4836", "0.662", "0.1184", "0.662" ], [ "0.5156", "0.652", "0.531", "0.652", "0.531", "0.66", "0.5156", "0.66" ], [ "0.523", "0.6484", "0.881", "0.651", "0.881", "0.6646", "0.523", "0.662" ], [ "0.1184", "0.665", "0.4895", "0.665", "0.4895", "0.6787", "0.1184", "0.6787" ], [ "0.511", "0.6665", "0.881", "0.6665", "0.881", "0.6797", "0.511", "0.6797" ], [ "0.1184", "0.6816", "0.4878", "0.6816", "0.4878", "0.6953", "0.1184", "0.6953" ], [ "0.511", "0.6816", "0.881", "0.6816", "0.881", "0.6953", "0.511", "0.6953" ], [ "0.1184", "0.6978", "0.4878", "0.6978", "0.4878", "0.7114", "0.1184", "0.7114" ], [ "0.51", "0.697", "0.881", "0.6978", "0.881", "0.712", "0.51", "0.7114" ], [ "0.1184", "0.712", "0.4878", "0.7134", "0.4878", "0.728", "0.1184", "0.727" ], [ "0.513", "0.7144", "0.8696", "0.7144", "0.8696", "0.728", "0.513", "0.728" ], [ "0.12006", "0.73", "0.4844", "0.73", "0.4844", "0.7407", "0.12006", "0.7407" ], [ "0.51", "0.73", "0.879", "0.73", "0.879", "0.7437", "0.51", "0.7437" ], [ "0.12006", "0.745", "0.486", "0.745", "0.486", "0.759", "0.12006", "0.759" ], [ "0.5146", "0.749", "0.879", "0.749", "0.879", "0.759", "0.5146", "0.759" ], [ "0.12006", "0.761", "0.4878", "0.761", "0.4878", "0.775", "0.12006", "0.775" ], [ "0.51", "0.761", "0.879", "0.761", "0.879", "0.775", "0.51", "0.775" ], [ "0.12006", "0.7793", "0.486", "0.7793", "0.486", "0.789", "0.12006", "0.789" ], [ "0.513", "0.7793", "0.881", "0.7793", "0.881", "0.792", "0.513", "0.792" ], [ "0.1184", "0.7935", "0.4878", "0.7935", "0.4878", "0.807", "0.1184", "0.807" ], [ "0.4895", "0.8", "0.4978", "0.7974", "0.4995", "0.8003", "0.4902", "0.8027" ], [ "0.511", "0.7944", "0.881", "0.7944", "0.881", "0.808", "0.511", "0.808" ], [ "0.12006", "0.811", "0.2275", "0.811", "0.2275", "0.8223", "0.12006", "0.8223" ], [ "0.513", "0.81", "0.879", "0.81", "0.879", "0.8237", "0.513", "0.8237" ], [ "0.1351", "0.8267", "0.482", "0.8267", "0.482", "0.8403", "0.1351", "0.8403" ], [ "0.5146", "0.829", "0.879", "0.829", "0.879", "0.839", "0.5146", "0.839" ], [ "0.1184", "0.8433", "0.4878", "0.8433", "0.4878", "0.857", "0.1184", "0.857" ], [ "0.513", "0.8433", "0.881", "0.8433", "0.881", "0.857", "0.513", "0.857" ], [ "0.1184", "0.8594", "0.4878", "0.8594", "0.4878", "0.8726", "0.1184", "0.8726" ], [ "0.513", "0.8594", "0.882", "0.8594", "0.882", "0.8726", "0.513", "0.8726" ], [ "0.12006", "0.877", "0.4844", "0.877", "0.4844", "0.887", "0.12006", "0.887" ], [ "0.5146", "0.877", "0.876", "0.877", "0.876", "0.887", "0.5146", "0.887" ], [ "0.1184", "0.891", "0.4844", "0.891", "0.4844", "0.905", "0.1184", "0.905" ], [ "0.511", "0.891", "0.879", "0.891", "0.879", "0.905", "0.511", "0.905" ], [ "0.1184", "0.9067", "0.4878", "0.9067", "0.4878", "0.9204", "0.1184", "0.9204" ], [ "0.513", "0.909", "0.5693", "0.909", "0.5693", "0.9204", "0.513", "0.9204" ] ]

[ "MLLM将LLM的高级语言处理与多模态感知", "评估，这将在E4.2节中讨论。在第一组中，我", "功能相结合(Zhang etal.,2023a;Yeet al.,2023;", "们随机选择了每个对象10到50个视频片段，", "Li et al., 2023a; Gao et al., 2023; Peng et al., 2023;", "从而产生了7442个来自EPIC-KITCHENS的", "Tangetal.，2023a）。这项研究的核心是将预训", "306个不同对象", "练的视觉编码器表示与LLM的输入嵌人空间", "融合在一起，这是通过使用交织图像和文本的", "我们利用最大的以自我为中心的视频数据集", "数据集预训练来实现的。（Lietal.,2023c；Zhu", "Ego4D，选择标记为“object_of_change”的", "etal.,2023a;Liuetal.,2023a)。在本文中，我", "视频片段来增强数据的多样性。此标签突出显", "们的目标是对这些模型进行全面的评估，特别", "示了展示对象状态更改的视频。通过收集这些", "是关注它们在对象状态变化识别中的性能", "特定的视频以及对象及其叙述的详细信息，我", "们为数据生成提供了信息，并编制了相关统计", "3The OSCaRDataset", "数据。从这个数据集中，我们为我们的OSCaR", "项目提取了5942个具有296个独特对象的段。", "本节概述了我们用于在对象状态上创建可视指", "令的管道。我们从从公共来源收集各种视觉数", "据的过程开始，详见3.1节。接下来，第3.2", "节介绍了我们在各种任务中使用简单的人工注", "3.2GPT-assisted Data Generation", "释来提高数据质量的方法，从而有助于更深入", "地了解对象状态。我们的方法可以生成详细的", "字幕，视觉问答和视觉对话。", "字幕生成：字幕在视觉理解中起着重要作用。", "了解对象状态需要详细且信息丰富的标题来捕", "3.1Video Collections", "获对象的确切状态。为厂实现这一自标，我们", "OSCaR是来自两个不同数据集的视频的精选", "利用GPT-4V和人类的弱注释为所有收集的视", "汇编: EPIC-KITCHENSE(Damen et al., 2018)", "频生成了字幕。此问题需要两种类型的注释，", "和Ego4DE（Graumanetal.,2021）。认识到对象", "包括1）事件期间视频中的开始和结束顿ID以", "状态的变化随着时间的推移而不是在单个帧内", "进行状态更改以及2）对视频中发生的情况的", "突然发生，我们有选择地包括有效地说明这些", "简短描述。简短描述可以是表示动作的动词和", "状态转换的视频剪辑。我们的选择过程可确保", "表示人类与之交互的对象的名词（e。g。，洗", "这些视频描绘对象状态的动态变化，并捕获对", "涤盘）。我们设计了自适应提示来注入此注释", "象在足够短的时间内保持静止的时刻。这种方", "作为上下文，以指导GPT-4V生成高质量的字", "法使我们能够编译一个全面的可视化数据集，", "幕。我们发现，GPT-4V在没有此指导的情况", "其中包含对象的静态和过渡状态", "下经常会出现歧义，并且生成的学幕的质量会", "我们最初分析了EPIC-KITCHENS数据集原", "降低。通过简单的人为指导，GPT-4V可以减", "始视频中的动词，以确保视频突出显示了经历", "少歧义并产生质量更好的字幕，", "状态更改的对象。我们将这些动词分为三组：", "多项选择QA生成：多项选择问题是一种呈现", "更改，不确定和不更改。更改组由可能更改对", "", "一组答案的方法，包括不正确的选项，以教导", "象状态的动词组成，包括“打开”、“关闭”、“清", "机器学习模型如何区分正确和不正确的答案。", "洗”、“剪切”和“混合”等操作。相反，不改变", "这种类型的问题也可以用作一种指令形式，其", "组包含引起状态改变的可能性最小的动词，例", "中问题作为提示，答案作为模型的响应。我们", "如Take，Put，Move，Check等。最后，不确定", "根据生成的字幕创建了多项选择题和答案集。", "的组包括具有模糊状态变化潜力的动词：涵盖", "对话生成：视觉对话是一项复杂的任务，需要", "诸如Shake，Flip，Use，Pull等动作。在过滤", "了解视觉内容和对话上下文，并且由于需要两", "了EPIC-KITCHENS数据集后，我们能够识别", "个人之间自然对话来查看相同的内容，因此在", "出总共由650个动词组成的69个动词类。使", "数据收集方面面临挑战。此过程耗时且资源密", "用这个动词列表，我们检索了包含这些动作的", "", "集，尤其是在涉及推理和解释时。随着机器学", "所有视频片段", "习模型的发展，生成视觉对话数据变得越来越", "在分析视频时，我们发现一些物体只出现了", "重要。我们开发了一种使用学幕创建视觉对话", "几次。因此，我们将视频分成两组。第一组包", "数据的方法，通过GPT-4V提供解释的能力得", "括聚焦于发生十次以上的对象的视频，它将", "到了增强，提供了灵活多样的数据。这种方法，", "用于构建我们的训练和测试集。第二组包括", "标记图像和视频的输入数据，具有成本效益", "发生次数少于十次的对象的视频。这些对象在", "比手动方法更快，为未来的模型生成大量的训", "EPIC-KITCHENS中很少见，可用于开放世界", "练数据" ]

{ "arxiv_id": "2402.17128", "language": "zh", "page_idx": 3 }

[ [ "0.11755", "0.0885", "0.3315", "0.0885", "0.3315", "0.0986", "0.11755", "0.0986" ], [ "0.771", "0.08905", "0.867", "0.08905", "0.867", "0.10095", "0.771", "0.10095" ], [ "0.11755", "0.11224", "0.486", "0.11224", "0.486", "0.1259", "0.11755", "0.1259" ], [ "0.1184", "0.133", "0.4895", "0.133", "0.4895", "0.1467", "0.1184", "0.1467" ], [ "0.11755", "0.1484", "0.486", "0.1478", "0.486", "0.1609", "0.11755", "0.1621" ], [ "0.12006", "0.1675", "0.4844", "0.1675", "0.4844", "0.1776", "0.12006", "0.1776" ], [ "0.1184", "0.1818", "0.4878", "0.1818", "0.4878", "0.1953", "0.1184", "0.1953" ], [ "0.1184", "0.1996", "0.482", "0.1996", "0.482", "0.2102", "0.1184", "0.2102" ], [ "0.11755", "0.2144", "0.4836", "0.2144", "0.4836", "0.228", "0.11755", "0.228" ], [ "0.11584", "0.2286", "0.486", "0.2299", "0.486", "0.2446", "0.11584", "0.2434" ], [ "0.1184", "0.2446", "0.1587", "0.2477", "0.1561", "0.2625", "0.11755", "0.26" ], [ "0.659", "0.2488", "0.684", "0.2488", "0.684", "0.253", "0.659", "0.253" ], [ "0.7363", "0.2488", "0.76", "0.2488", "0.76", "0.253", "0.7363", "0.253" ], [ "0.511", "0.2654", "0.879", "0.2666", "0.879", "0.2798", "0.511", "0.279" ], [ "0.11584", "0.2725", "0.4844", "0.2737", "0.4844", "0.288", "0.11584", "0.2874" ], [ "0.511", "0.2798", "0.881", "0.2798", "0.881", "0.2935", "0.511", "0.2935" ], [ "0.11755", "0.2935", "0.4844", "0.2935", "0.4844", "0.3071", "0.11755", "0.3071" ], [ "0.513", "0.2964", "0.879", "0.2964", "0.879", "0.3071", "0.513", "0.3071" ], [ "0.1184", "0.3113", "0.486", "0.3113", "0.486", "0.3247", "0.1184", "0.3247" ], [ "0.511", "0.3088", "0.881", "0.3088", "0.881", "0.3225", "0.511", "0.3225" ], [ "0.11755", "0.326", "0.4878", "0.3267", "0.4878", "0.3403", "0.11755", "0.339" ], [ "0.511", "0.3225", "0.881", "0.3225", "0.881", "0.3362", "0.511", "0.3362" ], [ "0.513", "0.3362", "0.879", "0.3374", "0.879", "0.3508", "0.513", "0.3499" ], [ "0.12006", "0.3457", "0.4836", "0.3457", "0.4836", "0.3557", "0.12006", "0.3557" ], [ "0.513", "0.3516", "0.882", "0.3516", "0.882", "0.3652", "0.513", "0.3652" ], [ "0.1184", "0.3599", "0.486", "0.3599", "0.486", "0.3735", "0.1184", "0.3735" ], [ "0.511", "0.364", "0.879", "0.364", "0.879", "0.3777", "0.511", "0.3777" ], [ "0.1184", "0.376", "0.4878", "0.376", "0.4878", "0.3896", "0.1184", "0.3896" ], [ "0.511", "0.379", "0.882", "0.3801", "0.882", "0.3938", "0.511", "0.3926" ], [ "0.1184", "0.3901", "0.486", "0.3901", "0.486", "0.4038", "0.1184", "0.4038" ], [ "0.513", "0.3943", "0.7095", "0.3955", "0.7095", "0.4067", "0.513", "0.4062" ], [ "0.12006", "0.408", "0.4844", "0.408", "0.4844", "0.4216", "0.12006", "0.4216" ], [ "0.1184", "0.4233", "0.486", "0.4233", "0.486", "0.437", "0.1184", "0.437" ], [ "0.1184", "0.4395", "0.486", "0.4395", "0.486", "0.4531", "0.1184", "0.4531" ], [ "0.513", "0.434", "0.881", "0.4353", "0.881", "0.449", "0.513", "0.4478" ], [ "0.1184", "0.456", "0.4878", "0.456", "0.4878", "0.4697", "0.1184", "0.4697" ], [ "0.513", "0.452", "0.881", "0.452", "0.881", "0.4656", "0.513", "0.4656" ], [ "0.1184", "0.4714", "0.4844", "0.4714", "0.4844", "0.485", "0.1184", "0.485" ], [ "0.511", "0.4673", "0.881", "0.4673", "0.881", "0.481", "0.511", "0.481" ], [ "0.1184", "0.488", "0.486", "0.488", "0.486", "0.502", "0.1184", "0.502" ], [ "0.513", "0.4827", "0.881", "0.4827", "0.881", "0.4963", "0.513", "0.4963" ], [ "0.12006", "0.505", "0.2091", "0.505", "0.2091", "0.5166", "0.12006", "0.5166" ], [ "0.513", "0.4995", "0.881", "0.4995", "0.881", "0.513", "0.513", "0.513" ], [ "0.1184", "0.521", "0.4878", "0.521", "0.4878", "0.5337", "0.1184", "0.5337" ], [ "0.513", "0.5156", "0.881", "0.5156", "0.881", "0.5293", "0.513", "0.5293" ], [ "0.11755", "0.536", "0.486", "0.535", "0.486", "0.549", "0.11755", "0.55" ], [ "0.5156", "0.533", "0.879", "0.533", "0.879", "0.5435", "0.5156", "0.5435" ], [ "0.1184", "0.552", "0.4844", "0.552", "0.4844", "0.5654", "0.1184", "0.5654" ], [ "0.513", "0.5474", "0.881", "0.5474", "0.881", "0.561", "0.513", "0.561" ], [ "0.1184", "0.5684", "0.4878", "0.5684", "0.4878", "0.582", "0.1184", "0.582" ], [ "0.513", "0.5635", "0.881", "0.5635", "0.881", "0.5767", "0.513", "0.5767" ], [ "0.12006", "0.586", "0.486", "0.586", "0.486", "0.5967", "0.12006", "0.5967" ], [ "0.5146", "0.5806", "0.5693", "0.5806", "0.5693", "0.593", "0.5146", "0.593" ], [ "0.1184", "0.6016", "0.486", "0.6016", "0.486", "0.615", "0.1184", "0.615" ], [ "0.513", "0.608", "0.6733", "0.608", "0.6733", "0.6216", "0.513", "0.6216" ], [ "0.11755", "0.616", "0.482", "0.616", "0.482", "0.63", "0.11755", "0.63" ], [ "0.1184", "0.633", "0.297", "0.633", "0.297", "0.6465", "0.1184", "0.6465" ], [ "0.5146", "0.6343", "0.881", "0.6343", "0.881", "0.648", "0.5146", "0.648" ], [ "0.1184", "0.6494", "0.486", "0.6494", "0.486", "0.663", "0.1184", "0.663" ], [ "0.511", "0.6484", "0.8774", "0.6484", "0.8774", "0.662", "0.511", "0.662" ], [ "0.11755", "0.6646", "0.486", "0.6646", "0.486", "0.6777", "0.11755", "0.6777" ], [ "0.511", "0.6665", "0.8774", "0.6665", "0.8774", "0.677", "0.511", "0.677" ], [ "0.12177", "0.683", "0.486", "0.683", "0.486", "0.6934", "0.12177", "0.6934" ], [ "0.511", "0.6816", "0.879", "0.6816", "0.879", "0.6953", "0.511", "0.6953" ], [ "0.12006", "0.6987", "0.486", "0.6987", "0.486", "0.712", "0.12006", "0.712" ], [ "0.513", "0.6997", "0.8774", "0.6997", "0.8774", "0.7104", "0.513", "0.7104" ], [ "0.1184", "0.7144", "0.4878", "0.7144", "0.4878", "0.728", "0.1184", "0.728" ], [ "0.51", "0.7134", "0.881", "0.7134", "0.881", "0.727", "0.51", "0.727" ], [ "0.1184", "0.73", "0.4895", "0.73", "0.4895", "0.7437", "0.1184", "0.7437" ], [ "0.513", "0.73", "0.881", "0.73", "0.881", "0.7437", "0.513", "0.7437" ], [ "0.1184", "0.745", "0.4878", "0.745", "0.4878", "0.759", "0.1184", "0.759" ], [ "0.513", "0.745", "0.881", "0.745", "0.881", "0.759", "0.513", "0.759" ], [ "0.12006", "0.7627", "0.4836", "0.7627", "0.4836", "0.7725", "0.12006", "0.7725" ], [ "0.513", "0.7627", "0.881", "0.7627", "0.881", "0.7754", "0.513", "0.7754" ], [ "0.1184", "0.778", "0.4844", "0.778", "0.4844", "0.7915", "0.1184", "0.7915" ], [ "0.513", "0.7793", "0.881", "0.7793", "0.881", "0.792", "0.513", "0.792" ], [ "0.1184", "0.7944", "0.4895", "0.7944", "0.4895", "0.808", "0.1184", "0.808" ], [ "0.511", "0.7935", "0.879", "0.7935", "0.879", "0.807", "0.511", "0.807" ], [ "0.1184", "0.81", "0.4895", "0.81", "0.4895", "0.8237", "0.1184", "0.8237" ], [ "0.513", "0.811", "0.881", "0.811", "0.881", "0.8247", "0.513", "0.8247" ], [ "0.11755", "0.8247", "0.2805", "0.826", "0.2805", "0.8403", "0.11755", "0.839" ], [ "0.5146", "0.8267", "0.881", "0.8267", "0.881", "0.8403", "0.5146", "0.8403" ], [ "0.513", "0.8433", "0.881", "0.8433", "0.881", "0.857", "0.513", "0.857" ], [ "0.11584", "0.8525", "0.361", "0.854", "0.361", "0.868", "0.11584", "0.8677" ], [ "0.5146", "0.8594", "0.79", "0.8594", "0.79", "0.8726", "0.5146", "0.8726" ], [ "0.1184", "0.8745", "0.486", "0.8745", "0.486", "0.888", "0.1184", "0.888" ], [ "0.531", "0.8745", "0.881", "0.8745", "0.881", "0.888", "0.531", "0.888" ], [ "0.1184", "0.8936", "0.4878", "0.8936", "0.4878", "0.904", "0.1184", "0.904" ], [ "0.513", "0.891", "0.881", "0.891", "0.881", "0.905", "0.513", "0.905" ], [ "0.11755", "0.9067", "0.4963", "0.9077", "0.4963", "0.9214", "0.11755", "0.9204" ], [ "0.492", "0.912", "0.5054", "0.912", "0.5054", "0.9165", "0.492", "0.9165" ], [ "0.513", "0.909", "0.875", "0.909", "0.875", "0.9194", "0.513", "0.9194" ] ]

[ "4OsCaRBenchmarks", "ShontAnswes(10 wrds)", "4.1 Evaluation with Text Generation Metrics", "我们提供的数据集由来自Ego4D和EPIC-", "KITCHENS数据集的500个视频组成，这些", "视频是专门为基准测试目的而设计的。每个视", "频都有四个详细的字幕注释，所有这些字幕都", "经过严格的人工验证，以确保此评估集的质量", "和可靠性。为确保全面准确地评估性能，可将", "BLEU、Rouge、LSA等文本生成指标用于评估", "目的", "ength", "twor", "Figure 3:Distribution of answer lengths.The figure", "4.2Open-world Object State Understanding", "shows how answers are distributed by length in the", "收集全球所有对象的数据，然后训练模型是", "dataset.It separates short answers (1-9 words)from", "不可行的。然而，人类可以描述新的或不熟悉", "long answers ( ≥ 10 words). The histogram displays", "的物体，这对人工智能来说可能是具有挑战性", "the number of answers on the y-axis based on increas-", "ing answer lengths on the x-axis. There is a category", "的，特别是当它们处于新的领域或服务于不同", "at 100 words for answers with lengths greater than or", "的目的时。幸运的是，最近在MLLM方面取", "equal to 100 words. This breakdown emphasizes the", "得的成就为人工智能开辟了拥有这种能力的潜", "balance between brief, direct answers and more de-", "力。在使用大量数据进行预训练期间，MLLM", "tailed,explanatoryresponses", "可以学习有关世界的一般知识。此外，模型将", "学习如何在视觉指令调整过程中执行任务。在", "这两个过程中，模型可能已经或可能没有暴露", "重点是特定于框架的描述准确性，两个用于评", "于不在对象状态理解训练集中的对象。问题是", "估状态变化的准确性，两个用于识别幻觉，三", "模型是否可以推广到这种类型的对象。为了回", "个用于识别不完整的描述。注释者被要求将每", "答这个问题，我们提供了两个评估集来测试模", "个描述归类为以下四个标签之一：1）完全详", "型的通用性", "细和全面，2）一般完整但有轻微遗漏，3）缺", "烹饪域对象未出现在训练集中以进行对象状态", "少重要细节或包含错误，或4）不完整，误导", "理解：对于此评估，我们希望调查模型在与训", "或幻觉，并提供理由阻正随机反应。这项研究", "练域类似的场景中理解未出现在训练集中的对", "使用了来自EPIC-KITCHENS和Ego4D数据", "象的能力。我们提供了一组2,485个视频，其", "集的500个样本，从而验证了2000种自然语", "中1024个对象未出现在对象状态训练集中。", "言描述", "此测试集将评估域内知识如何帮助模型理解对", "5Data Statistics", "象状态和状态更改。我们使用GPT-4V注释了", "344个视频以进行评估", "为了帮助模型生成简洁而翔实的答案，我们将", "域外对象状态理解：此评估着重于判断模型理", "简短的答案定义为少于十个单词的答案，将长", "解训练域以外的对象的能力。我们的训练集只", "的答案定义为多于十个单词的答案。简短的答", "包含烹饪领域的数据，而这个测试集有不同的", "案提供简洁，而允长的答案提供详细和翔实的", "领域，如面包师，家庭管理，清洁/洗衣，自行", "信息。这两类答案的分布情况可以在图3中看", "车机械师等。此集合是从Ego4D数据集中提", "到。数据集中的平均答案长度为47。06words。", "取的，包含43,367个视频，其中包含500多个", "长答案约占数据的75个%，平均长度为63个", "对象。此测试不仅可用于评估，而且有可能使", "单词，而短答案约占数据的25个%，平均长度", "用我们的管道扩展其他特定领域的对象状态理", "为3。32个字。通过相应地拆分数据，未来的", "解。对于此评估集，我们从51个不同域中的", "模型可以提供简短、直接和信息丰富的答案和", "每个域中选择了10个视频，总计356个视频。", "解释。为了展示我们的OSCaR数据集的独特", "少于10个视频的域名包含其所有视频。此设", "性，我们在表1中展示了OSCaR和其他相关", "置也由GPT-4V注释", "数据集之间的比较。OSCaR数据集包含大量", "指令以及图像和视频。此外，它还为对象状态", "4.3Data Quality Verification", "字幕和对象状态更改字幕提供数据", "我们使用AmazonMTurk评估了跨视频帧的对", "在E4.2节中，我们讨论了两种用于对象状", "象状态和活动的描述质量，以获得人上反馈。", "态理解的开放世界数据集：域内烹饪和开放域。", "我们的评估框架包括五个发现不准确的指南，", "", "尽管我们在具有对象状态变化的视频上进行了" ]

{ "arxiv_id": "2402.17128", "language": "zh", "page_idx": 4 }

[ [ "0.1276", "0.0956", "0.8696", "0.0956", "0.8696", "0.10986", "0.1276", "0.10986" ], [ "0.3047", "0.1372", "0.4768", "0.1372", "0.4768", "0.1509", "0.3047", "0.1509" ], [ "0.5283", "0.1354", "0.7", "0.133", "0.7", "0.152", "0.5283", "0.1538" ], [ "0.2778", "0.152", "0.7246", "0.1526", "0.7246", "0.1687", "0.2778", "0.1675" ], [ "0.3687", "0.1704", "0.635", "0.1704", "0.635", "0.1841", "0.3687", "0.1841" ], [ "0.1873", "0.1853", "0.8154", "0.187", "0.8154", "0.2019", "0.1873", "0.1996" ], [ "0.2494", "0.2031", "0.763", "0.2037", "0.763", "0.2185", "0.2494", "0.2173" ], [ "0.2627", "0.2625", "0.3418", "0.2625", "0.3418", "0.2766", "0.2627", "0.2766" ], [ "0.7446", "0.2654", "0.784", "0.2654", "0.784", "0.2778", "0.7446", "0.2778" ], [ "0.552", "0.2778", "0.6978", "0.2778", "0.6978", "0.2922", "0.552", "0.2922" ], [ "0.1453", "0.2903", "0.4575", "0.2903", "0.4575", "0.304", "0.1453", "0.304" ], [ "0.1453", "0.3047", "0.4592", "0.3047", "0.4592", "0.3184", "0.1453", "0.3184" ], [ "0.7515", "0.3018", "0.7866", "0.3018", "0.7866", "0.3135", "0.7515", "0.3135" ], [ "0.553", "0.3164", "0.613", "0.3164", "0.613", "0.3206", "0.553", "0.3206" ], [ "0.663", "0.3164", "0.714", "0.3164", "0.714", "0.3237", "0.663", "0.3237" ], [ "0.553", "0.326", "0.5884", "0.326", "0.5884", "0.3289", "0.553", "0.3289" ], [ "0.1453", "0.3184", "0.4592", "0.3184", "0.4592", "0.332", "0.1453", "0.332" ], [ "0.5938", "0.326", "0.6187", "0.326", "0.6187", "0.3289", "0.5938", "0.3289" ], [ "0.1453", "0.3333", "0.4592", "0.3333", "0.4592", "0.3467", "0.1453", "0.3467" ], [ "0.1436", "0.3457", "0.4592", "0.3467", "0.4592", "0.361", "0.1436", "0.3599" ], [ "0.635", "0.3416", "0.7715", "0.3416", "0.7715", "0.3557", "0.635", "0.3557" ], [ "0.1436", "0.361", "0.4575", "0.3599", "0.4575", "0.3735", "0.1436", "0.3748" ], [ "0.546", "0.3623", "0.708", "0.3623", "0.708", "0.376", "0.546", "0.376" ], [ "0.1453", "0.3748", "0.4568", "0.3748", "0.4568", "0.3884", "0.1453", "0.3884" ], [ "0.614", "0.376", "0.777", "0.3818", "0.775", "0.4038", "0.613", "0.398" ], [ "0.1453", "0.3914", "0.4568", "0.3914", "0.4568", "0.402", "0.1453", "0.402" ], [ "0.1486", "0.4062", "0.455", "0.4062", "0.455", "0.4163", "0.1486", "0.4163" ], [ "0.5483", "0.4038", "0.567", "0.4038", "0.567", "0.412", "0.5483", "0.412" ], [ "0.702", "0.4092", "0.7773", "0.4092", "0.7773", "0.4204", "0.702", "0.4204" ], [ "0.1453", "0.4175", "0.4575", "0.4175", "0.4575", "0.4312", "0.1453", "0.4312" ], [ "0.1453", "0.4316", "0.4568", "0.4316", "0.4568", "0.4453", "0.1453", "0.4453" ], [ "0.513", "0.4395", "0.881", "0.4395", "0.881", "0.4531", "0.513", "0.4531" ], [ "0.1486", "0.4478", "0.4568", "0.4478", "0.4568", "0.4578", "0.1486", "0.4578" ], [ "0.511", "0.4531", "0.881", "0.4531", "0.881", "0.466", "0.511", "0.466" ], [ "0.1453", "0.4602", "0.4575", "0.4602", "0.4575", "0.4739", "0.1453", "0.4739" ], [ "0.513", "0.4685", "0.8774", "0.4685", "0.8774", "0.4785", "0.513", "0.4785" ], [ "0.147", "0.4744", "0.4575", "0.4744", "0.4575", "0.488", "0.147", "0.488" ], [ "0.511", "0.481", "0.881", "0.481", "0.881", "0.4954", "0.511", "0.4954" ], [ "0.1453", "0.4893", "0.4575", "0.4893", "0.4575", "0.503", "0.1453", "0.503" ], [ "0.5146", "0.4963", "0.621", "0.4963", "0.621", "0.5083", "0.5146", "0.5083" ], [ "0.1453", "0.503", "0.4575", "0.503", "0.4575", "0.5166", "0.1453", "0.5166" ], [ "0.1453", "0.517", "0.4575", "0.517", "0.4575", "0.531", "0.1453", "0.531" ], [ "0.1486", "0.533", "0.4568", "0.533", "0.4568", "0.5435", "0.1486", "0.5435" ], [ "0.51", "0.535", "0.882", "0.536", "0.882", "0.55", "0.51", "0.549" ], [ "0.147", "0.546", "0.4568", "0.546", "0.4568", "0.5586", "0.147", "0.5586" ], [ "0.513", "0.5527", "0.881", "0.5527", "0.881", "0.5664", "0.513", "0.5664" ], [ "0.1453", "0.56", "0.4575", "0.56", "0.4575", "0.5737", "0.1453", "0.5737" ], [ "0.1453", "0.5723", "0.4055", "0.5737", "0.4055", "0.5874", "0.1453", "0.586" ], [ "0.513", "0.5684", "0.8716", "0.5684", "0.8716", "0.582", "0.513", "0.582" ], [ "0.534", "0.585", "0.881", "0.585", "0.881", "0.5986", "0.534", "0.5986" ], [ "0.11584", "0.6025", "0.1453", "0.6025", "0.1453", "0.6143", "0.11584", "0.6143" ], [ "0.1511", "0.6025", "0.2612", "0.6025", "0.2612", "0.6143", "0.1511", "0.6143" ], [ "0.511", "0.5996", "0.879", "0.5996", "0.879", "0.6133", "0.511", "0.6133" ], [ "0.1436", "0.6055", "0.1486", "0.6055", "0.1486", "0.611", "0.1436", "0.611" ], [ "0.5146", "0.6177", "0.879", "0.6177", "0.879", "0.6313", "0.5146", "0.6313" ], [ "0.11755", "0.626", "0.4836", "0.626", "0.4836", "0.6396", "0.11755", "0.6396" ], [ "0.513", "0.633", "0.881", "0.633", "0.881", "0.6465", "0.513", "0.6465" ], [ "0.1184", "0.6436", "0.4844", "0.6436", "0.4844", "0.6567", "0.1184", "0.6567" ], [ "0.5146", "0.6494", "0.881", "0.6494", "0.881", "0.663", "0.5146", "0.663" ], [ "0.12006", "0.66", "0.486", "0.66", "0.486", "0.6733", "0.12006", "0.6733" ], [ "0.511", "0.6646", "0.881", "0.6646", "0.881", "0.6777", "0.511", "0.6777" ], [ "0.11755", "0.675", "0.4844", "0.675", "0.4844", "0.6885", "0.11755", "0.6885" ], [ "0.513", "0.6816", "0.66", "0.6816", "0.66", "0.6953", "0.513", "0.6953" ], [ "0.12006", "0.6934", "0.4844", "0.6934", "0.4844", "0.7036", "0.12006", "0.7036" ], [ "0.5283", "0.6953", "0.882", "0.697", "0.882", "0.7114", "0.5283", "0.7104" ], [ "0.11584", "0.707", "0.4836", "0.707", "0.4836", "0.72", "0.11584", "0.72" ], [ "0.513", "0.7134", "0.881", "0.7134", "0.881", "0.728", "0.513", "0.728" ], [ "0.1184", "0.7246", "0.4668", "0.7246", "0.4668", "0.7383", "0.1184", "0.7383" ], [ "0.513", "0.7285", "0.881", "0.7285", "0.881", "0.7437", "0.513", "0.7437" ], [ "0.1368", "0.7407", "0.4836", "0.7407", "0.4836", "0.7544", "0.1368", "0.7544" ], [ "0.513", "0.745", "0.881", "0.745", "0.881", "0.759", "0.513", "0.759" ], [ "0.1184", "0.756", "0.4844", "0.756", "0.4844", "0.7695", "0.1184", "0.7695" ], [ "0.5156", "0.764", "0.534", "0.764", "0.534", "0.7725", "0.5156", "0.7725" ], [ "0.5283", "0.761", "0.881", "0.761", "0.881", "0.7754", "0.5283", "0.7754" ], [ "0.1184", "0.7715", "0.486", "0.7715", "0.486", "0.785", "0.1184", "0.785" ], [ "0.511", "0.777", "0.879", "0.777", "0.879", "0.7905", "0.511", "0.7905" ], [ "0.11755", "0.7876", "0.486", "0.788", "0.486", "0.8027", "0.11755", "0.802" ], [ "0.5146", "0.7944", "0.679", "0.7944", "0.679", "0.808", "0.5146", "0.808" ], [ "0.1184", "0.8047", "0.486", "0.8047", "0.486", "0.8184", "0.1184", "0.8184" ], [ "0.531", "0.81", "0.881", "0.81", "0.881", "0.8237", "0.531", "0.8237" ], [ "0.1184", "0.821", "0.4844", "0.821", "0.4844", "0.8345", "0.1184", "0.8345" ], [ "0.51", "0.8247", "0.881", "0.826", "0.881", "0.8403", "0.51", "0.839" ], [ "0.1184", "0.8374", "0.4844", "0.8374", "0.4844", "0.851", "0.1184", "0.851" ], [ "0.5146", "0.843", "0.879", "0.843", "0.879", "0.8555", "0.5146", "0.8555" ], [ "0.1411", "0.86", "0.4844", "0.86", "0.4844", "0.8735", "0.1411", "0.8735" ], [ "0.5146", "0.8594", "0.881", "0.8594", "0.881", "0.8726", "0.5146", "0.8726" ], [ "0.11584", "0.8706", "0.4878", "0.8716", "0.4878", "0.8857", "0.11584", "0.8853" ], [ "0.513", "0.8745", "0.879", "0.8745", "0.879", "0.888", "0.513", "0.888" ], [ "0.11584", "0.8843", "0.486", "0.8843", "0.486", "0.898", "0.11584", "0.898" ], [ "0.11755", "0.8965", "0.486", "0.8965", "0.486", "0.91", "0.11755", "0.91" ], [ "0.5146", "0.891", "0.881", "0.891", "0.881", "0.905", "0.5146", "0.905" ], [ "0.11584", "0.9077", "0.361", "0.9077", "0.361", "0.9214", "0.11584", "0.9214" ], [ "0.5156", "0.9067", "0.881", "0.9067", "0.881", "0.9204", "0.5156", "0.9204" ] ]

[ "采取行动实现优雅的交互：对指令澄清请求的建模策略执行操作的影响", "Brielen Madureiral", "David Schlangen1,2", "1Computational Linguistics, Department of Linguistics", "University of Potsdam, Germany", "2GermanResearchCenterfor Artificial Intelligence (DFKI),Berlin,Germany", "{madureiralasota,david.schlangen } @uni-potsdam.de", "Abstract", "ready", "thbigleoolftic", "Clarification requests are a mechanism to help", "solve communication problems, e.g.due to", "okay", "thatroe.ic.intha", "and theboy", "noa", "ambiguity or underspecification, in instruction-", "whr", "following interactions. Despite their impor-", "tance, even skilful models struggle with pro-", "middoofnhtrikabutoa", "ducing or interpreting such repair acts. In this", "hsitinhm", "work, we test three hypotheses concerning the", "oyotey", "effects of actiontaking as an auxiliarytask in", "modellingiCRpolicies.Contrarytoinitialex", "yes", "okay, what elsee", "pectations, we conclude that its contribution", "to learning an iCR policy is limited, but some", "Figure l: Clarification requests posed by an instruction", "informationcanstillbeextractedfrompredic", "follower, demonstrating uncertainty on deciding what", "tion uncertainty. We present further evidence", "actionstotakeduetoambiguityorunderspecification", "that even well-motivated, Transformer-based", "From: CoDraw dialogue game 8198, CC BY-NC 4.0,", "models fail to learn good policies for when", "cliparts from ?.", "to ask Instruction CRs (iCRs), while the task", "of determining what to ask about can be more", "successfullymodelled.Consideringtheimpli", "4个%轮转到严格遵循指令的交互中的11个%", "cations of these findings, we further discuss", "轮转（??？）。因此，目前还不清楚使用数据驱", "the shortcomings of the data-driven paradigm", "for learning meta-communication acts.", "动的方法（?）可以在多大程度上学习CR策略", "许多现有的CR数据集，尽管它们对会话搜", "1", "Introduction", "索（？？）等应用程序很有用，但它们不是通过", "", "真实的交互收集的，或者是合成的，因此学习", "优雅互动（?？）的概念是作为一组技能提出的", "的CR策略可能不符合真正的人类行为。而且，", "机器应该展示这些技能，以正确地与人类进行", "当前性能最好的数据驱动模型在决定何时请求", "合作对话，其中包括能够要求，理解和提供澄", "澄清方面仍然做得不是很好（请参阅$2），我", "清。四十多年后，大型语言模型和语音助手无", "们必须了解为什么", "法处理规格不足以及正确处理或产生澄清请求", "CR可以在所有四个级别的通信（？）中发生：", "（CR）的能力仍在记录中（?????）。这也是目前", "注意（由于通道中的问题），识别（由于声学", "流行的商业聊天优化LLM的公认限制之一。", "障碍），识别（当理解信号但出现词汇，解析", "鉴于它们是为指令而调制的，这似乎是一个", "或参考问题时）和考虑（当意图不清楚时）（？)", "特殊的错误：CR是一种至关重要的机制，用", "", "指令CR（iCR）主要出现在Clark的第4级", "于修复交互（?）之后的指令中的误解，如图1", "通信（?），i。e。在摄取（?)的水平上，解决歧", "所示。转念一想，这并不奇怪。澄清交换是元", "义和规格不足的问题", "通信行为，通常不会出现在非交互式数据（?)", "最近，？？认为多模式CoDraw游戏（?）是", "中，在对话数据中也相对罕见。作为一种特定", "iCR的丰富资源，自然是通过动作玩游戏的副", "的对话现象，CR的经验频率为自发对话中的", "产品，如图1中的示例所示。该数据集提供了", "'In the blogpost releasing chatGPT, the limitations sec", "大小（与精心策划但较小的语料库相比）和保", "tion says: “\" Ideally, the modei would ask clarifying questions", "留生态有效性（与人工收集或制作的大量数据", "when the user provided an ambiguous query. Instead, our cur", "rent models usually guess what the user intended. \". Source:", "集相反）之间的平衡。。假设潜在的iCR策略", "https://openai.com/blog/chatgpt.", "可以从数据中出现，我们可以合理地假设采取" ]

{ "arxiv_id": "2401.17039", "language": "zh", "page_idx": 1 }

[ [ "0.1184", "0.0873", "0.4878", "0.0873", "0.4878", "0.10095", "0.1184", "0.10095" ], [ "0.5073", "0.0849", "0.879", "0.0873", "0.879", "0.10095", "0.5073", "0.0986" ], [ "0.1184", "0.1027", "0.4106", "0.1027", "0.4106", "0.1164", "0.1184", "0.1164" ], [ "0.513", "0.1027", "0.7153", "0.1027", "0.7153", "0.1164", "0.513", "0.1164" ], [ "0.1394", "0.1217", "0.4836", "0.1217", "0.4836", "0.1318", "0.1394", "0.1318" ], [ "0.531", "0.1194", "0.881", "0.1194", "0.881", "0.133", "0.531", "0.133" ], [ "0.1184", "0.136", "0.4878", "0.136", "0.4878", "0.1497", "0.1184", "0.1497" ], [ "0.51", "0.1342", "0.8774", "0.1342", "0.8774", "0.1478", "0.51", "0.1478" ], [ "0.12006", "0.152", "0.4844", "0.152", "0.4844", "0.165", "0.12006", "0.165" ], [ "0.513", "0.1509", "0.881", "0.1509", "0.881", "0.1644", "0.513", "0.1644" ], [ "0.1184", "0.1687", "0.4844", "0.1687", "0.4844", "0.1818", "0.1184", "0.1818" ], [ "0.511", "0.1663", "0.879", "0.1687", "0.879", "0.1818", "0.511", "0.1799" ], [ "0.1184", "0.1853", "0.4844", "0.1853", "0.4844", "0.199", "0.1184", "0.199" ], [ "0.513", "0.1841", "0.881", "0.1841", "0.881", "0.1978", "0.513", "0.1978" ], [ "0.1184", "0.2007", "0.4878", "0.2007", "0.4878", "0.2144", "0.1184", "0.2144" ], [ "0.511", "0.1996", "0.879", "0.1996", "0.879", "0.2131", "0.511", "0.2131" ], [ "0.11755", "0.2173", "0.486", "0.2173", "0.486", "0.231", "0.11755", "0.231" ], [ "0.511", "0.2144", "0.881", "0.2156", "0.881", "0.2299", "0.511", "0.2286" ], [ "0.12006", "0.2328", "0.486", "0.2328", "0.486", "0.2465", "0.12006", "0.2465" ], [ "0.513", "0.2322", "0.881", "0.2322", "0.881", "0.2452", "0.513", "0.2452" ], [ "0.11584", "0.2477", "0.3594", "0.2494", "0.3594", "0.2632", "0.11584", "0.2612" ], [ "0.513", "0.2477", "0.881", "0.2477", "0.881", "0.2612", "0.513", "0.2612" ], [ "0.513", "0.2642", "0.5415", "0.2642", "0.5415", "0.2766", "0.513", "0.2766" ], [ "0.12427", "0.2778", "0.4844", "0.2778", "0.4844", "0.2915", "0.12427", "0.2915" ], [ "0.5283", "0.279", "0.882", "0.2798", "0.882", "0.2944", "0.5283", "0.2935" ], [ "0.1184", "0.2944", "0.4878", "0.2944", "0.4878", "0.308", "0.1184", "0.308" ], [ "0.511", "0.2964", "0.881", "0.2964", "0.881", "0.31", "0.511", "0.31" ], [ "0.1184", "0.31", "0.4844", "0.31", "0.4844", "0.3237", "0.1184", "0.3237" ], [ "0.51", "0.3113", "0.879", "0.3113", "0.879", "0.3247", "0.51", "0.3247" ], [ "0.12006", "0.328", "0.486", "0.328", "0.486", "0.3384", "0.12006", "0.3384" ], [ "0.5146", "0.3289", "0.8774", "0.3289", "0.8774", "0.339", "0.5146", "0.339" ], [ "0.1184", "0.3425", "0.486", "0.3425", "0.486", "0.3557", "0.1184", "0.3557" ], [ "0.511", "0.3445", "0.879", "0.3445", "0.879", "0.3582", "0.511", "0.3582" ], [ "0.1184", "0.3582", "0.486", "0.3582", "0.486", "0.3718", "0.1184", "0.3718" ], [ "0.513", "0.361", "0.881", "0.361", "0.881", "0.3748", "0.513", "0.3748" ], [ "0.1184", "0.3748", "0.486", "0.3748", "0.486", "0.3884", "0.1184", "0.3884" ], [ "0.513", "0.3772", "0.881", "0.3772", "0.881", "0.3901", "0.513", "0.3901" ], [ "0.1184", "0.3914", "0.4878", "0.3914", "0.4878", "0.405", "0.1184", "0.405" ], [ "0.513", "0.3926", "0.881", "0.3926", "0.881", "0.4062", "0.513", "0.4062" ], [ "0.1184", "0.4067", "0.3015", "0.4067", "0.3015", "0.4204", "0.1184", "0.4204" ], [ "0.513", "0.4092", "0.879", "0.4092", "0.879", "0.4229", "0.513", "0.4229" ], [ "0.5146", "0.4246", "0.881", "0.4246", "0.881", "0.4382", "0.5146", "0.4382" ], [ "0.11755", "0.4395", "0.1587", "0.4395", "0.1587", "0.4531", "0.11755", "0.4531" ], [ "0.1503", "0.4382", "0.2737", "0.4382", "0.2737", "0.4531", "0.1503", "0.4531" ], [ "0.513", "0.4407", "0.7305", "0.4412", "0.7305", "0.456", "0.513", "0.4548" ], [ "0.12006", "0.4673", "0.486", "0.4673", "0.486", "0.481", "0.12006", "0.481" ], [ "0.5156", "0.4673", "0.881", "0.4673", "0.881", "0.481", "0.5156", "0.481" ], [ "0.11755", "0.4822", "0.486", "0.484", "0.486", "0.4976", "0.11755", "0.4954" ], [ "0.531", "0.4827", "0.879", "0.4827", "0.879", "0.4963", "0.531", "0.4963" ], [ "0.12006", "0.4995", "0.4878", "0.4995", "0.4878", "0.513", "0.12006", "0.513" ], [ "0.51", "0.4976", "0.8774", "0.4988", "0.8774", "0.5127", "0.51", "0.511" ], [ "0.12177", "0.517", "0.486", "0.517", "0.486", "0.528", "0.12177", "0.528" ], [ "0.511", "0.5156", "0.881", "0.5156", "0.881", "0.5293", "0.511", "0.5293" ], [ "0.1184", "0.532", "0.492", "0.532", "0.492", "0.546", "0.1184", "0.546" ], [ "0.511", "0.531", "0.881", "0.531", "0.881", "0.5444", "0.511", "0.5444" ], [ "0.1184", "0.5474", "0.4878", "0.5474", "0.4878", "0.561", "0.1184", "0.561" ], [ "0.513", "0.5474", "0.881", "0.5474", "0.881", "0.561", "0.513", "0.561" ], [ "0.1184", "0.564", "0.4895", "0.564", "0.4895", "0.5776", "0.1184", "0.5776" ], [ "0.513", "0.5635", "0.881", "0.5635", "0.881", "0.5767", "0.513", "0.5767" ], [ "0.1184", "0.58", "0.309", "0.58", "0.309", "0.5933", "0.1184", "0.5933" ], [ "0.5146", "0.579", "0.881", "0.579", "0.881", "0.593", "0.5146", "0.593" ], [ "0.5156", "0.5957", "0.8774", "0.5957", "0.8774", "0.6055", "0.5156", "0.6055" ], [ "0.12177", "0.6094", "0.3762", "0.6094", "0.3762", "0.6226", "0.12177", "0.6226" ], [ "0.3938", "0.611", "0.4844", "0.611", "0.4844", "0.6226", "0.3938", "0.6226" ], [ "0.513", "0.61", "0.8774", "0.611", "0.8774", "0.626", "0.513", "0.6245" ], [ "0.11755", "0.6245", "0.4878", "0.6245", "0.4878", "0.638", "0.11755", "0.638" ], [ "0.513", "0.6274", "0.881", "0.6274", "0.881", "0.641", "0.513", "0.641" ], [ "0.1184", "0.641", "0.4878", "0.641", "0.4878", "0.655", "0.1184", "0.655" ], [ "0.513", "0.6445", "0.8774", "0.6445", "0.8774", "0.655", "0.513", "0.655" ], [ "0.12177", "0.659", "0.4836", "0.659", "0.4836", "0.6694", "0.12177", "0.6694" ], [ "0.513", "0.66", "0.881", "0.66", "0.881", "0.6733", "0.513", "0.6733" ], [ "0.1184", "0.6733", "0.4844", "0.6733", "0.4844", "0.687", "0.1184", "0.687" ], [ "0.511", "0.675", "0.881", "0.676", "0.881", "0.6895", "0.511", "0.6885" ], [ "0.11755", "0.6885", "0.4895", "0.6895", "0.4895", "0.7036", "0.11755", "0.703" ], [ "0.51", "0.6914", "0.66", "0.6914", "0.66", "0.705", "0.51", "0.705" ], [ "0.1184", "0.706", "0.486", "0.706", "0.486", "0.7197", "0.1184", "0.7197" ], [ "0.1184", "0.7217", "0.4895", "0.7217", "0.4895", "0.7354", "0.1184", "0.7354" ], [ "0.4902", "0.7256", "0.5054", "0.7236", "0.5073", "0.7285", "0.492", "0.731" ], [ "0.5073", "0.72", "0.642", "0.72", "0.642", "0.734", "0.5073", "0.734" ], [ "0.1184", "0.7383", "0.346", "0.7383", "0.346", "0.752", "0.1184", "0.752" ], [ "0.513", "0.7466", "0.881", "0.7466", "0.881", "0.7603", "0.513", "0.7603" ], [ "0.1368", "0.756", "0.4878", "0.756", "0.4878", "0.7695", "0.1368", "0.7695" ], [ "0.513", "0.7627", "0.881", "0.7627", "0.881", "0.7754", "0.513", "0.7754" ], [ "0.11755", "0.771", "0.486", "0.771", "0.486", "0.7837", "0.11755", "0.7837" ], [ "0.5146", "0.7793", "0.881", "0.7793", "0.881", "0.792", "0.5146", "0.792" ], [ "0.11755", "0.7876", "0.486", "0.7876", "0.486", "0.8003", "0.11755", "0.8003" ], [ "0.51", "0.7935", "0.879", "0.7935", "0.879", "0.807", "0.51", "0.807" ], [ "0.12006", "0.804", "0.4878", "0.804", "0.4878", "0.818", "0.12006", "0.818" ], [ "0.513", "0.811", "0.881", "0.811", "0.881", "0.8247", "0.513", "0.8247" ], [ "0.1184", "0.8193", "0.486", "0.8193", "0.486", "0.833", "0.1184", "0.833" ], [ "0.511", "0.8267", "0.879", "0.8267", "0.879", "0.8403", "0.511", "0.8403" ], [ "0.1184", "0.835", "0.4878", "0.8345", "0.4878", "0.8486", "0.1184", "0.8496" ], [ "0.513", "0.8433", "0.881", "0.8433", "0.881", "0.857", "0.513", "0.857" ], [ "0.1184", "0.8516", "0.4878", "0.8516", "0.4878", "0.865", "0.1184", "0.865" ], [ "0.513", "0.8594", "0.881", "0.8594", "0.881", "0.8726", "0.513", "0.8726" ], [ "0.1184", "0.8677", "0.486", "0.8677", "0.486", "0.8813", "0.1184", "0.8813" ], [ "0.5146", "0.877", "0.8774", "0.877", "0.8774", "0.887", "0.5146", "0.887" ], [ "0.1368", "0.8945", "0.4844", "0.8955", "0.4844", "0.91", "0.1368", "0.909" ], [ "0.513", "0.891", "0.881", "0.891", "0.881", "0.905", "0.513", "0.905" ], [ "0.5156", "0.909", "0.879", "0.909", "0.879", "0.9194", "0.5156", "0.9194" ] ]

[ "行动是建模策略的关键组成部分，用于决定何", "。这些政策未能从数据中完全捕捉到人类的行", "时以及在这种类型的游戏中修复什么", "为，但原因仍然模糊不清", "然而，所提出的基线模型的一个主要缺点是", "另一个悬而未决的问题是如何收集足够数量", "超越性范式：模型没有被训练成真正的对话参", "的高质量CR数据以用于机器学习目的。在带", "与者。相反，他们处理其他人的互动，并且在", "注释的Minecraft对话语料库（??？），TEACh数", "某些时候必须预测何时询问iCR，这是一个与", "据集（??）和CoDraw（???）中，CR是由玩家自", "游戏所需的实际行动脱节的决定。对于听众和", "已主动在真实的多轮互动中发生的，范围从数", "收件人的理解是不同的，后者在建立共同基础", "百到少于一万个已识别的CR话语。仍然在相", "(?）方面具有优势。？认为，心理语言学中的主", "同的大小范围内，IGLU数据集（?？）已在避免", "题实际上通常被视为听者；我们补充说，许多", "配对玩家的设置中收集，并有一次机会要求澄", "NLP方法也在对听者进行建模", "清（并且没有合作伙伴回答并允许进一步的操", "作）", "Contributions在这种背景下，这项工作旨", "其他程序已用于收集大量CR数据。大量数", "在扩展从人类数据中学习元通信行为的开放问", "据集是DialFRED（？），它是通过众包创建的，", "题的边界。为此，我们通过（i）实施更积极的", "这些工人被明确要求生成问题并回答问题，因", "模型来学习何时在CoDraw中询问iCR；（ii）", "为他们实际上并没有参与。在虚拟辅助，会话", "通过定义和建模要询问的任务，朝着更现实的", "搜索和代码生成等相邻领域中，已使用数据增", "代理迈出了一步；最重要的是，（ii）测试三", "强方法（?），用户模拟（?），模板（?）和爬行", "个假设，以研究采取行动对学习iCR政策的影", "QA在线论坛（??）构建了包含CR的大规模数", "响，验证是否可以使用确定性措施来探索iCR", "据集。这些策略可以反映CR形式并促进数据", "能力并为预测提供信息", "收集，但抽象出指令CR的基本触发因素（共", "同努力，实时交互和采取行动），可以说不适", "2.1", "Related Work", "合学习CR策略以遵循指令", "Learningwhentoaskquestions知道何时在", "Evaluating CR mechanisms in dialogue models", "交互中提问的问题出现在各种上下文中。在语", "我们需要更多的评估活动和方法来阐明模型", "言辅助的视觉导航（????）中已经完成了相关", "在CR策略方面实际学到了什么以及为什么失", "工作，其中代理必须在环境中采取行动并决定", "败。一些旨在进行更详细评估的举措正在进行", "何时寻求帮助，其中RL是一种合适的方法。", "中。？通过将CR前后的对象F1得分差异解释", "类似的策略在交互式设置中是必要的，如视觉", "为通过合并澄清而提供的改进，来评估多模态", "对话游戏，需要决定何时停止询问（？）或何时", "模型在共参考分辨率下处理CR的能力；他们", "回答问题的增量预测（?）", "还通过考虑各种CR属性来分析结果。在LLM", "领域，最近的研究通过提示使用评估技术来测", "Modelling clarification requests", "关于描述和", "试模型的能力，得出的结论是，它们可以在一", "建模澄清策略（??????，interalia）的大量文献。", "定程度上检测到歧义，但即使如此，通常也不", "在基于神经网络的NLP时代，该问题通常被", "会尝试修复歧义，并目当它们确实要求澄清时", "分解为从数据中学习的各种任务：何时询问", "与人类策略（??）。当？首次诱导LLM预测适", "（???？），询问（??？）的内容以及如何生成（??？）", "当的对话行为是否要求澄清时，最佳LLM仅", "或选择/排名适当的CRs（???）。理想情况下，", "达到0。28F1得分", "这些任务应该绑定到一个代理中，但是在没有", "集成所有功能（?）的情况下，一些工作仍在以", "", "3Definitions", "“任务框架”的方式解决问题", "CoDraw（?）是一个多模式对话游戏，其中指", "关于何时在指令跟随中要求澄清的建模策", "令跟随器（IF）使用28个（共58个）剪辑库", "略远未解决，因为模型的性能远低于上限。", "重建他们无法看到的场景（来自抽象场景数据", "Minecraft对话数据集中的性能为0。CR类（?)", "集（？））。他们以回合制的方式与指导者（IG）", "的精度为63。在最近的IGLU挑战（？）中，排", "交换短信，该指导者看到原始场景，但无法访", "行榜2中的最佳模型达到0。75加权平均F1得", "问重建场景的状态，除了在游戏过程中有一次", "分。在预测代码生成的规格不足时，最高性能", "机会窥视它。可用的操作是在画布中添加或删", "为0。78二进制F1Score（?）。在Codraw-iCR", "除、移动、翻转和调整剪贴大小。游戏成功是", "中，基线达到了类似的次优0。34平均精度（?)", "通过基于其符号表示的场景相似性得分来衡量", "2Reported in the NeurIPS 2022 IGLU challenge platform.", "的。作者收集了9。9k个这样的英语对话，包", "含大约8k个iCR（11。3%的游戏回合），由？？" ]

{ "arxiv_id": "2401.17039", "language": "zh", "page_idx": 2 }

[ [ "0.1184", "0.0873", "0.3914", "0.0873", "0.3914", "0.10095", "0.1184", "0.10095" ], [ "0.509", "0.0849", "0.6484", "0.0861", "0.6484", "0.10095", "0.509", "0.0998" ], [ "0.1368", "0.1069", "0.486", "0.1069", "0.486", "0.12054", "0.1368", "0.12054" ], [ "0.511", "0.1152", "0.879", "0.1152", "0.879", "0.1289", "0.511", "0.1289" ], [ "0.12006", "0.1223", "0.4878", "0.1223", "0.4878", "0.136", "0.12006", "0.136" ], [ "0.511", "0.1312", "0.881", "0.1312", "0.881", "0.1443", "0.511", "0.1443" ], [ "0.11755", "0.1372", "0.4844", "0.1372", "0.4844", "0.1509", "0.11755", "0.1509" ], [ "0.511", "0.1478", "0.882", "0.1478", "0.882", "0.1609", "0.511", "0.1609" ], [ "0.12006", "0.1561", "0.486", "0.1561", "0.486", "0.1663", "0.12006", "0.1663" ], [ "0.511", "0.1621", "0.879", "0.1621", "0.879", "0.1758", "0.511", "0.1758" ], [ "0.12006", "0.1716", "0.4844", "0.1716", "0.4844", "0.1818", "0.12006", "0.1818" ], [ "0.511", "0.1787", "0.879", "0.1787", "0.879", "0.1924", "0.511", "0.1924" ], [ "0.11755", "0.1853", "0.4844", "0.1853", "0.4844", "0.199", "0.11755", "0.199" ], [ "0.513", "0.1953", "0.5356", "0.1953", "0.5356", "0.2079", "0.513", "0.2079" ], [ "0.1184", "0.2031", "0.4844", "0.2031", "0.4844", "0.2162", "0.1184", "0.2162" ], [ "0.529", "0.2131", "0.879", "0.2131", "0.879", "0.2268", "0.529", "0.2268" ], [ "0.11755", "0.2197", "0.1394", "0.2197", "0.1394", "0.2322", "0.11755", "0.2322" ], [ "0.513", "0.2286", "0.882", "0.2286", "0.882", "0.2423", "0.513", "0.2423" ], [ "0.1368", "0.2382", "0.486", "0.2382", "0.486", "0.2517", "0.1368", "0.2517" ], [ "0.511", "0.2434", "0.879", "0.2446", "0.879", "0.2583", "0.511", "0.257" ], [ "0.1142", "0.253", "0.241", "0.2546", "0.241", "0.2683", "0.1142", "0.2666" ], [ "0.513", "0.2612", "0.879", "0.2612", "0.879", "0.275", "0.513", "0.275" ], [ "0.513", "0.2766", "0.882", "0.2778", "0.882", "0.2915", "0.513", "0.2903" ], [ "0.1351", "0.284", "0.486", "0.284", "0.486", "0.2976", "0.1351", "0.2976" ], [ "0.51", "0.2922", "0.881", "0.2922", "0.881", "0.306", "0.51", "0.306" ], [ "0.1184", "0.3005", "0.4844", "0.3005", "0.4844", "0.3113", "0.1184", "0.3113" ], [ "0.513", "0.308", "0.881", "0.3088", "0.881", "0.3237", "0.513", "0.3225" ], [ "0.1184", "0.3164", "0.4114", "0.3164", "0.4114", "0.33", "0.1184", "0.33" ], [ "0.5264", "0.3267", "0.613", "0.3267", "0.613", "0.3384", "0.5264", "0.3384" ], [ "0.1351", "0.3362", "0.4844", "0.3362", "0.4844", "0.3499", "0.1351", "0.3499" ], [ "0.1184", "0.3516", "0.3972", "0.3516", "0.3972", "0.3652", "0.1184", "0.3652" ], [ "0.511", "0.3486", "0.882", "0.3499", "0.882", "0.364", "0.511", "0.3635" ], [ "0.1351", "0.3718", "0.486", "0.3718", "0.486", "0.3855", "0.1351", "0.3855" ], [ "0.513", "0.3665", "0.882", "0.3665", "0.882", "0.3801", "0.513", "0.3801" ], [ "0.1184", "0.3872", "0.3198", "0.3872", "0.3198", "0.401", "0.1184", "0.401" ], [ "0.513", "0.3818", "0.702", "0.3818", "0.702", "0.3955", "0.513", "0.3955" ], [ "0.1351", "0.4067", "0.4878", "0.4067", "0.4878", "0.4204", "0.1351", "0.4204" ], [ "0.531", "0.408", "0.881", "0.408", "0.881", "0.4216", "0.531", "0.4216" ], [ "0.1184", "0.4216", "0.2737", "0.4229", "0.2737", "0.437", "0.1184", "0.4358" ], [ "0.513", "0.4233", "0.881", "0.4233", "0.881", "0.437", "0.513", "0.437" ], [ "0.5146", "0.4395", "0.881", "0.4395", "0.881", "0.4531", "0.5146", "0.4531" ], [ "0.1368", "0.4531", "0.4878", "0.4531", "0.4878", "0.466", "0.1368", "0.466" ], [ "0.51", "0.456", "0.881", "0.456", "0.881", "0.4697", "0.51", "0.4697" ], [ "0.11755", "0.4685", "0.486", "0.4685", "0.486", "0.4822", "0.11755", "0.4822" ], [ "0.513", "0.4714", "0.881", "0.4714", "0.881", "0.485", "0.513", "0.485" ], [ "0.11755", "0.485", "0.486", "0.485", "0.486", "0.4988", "0.11755", "0.4988" ], [ "0.513", "0.4893", "0.8774", "0.4893", "0.8774", "0.4995", "0.513", "0.4995" ], [ "0.1184", "0.503", "0.4844", "0.503", "0.4844", "0.513", "0.1184", "0.513" ], [ "0.513", "0.5044", "0.881", "0.5044", "0.881", "0.517", "0.513", "0.517" ], [ "0.1184", "0.5166", "0.4895", "0.5166", "0.4895", "0.53", "0.1184", "0.53" ], [ "0.511", "0.5195", "0.881", "0.5195", "0.881", "0.533", "0.511", "0.533" ], [ "0.11755", "0.533", "0.4768", "0.533", "0.4768", "0.5464", "0.11755", "0.5464" ], [ "0.511", "0.535", "0.6035", "0.533", "0.6035", "0.5474", "0.511", "0.55" ], [ "0.531", "0.554", "0.882", "0.554", "0.882", "0.568", "0.531", "0.568" ], [ "0.12427", "0.5693", "0.4878", "0.5693", "0.4878", "0.583", "0.12427", "0.583" ], [ "0.511", "0.5693", "0.881", "0.5693", "0.881", "0.583", "0.511", "0.583" ], [ "0.11755", "0.586", "0.4844", "0.586", "0.4844", "0.5967", "0.11755", "0.5967" ], [ "0.513", "0.585", "0.881", "0.585", "0.881", "0.5986", "0.513", "0.5986" ], [ "0.1184", "0.6016", "0.4844", "0.6016", "0.4844", "0.615", "0.1184", "0.615" ], [ "0.511", "0.601", "0.879", "0.601", "0.879", "0.6143", "0.511", "0.6143" ], [ "0.11755", "0.616", "0.4836", "0.618", "0.4836", "0.633", "0.11755", "0.6313" ], [ "0.5146", "0.618", "0.881", "0.618", "0.881", "0.632", "0.5146", "0.632" ], [ "0.12006", "0.6357", "0.4844", "0.6357", "0.4844", "0.6465", "0.12006", "0.6465" ], [ "0.513", "0.6343", "0.881", "0.6343", "0.881", "0.648", "0.513", "0.648" ], [ "0.1184", "0.651", "0.486", "0.651", "0.486", "0.6646", "0.1184", "0.6646" ], [ "0.513", "0.651", "0.881", "0.651", "0.881", "0.6646", "0.513", "0.6646" ], [ "0.1184", "0.6665", "0.4878", "0.6665", "0.4878", "0.6797", "0.1184", "0.6797" ], [ "0.511", "0.665", "0.881", "0.6665", "0.881", "0.6797", "0.511", "0.6787" ], [ "0.1184", "0.683", "0.339", "0.683", "0.339", "0.697", "0.1184", "0.697" ], [ "0.5146", "0.683", "0.881", "0.683", "0.881", "0.697", "0.5146", "0.697" ], [ "0.513", "0.6987", "0.879", "0.6987", "0.879", "0.712", "0.513", "0.712" ], [ "0.1234", "0.7197", "0.486", "0.7197", "0.486", "0.733", "0.1234", "0.733" ], [ "0.513", "0.7163", "0.879", "0.7163", "0.879", "0.727", "0.513", "0.727" ], [ "0.1184", "0.7354", "0.4844", "0.7354", "0.4844", "0.749", "0.1184", "0.749" ], [ "0.513", "0.7324", "0.5576", "0.7324", "0.5576", "0.7437", "0.513", "0.7437" ], [ "0.1184", "0.752", "0.486", "0.752", "0.486", "0.7656", "0.1184", "0.7656" ], [ "0.513", "0.755", "0.882", "0.755", "0.882", "0.7686", "0.513", "0.7686" ], [ "0.11755", "0.767", "0.4836", "0.767", "0.4836", "0.781", "0.11755", "0.781" ], [ "0.513", "0.771", "0.879", "0.771", "0.879", "0.7837", "0.513", "0.7837" ], [ "0.1184", "0.7837", "0.4836", "0.7837", "0.4836", "0.7944", "0.1184", "0.7944" ], [ "0.511", "0.7876", "0.8774", "0.7876", "0.8774", "0.7974", "0.511", "0.7974" ], [ "0.11755", "0.8", "0.4836", "0.8", "0.4836", "0.8135", "0.11755", "0.8135" ], [ "0.513", "0.8027", "0.882", "0.8027", "0.882", "0.8164", "0.513", "0.8164" ], [ "0.1184", "0.8154", "0.4844", "0.8154", "0.4844", "0.829", "0.1184", "0.829" ], [ "0.51", "0.821", "0.552", "0.821", "0.552", "0.832", "0.51", "0.832" ], [ "0.1184", "0.832", "0.486", "0.832", "0.486", "0.8457", "0.1184", "0.8457" ], [ "0.11755", "0.8467", "0.1663", "0.8467", "0.1663", "0.861", "0.11755", "0.861" ], [ "0.532", "0.8457", "0.881", "0.847", "0.881", "0.8623", "0.532", "0.86" ], [ "0.5146", "0.861", "0.6987", "0.861", "0.6987", "0.8716", "0.5146", "0.8716" ], [ "0.532", "0.8706", "0.8774", "0.8726", "0.8774", "0.887", "0.532", "0.8853" ], [ "0.1368", "0.882", "0.4878", "0.8853", "0.4878", "0.8994", "0.1368", "0.8965" ], [ "0.511", "0.8853", "0.879", "0.8853", "0.879", "0.8984", "0.511", "0.8984" ], [ "0.11755", "0.8955", "0.4895", "0.8965", "0.4895", "0.911", "0.11755", "0.91" ], [ "0.513", "0.8965", "0.881", "0.8965", "0.881", "0.91", "0.513", "0.91" ], [ "0.1184", "0.909", "0.367", "0.909", "0.367", "0.923", "0.1184", "0.923" ], [ "0.513", "0.9077", "0.66", "0.909", "0.66", "0.923", "0.513", "0.9214" ] ]

[ "注释，均在许可CCBY-NC4.0下", "4Hypotheses", "请注意，并非所有iCR都是问题。就情绪而", "在本节中，我们激发并陈述了我们测试的三", "言，大多数CoDraw-iCR都是极性问题，其次", "个假设，作为我们的主要贡献。我们参考了", "是who-和替代问题，但也有陈述式和命令式", "Minecraft游戏中的相关发现，但请注意，Co-", "形式。几乎60个%实例仅引用一个对象，大", "Draw在玩家的共同点方面具有更具挑战性的", "约33个%引用两个对象。被澄清的属性按频", "不对称性：IG在整个游戏中不会观察IF的动", "率顺序是对象之间的关系，场景中的位置，人", "作", "的歧义消除，方向，大小和歧义消除ob对象", "？认为，在场景（?）中检测对象属性的辅助", "(?)", "学习目标对于在参考解析过程中得出的Clark", "我们可以针对此游戏的CR功能拆分可能的", "第三层的参考CR很有用。4我们的期望是行动", "IF模型的空间：", "预测应该与第4级iCR等效相关，这在决定如", "何采取行动时出现。更具体地说，iCR-Action", "??？。Overhearer：观察当前游戏状态（对话", "Takers应该有更真实的动机来决定要求澄清，", "上下文和场景）以预测何时询问iCR的模型", "与Overhearers相比。5为了研究它，我们的第", "而无需任何其他游戏操作或语言决定", "一个假设是：", "???。Action-Taker：一种通过仅采取clipart动", "作而无需iCR决策来玩游戏的模型", "Hypothesis 1: iCR-Action-Takers can learn a", "???。iCR-Action-Taker:一个Action-Taker，可", "more accurate policy for predicting when to ask", "以额外决定何时询问iCR", "an iCR than Overhearers.", "??？。完全代理：做出所有游戏决策的模型，", "在这里，我们还可以测试动作检测是否具有", "包括自然语言生成。", "类似的效果，方法是让模型学习检测给定场景", "前后的动作，就像在？中一样。它是一个框", "Overhearer是NLP中的一个常见范例，其", "架，甚至更等同于？，因为在他们的模型中，", "中模型类似于实际玩家的观察者，决定做", "属性已经在图像中可用。可以在此对话游戏中", "什么，就好像它在他们的鞋子里一样。然", "检香对动作后场景的访问，因为它是基于回合", "而，这是对成熟代理的相当粗略的简化，这", "的：IF会在按下按钮发送下一个消息时完成他", "是一个尚未达到的理想化目标。(iCR-)Action-", "们想要的所有操作（从而看到新编辑的场景）", "Takersareanintermediatestepexaminedinthiswork", "消息或iCR", "接下来，我们的目标是调查在没有任何显式", "Task1我们遵循？何时在CoDraw中要求", "iCR信号的情况下训练的Action-Takers是否仍", "iCR的任务的形式化。简而言之，鉴于游戏状", "然构建对修复需求进行编码的表示。？对量化", "态直到最后一次IG发言，IF必须决定是否要", "NLP任务中的不确定性所做的研究表明，所检", "求澄清。此策略被建模为一个函数fwhen：s→", "查的模型对于更困难的预测会输出更高的数据", "0.1：该函数将当前回合t的游戏状态St映", "不确定性。此外，？提出了这样的假设：如果", "射到此时询问iCR的概率，在游戏的每个回合", "模型对预测不确定，则更有可能是错误的，并", "执行二进制决策任务。在这里，状态s包括对", "使用不确定性作为分数来决定预测是否需要在", "话历史，画廊和场景的情况", "语义解析中进行用户澄清。在此基础上，我们", "推测修复的需求应该表现为行动制定者决策中", "Task2此外，一旦做出询问的决定，玩家", "的不确定性较低。因此，我们检验的第二个假", "还应该知道在这一点上哪些对象需要澄清。因", "设是：", "此，我们定义了要询问的后续任务：在iCR转", "Hypothesis 2: At iCR turns, Action-Takers pre-", "弯t时，函数fwhat：（oi,s）→[0,1]为图库中", "dict actions with less certainty than at other turns.", "的28个对象0；中的每个对象输出询问iCR的", "Similarly less certainty is expected for actions", "概率，给定状态st。这些是对图库中每个可用", "upon objects subject to iCRs than for other ob-", "对象的二进制决策。这两个任务都是在实际生", "jects.", "成之前发生的步骤，我们在这项工作中没有解", "决。3", "4CoDraw-iCR also contains referential CRs, but directly", "relatedtouptakeofinstruction", "'Experiments in the Minecraft dataset point to the oppo", "3we leave the additional decisions of what attributes to", "site direction: Generating action sequences slightly harmed", "mention and which form to realise for ongoing parallel work", "the accuracy on when to ask (?) . We seek to dive deeper into", "dealing specifically with iCR generation.", "understanding this issue" ]

{ "arxiv_id": "2401.17039", "language": "zh", "page_idx": 3 }

[ [ "0.1368", "0.0861", "0.4844", "0.0861", "0.4844", "0.0998", "0.1368", "0.0998" ], [ "0.529", "0.08905", "0.613", "0.08905", "0.613", "0.0968", "0.529", "0.0968" ], [ "0.812", "0.0903", "0.8364", "0.0903", "0.8364", "0.0974", "0.812", "0.0974" ], [ "0.1184", "0.10156", "0.4844", "0.10156", "0.4844", "0.1164", "0.1184", "0.1164" ], [ "0.804", "0.10925", "0.8423", "0.10925", "0.8423", "0.1217", "0.804", "0.1217" ], [ "0.1184", "0.11816", "0.4844", "0.11816", "0.4844", "0.1318", "0.1184", "0.1318" ], [ "0.5264", "0.1271", "0.5938", "0.1271", "0.5938", "0.1342", "0.5264", "0.1342" ], [ "0.6035", "0.1223", "0.642", "0.1247", "0.6406", "0.136", "0.602", "0.1342" ], [ "0.6665", "0.1271", "0.712", "0.1271", "0.712", "0.1342", "0.6665", "0.1342" ], [ "0.738", "0.1271", "0.771", "0.1271", "0.771", "0.1342", "0.738", "0.1342" ], [ "0.11755", "0.133", "0.4836", "0.133", "0.4836", "0.1467", "0.11755", "0.1467" ], [ "0.5415", "0.1437", "0.58", "0.1437", "0.58", "0.155", "0.5415", "0.155" ], [ "0.602", "0.1467", "0.6465", "0.1467", "0.6465", "0.158", "0.602", "0.158" ], [ "0.617", "0.1443", "0.636", "0.1443", "0.636", "0.1478", "0.617", "0.1478" ], [ "0.6675", "0.1467", "0.711", "0.1467", "0.711", "0.158", "0.6675", "0.158" ], [ "0.733", "0.1467", "0.775", "0.1467", "0.775", "0.158", "0.733", "0.158" ], [ "0.552", "0.1561", "0.565", "0.1561", "0.565", "0.1603", "0.552", "0.1603" ], [ "0.11584", "0.1484", "0.3972", "0.1455", "0.3972", "0.1644", "0.11584", "0.1675" ], [ "0.6094", "0.158", "0.644", "0.158", "0.644", "0.1663", "0.6094", "0.1663" ], [ "0.6753", "0.158", "0.7095", "0.158", "0.7095", "0.165", "0.6753", "0.165" ], [ "0.7407", "0.158", "0.771", "0.158", "0.771", "0.165", "0.7407", "0.165" ], [ "0.1368", "0.1675", "0.486", "0.1675", "0.486", "0.1812", "0.1368", "0.1812" ], [ "0.1184", "0.1829", "0.4727", "0.1829", "0.4727", "0.1965", "0.1184", "0.1965" ], [ "0.6543", "0.1859", "0.802", "0.1882", "0.802", "0.2102", "0.6543", "0.2079" ], [ "0.8564", "0.1818", "0.8716", "0.1841", "0.8696", "0.1895", "0.8555", "0.187" ], [ "0.11755", "0.1978", "0.2007", "0.1996", "0.2007", "0.2144", "0.11755", "0.212" ], [ "0.735", "0.1978", "0.7573", "0.1978", "0.7573", "0.2019", "0.735", "0.2019" ], [ "0.11755", "0.2214", "0.3225", "0.2214", "0.3225", "0.2351", "0.11755", "0.2351" ], [ "0.346", "0.2227", "0.486", "0.2227", "0.486", "0.2328", "0.346", "0.2328" ], [ "0.771", "0.2173", "0.8237", "0.2173", "0.8237", "0.228", "0.771", "0.228" ], [ "0.6724", "0.2239", "0.706", "0.2239", "0.706", "0.2322", "0.6724", "0.2322" ], [ "0.1142", "0.2369", "0.486", "0.2382", "0.486", "0.253", "0.1142", "0.2517" ], [ "0.7925", "0.2405", "0.839", "0.2405", "0.839", "0.2505", "0.7925", "0.2505" ], [ "0.11584", "0.2542", "0.482", "0.2542", "0.482", "0.2673", "0.11584", "0.2673" ], [ "0.11755", "0.284", "0.1511", "0.284", "0.1511", "0.2964", "0.11755", "0.2964" ], [ "0.147", "0.2808", "0.22", "0.2832", "0.22", "0.2976", "0.147", "0.2957" ], [ "0.634", "0.2976", "0.714", "0.2976", "0.714", "0.3047", "0.634", "0.3047" ], [ "0.769", "0.2957", "0.8237", "0.2976", "0.8237", "0.3088", "0.769", "0.3071" ], [ "0.6406", "0.3027", "0.7095", "0.3027", "0.7095", "0.3135", "0.6406", "0.3135" ], [ "0.561", "0.2986", "0.6094", "0.2998", "0.6094", "0.3071", "0.561", "0.306" ], [ "0.5425", "0.304", "0.6094", "0.304", "0.6094", "0.3142", "0.5425", "0.3142" ], [ "0.748", "0.304", "0.797", "0.304", "0.797", "0.3142", "0.748", "0.3142" ], [ "0.818", "0.304", "0.846", "0.306", "0.8447", "0.3135", "0.817", "0.3113" ], [ "0.1184", "0.31", "0.486", "0.31", "0.486", "0.3237", "0.1184", "0.3237" ], [ "0.7656", "0.3113", "0.827", "0.3088", "0.827", "0.3206", "0.7656", "0.3225" ], [ "0.11755", "0.3247", "0.486", "0.3267", "0.486", "0.3403", "0.11755", "0.3384" ], [ "0.51", "0.3342", "0.881", "0.3342", "0.881", "0.3474", "0.51", "0.3474" ], [ "0.1184", "0.3425", "0.486", "0.3425", "0.486", "0.3557", "0.1184", "0.3557" ], [ "0.511", "0.3499", "0.5557", "0.3499", "0.5557", "0.361", "0.511", "0.361" ], [ "0.5483", "0.3486", "0.882", "0.3486", "0.882", "0.3623", "0.5483", "0.3623" ], [ "0.12006", "0.3599", "0.4836", "0.3599", "0.4836", "0.3706", "0.12006", "0.3706" ], [ "0.51", "0.3635", "0.882", "0.3635", "0.882", "0.3772", "0.51", "0.3772" ], [ "0.1184", "0.3748", "0.486", "0.3748", "0.486", "0.3884", "0.1184", "0.3884" ], [ "0.513", "0.379", "0.881", "0.379", "0.881", "0.3896", "0.513", "0.3896" ], [ "0.11755", "0.3901", "0.4844", "0.3901", "0.4844", "0.405", "0.11755", "0.405" ], [ "0.511", "0.3926", "0.8237", "0.3926", "0.8237", "0.4026", "0.511", "0.4026" ], [ "0.12006", "0.4067", "0.4836", "0.4067", "0.4836", "0.4204", "0.12006", "0.4204" ], [ "0.1335", "0.4246", "0.4836", "0.4246", "0.4836", "0.4353", "0.1335", "0.4353" ], [ "0.11755", "0.4395", "0.486", "0.4395", "0.486", "0.4531", "0.11755", "0.4531" ], [ "0.513", "0.4412", "0.879", "0.4412", "0.879", "0.4548", "0.513", "0.4548" ], [ "0.11755", "0.4573", "0.3745", "0.4573", "0.3745", "0.4673", "0.11755", "0.4673" ], [ "0.5156", "0.4578", "0.881", "0.4578", "0.881", "0.4714", "0.5156", "0.4714" ], [ "0.1335", "0.4714", "0.4844", "0.4714", "0.4844", "0.485", "0.1335", "0.485" ], [ "0.5146", "0.4739", "0.881", "0.4739", "0.881", "0.4868", "0.5146", "0.4868" ], [ "0.11755", "0.4863", "0.486", "0.4868", "0.486", "0.502", "0.11755", "0.5005" ], [ "0.513", "0.4905", "0.881", "0.4905", "0.881", "0.5044", "0.513", "0.5044" ], [ "0.1184", "0.5044", "0.486", "0.5044", "0.486", "0.517", "0.1184", "0.517" ], [ "0.51", "0.506", "0.881", "0.506", "0.881", "0.5195", "0.51", "0.5195" ], [ "0.12006", "0.5195", "0.4844", "0.5195", "0.4844", "0.533", "0.12006", "0.533" ], [ "0.51", "0.521", "0.881", "0.5215", "0.881", "0.536", "0.51", "0.535" ], [ "0.12006", "0.536", "0.4844", "0.536", "0.4844", "0.55", "0.12006", "0.55" ], [ "0.513", "0.539", "0.55", "0.539", "0.55", "0.552", "0.513", "0.552" ], [ "0.12006", "0.5527", "0.4844", "0.5527", "0.4844", "0.5664", "0.12006", "0.5664" ], [ "0.529", "0.5586", "0.881", "0.5586", "0.881", "0.5723", "0.529", "0.5723" ], [ "0.1184", "0.5684", "0.4844", "0.5684", "0.4844", "0.582", "0.1184", "0.582" ], [ "0.5146", "0.5747", "0.881", "0.576", "0.881", "0.589", "0.5146", "0.5884" ], [ "0.1184", "0.585", "0.3838", "0.585", "0.3838", "0.5986", "0.1184", "0.5986" ], [ "0.5146", "0.5913", "0.881", "0.5913", "0.881", "0.605", "0.5146", "0.605" ], [ "0.1377", "0.6025", "0.486", "0.6025", "0.486", "0.6133", "0.1377", "0.6133" ], [ "0.513", "0.607", "0.8774", "0.607", "0.8774", "0.6206", "0.513", "0.6206" ], [ "0.11755", "0.6177", "0.4844", "0.6177", "0.4844", "0.6313", "0.11755", "0.6313" ], [ "0.513", "0.6235", "0.881", "0.6235", "0.881", "0.637", "0.513", "0.637" ], [ "0.11755", "0.633", "0.4844", "0.633", "0.4844", "0.6465", "0.11755", "0.6465" ], [ "0.511", "0.638", "0.881", "0.64", "0.881", "0.654", "0.511", "0.652" ], [ "0.1184", "0.6484", "0.486", "0.6484", "0.486", "0.662", "0.1184", "0.662" ], [ "0.513", "0.6562", "0.882", "0.6562", "0.882", "0.6694", "0.513", "0.6694" ], [ "0.1184", "0.665", "0.4844", "0.665", "0.4844", "0.6787", "0.1184", "0.6787" ], [ "0.5146", "0.673", "0.769", "0.673", "0.769", "0.683", "0.5146", "0.683" ], [ "0.11755", "0.681", "0.4844", "0.6797", "0.4844", "0.695", "0.11755", "0.6953" ], [ "0.529", "0.6914", "0.881", "0.6924", "0.881", "0.707", "0.529", "0.706" ], [ "0.12006", "0.6978", "0.3938", "0.6978", "0.3938", "0.7114", "0.12006", "0.7114" ], [ "0.1351", "0.7144", "0.486", "0.7144", "0.486", "0.728", "0.1351", "0.728" ], [ "0.513", "0.709", "0.881", "0.709", "0.881", "0.7227", "0.513", "0.7227" ], [ "0.511", "0.7236", "0.879", "0.7236", "0.879", "0.737", "0.511", "0.737" ], [ "0.1184", "0.731", "0.486", "0.731", "0.486", "0.7446", "0.1184", "0.7446" ], [ "0.511", "0.7393", "0.881", "0.7407", "0.881", "0.755", "0.511", "0.7544" ], [ "0.1184", "0.7466", "0.4844", "0.7466", "0.4844", "0.7603", "0.1184", "0.7603" ], [ "0.513", "0.7573", "0.8774", "0.7573", "0.8774", "0.771", "0.513", "0.771" ], [ "0.11755", "0.763", "0.486", "0.763", "0.486", "0.777", "0.11755", "0.777" ], [ "0.511", "0.7695", "0.881", "0.771", "0.881", "0.788", "0.511", "0.7876" ], [ "0.1368", "0.786", "0.4895", "0.7876", "0.4895", "0.802", "0.1368", "0.8003" ], [ "0.511", "0.788", "0.879", "0.788", "0.879", "0.802", "0.511", "0.802" ], [ "0.11584", "0.8", "0.4895", "0.8", "0.4895", "0.8135", "0.11584", "0.8135" ], [ "0.511", "0.8057", "0.881", "0.8057", "0.881", "0.8193", "0.511", "0.8193" ], [ "0.1184", "0.814", "0.4878", "0.814", "0.4878", "0.8247", "0.1184", "0.8247" ], [ "0.1184", "0.826", "0.4878", "0.826", "0.4878", "0.836", "0.1184", "0.836" ], [ "0.513", "0.822", "0.882", "0.822", "0.882", "0.835", "0.513", "0.835" ], [ "0.11755", "0.836", "0.4878", "0.836", "0.4878", "0.8496", "0.11755", "0.8496" ], [ "0.511", "0.836", "0.881", "0.8374", "0.881", "0.8516", "0.511", "0.851" ], [ "0.11584", "0.847", "0.3745", "0.847", "0.3745", "0.861", "0.11584", "0.861" ], [ "0.1411", "0.8594", "0.4878", "0.86", "0.4878", "0.8735", "0.1411", "0.8726" ], [ "0.513", "0.854", "0.712", "0.854", "0.712", "0.8677", "0.513", "0.8677" ], [ "0.11584", "0.8716", "0.486", "0.8706", "0.486", "0.8843", "0.11584", "0.8853" ], [ "0.11755", "0.8843", "0.4878", "0.8843", "0.4878", "0.898", "0.11755", "0.898" ], [ "0.1184", "0.898", "0.4878", "0.898", "0.4878", "0.9077", "0.1184", "0.9077" ], [ "0.532", "0.8936", "0.882", "0.8955", "0.882", "0.91", "0.532", "0.9077" ], [ "0.11755", "0.9067", "0.1545", "0.909", "0.1528", "0.9204", "0.11584", "0.9185" ], [ "0.511", "0.9077", "0.881", "0.9077", "0.881", "0.9214", "0.511", "0.9214" ] ]

[ "对于此步骤，我们设置了联系假设，即确定", "bsent in Overhearel", "P(iCR", "性以模型分配给在给定回合处采取行动或不采", "oredcRte", "取行动的概率表示。这是一个合理的假设，因", "P(add/remove)", "P(fLip)", "P(resize)", "P(move)", "为期望目标函数将预测推至0或1，因此预测", "redicto", "practicoto", "", "prectioto", "redicoto", "ado", "接近0。5可以看作是优柔寡断。6", "(flip)", "resize", "(move)", "最后，关于何时询问的iCR政策应以精确表", "示不清楚的内容为基础。因此，我们的最后", "contextapl.achasijea", "aker", "个假设是：", "", "Hypothesis 3: Pre-trained", "iCR-Action-Takers", "Transformer", "memory", "can learn a more accurate policy for predicting", "+target", "what to ask about in iCR turns than Overhearers.", "5", "Models", "dialoque contex", "28objects", "(pretrained)", "e features", "(pretrained)", "in galler", "scene", "在本节中，我们将介绍我们在实验中分析的模", "(symbolic)", "型。我们不打算提出一个新的架构，因为我们", "Figure 2: The basic structure of our iCR policy mod", "的目的是了解当前SotA模型失败的原因以及", "els.", "The full structure represents the iCR-Action-", "学习采取行动对它们的影响。我们实现了一个", "Taker. The Overhearer contains no action predictor", "模型，通过在最近的多模式对话挑战中结合顶", "(area shaded in grey), whereas the Action-Taker con-", "级模型的技术，即IGLU（?）和SIMMC2，解决", "tains noiCRpredictor (areainthedottedbox)", "了？基线模型（iCR-baseline）的局限性。（？)", "Overhearer和（iCR-)Action-Taker的基本体系", "结构如图2所示。我们在此提供其信息流的概", "目前为止的对话，由BERT（?）构建的令牌级", "述；有关详细规范，请参见附录", "上下文嵌人表示，当前场景，由ResNet（?）主", "CoDrawIF可以访问28个对象的图库，这是", "干构建的图像特征表示，然后是可训练的卷积", "游戏中的信息来源（例如。g。如果它只包含", "层，以减少通道数，如DETR模型（?）。我们", "三个树片段中的一个，则不太可能需要消除歧", "将文本和场景作为一个序列提供，如？。iCR", "义)，但在iCR基线中不存在。我们遵循一种符", "Action-Detecters的变体在操作之前和之后访问", "号方法来表示对象的属性（场景中的存在，方", "场景", "向，位置，大小，姿势，面部表情）基于？中", "Transformer输出每个对象的上下文表示。到", "的原始抽屉（但是，它对数据库中所有可能的", "目前为止的步骤在图2的下部表示。现在我们", "对象都具有不切实际的访问权限）", "继续进行上半部分的预测，根据模型的类型而", "以前的作品没有使用Transformers（?）在Co-", "有所不同。为了检验我们的假设，我们实现了", "Draw中对iCR策略进行建模。鉴于其在多个", "预测游戏动作（或检测它们，如果使用更新的", "场景中的领先性能，我们以DETR（?）启发的", "图像）和/或通过多任务学习做出iCR决策的", "方式将它们带到现场。我们使用Transformer", "模型。我们从？中获得灵感，将上下文对象嵌", "解码器7模块来创建当前游戏状态下每个对象", "入训练为所有分类器的联合编码", "的上下文嵌人，即。e。通过建立一个考虑到", "动作预测器和iCR预测器被实现为具有丢", "目前为止的对话和实际场景的表示", "这是通过将每个对象传递给Transformer解", "失的2层前馈网络，该网络将表示作为输入并", "输出概率。在（iCR-）Action-Takers中，我们", "码器（“目标”）来完成的，以允许自我关注", "将每个动作预测（添加/删除，翻转，调整大", "画廊的状态，以及随后与游戏状态表示（“内", "小，移动）建模为在每个对象嵌人时完成的二", "存”）的交叉关注。状态有两个组成部分：到", "进制分类。8iCR决策也作为二进制分类任务", "6An investigation of the predictive uncertainty of the IF", "执行。在任务1（何时询问）中，它预测是否", "model in the Minecraft data has been done by ? using length-", "应在当前回合进行iCR。在任务2中，（问什么）", "normalized log-likelihood and entropy ofgenerated action se-", "quencesNegative results are reported in an unpublished", "它预测，对于每个对象，它是否受到和iCR。", "short manuscript concluding that uncertainty is not a direct", "在iCR-Action-Takers中，我们让动作日志成为", "signal for when to ask CRs in their setting.", "Thefull Transformerencoder-decoder was detrimental in", "iCR预测器输入的一部分", "almost all cases, so we report results using only the decoder", "component. This is probably due to the fact that the scene and", "dialoguehad alreadybeen encoded by thepretrained compo-", "8To facilitate evaluation, we add an additional meta-action", "nents", "prediction which is I whenever any action is made to a clipart." ]

{ "arxiv_id": "2401.17039", "language": "zh", "page_idx": 4 }

[ [ "0.2852", "0.1351", "0.687", "0.1351", "0.687", "0.1497", "0.2852", "0.1497" ], [ "0.1487", "0.1818", "0.3718", "0.1812", "0.3718", "0.1989", "0.1487", "0.1995" ], [ "0.4167", "0.1831", "0.607", "0.1818", "0.607", "0.1963", "0.4167", "0.1976" ], [ "0.6562", "0.1831", "0.819", "0.1831", "0.819", "0.1976", "0.6562", "0.1976" ], [ "0.2009", "0.2203", "0.3694", "0.2197", "0.3694", "0.2374", "0.2009", "0.238" ], [ "0.437", "0.2229", "0.5703", "0.2229", "0.5703", "0.2374", "0.437", "0.2374" ], [ "0.6406", "0.2203", "0.7656", "0.2184", "0.7656", "0.2374", "0.6406", "0.2393" ], [ "0.2263", "0.238", "0.3342", "0.2405", "0.3342", "0.2551", "0.2263", "0.2524" ], [ "0.4355", "0.238", "0.5596", "0.2412", "0.5596", "0.2556", "0.4355", "0.2524" ], [ "0.636", "0.2361", "0.767", "0.2405", "0.7656", "0.2583", "0.6357", "0.254" ], [ "0.1871", "0.26", "0.3865", "0.26", "0.3865", "0.2734", "0.1871", "0.2734" ], [ "0.4182", "0.26", "0.6113", "0.26", "0.6113", "0.2734", "0.4182", "0.2734" ], [ "0.653", "0.2615", "0.7466", "0.2615", "0.7466", "0.2727", "0.653", "0.2727" ], [ "0.4617", "0.2747", "0.574", "0.2747", "0.574", "0.2854", "0.4617", "0.2854" ], [ "0.2385", "0.324", "0.312", "0.324", "0.312", "0.3384", "0.2385", "0.3384" ], [ "0.6055", "0.3245", "0.7646", "0.3245", "0.7646", "0.3352", "0.6055", "0.3352" ], [ "0.638", "0.336", "0.7563", "0.336", "0.7563", "0.3435", "0.638", "0.3435" ], [ "0.794", "0.3352", "0.8784", "0.3352", "0.8784", "0.3423", "0.794", "0.3423" ], [ "0.6055", "0.3435", "0.7656", "0.3435", "0.7656", "0.3567", "0.6055", "0.3567" ], [ "0.806", "0.3435", "0.8677", "0.3435", "0.8677", "0.351", "0.806", "0.351" ], [ "0.10046", "0.356", "0.4656", "0.356", "0.4656", "0.3694", "0.10046", "0.3694" ], [ "0.5986", "0.3662", "0.6846", "0.3662", "0.6846", "0.377", "0.5986", "0.377" ], [ "0.7017", "0.3662", "0.7563", "0.3662", "0.7563", "0.3774", "0.7017", "0.3774" ], [ "0.819", "0.3694", "0.8325", "0.3694", "0.8325", "0.3738", "0.819", "0.3738" ], [ "0.0809", "0.3865", "0.468", "0.3865", "0.468", "0.401", "0.0809", "0.401" ], [ "0.0784", "0.3706", "0.4673", "0.3718", "0.4673", "0.3865", "0.0784", "0.385" ], [ "0.7017", "0.3926", "0.7573", "0.3926", "0.7573", "0.4028", "0.7017", "0.4028" ], [ "0.796", "0.3914", "0.884", "0.3914", "0.884", "0.3984", "0.796", "0.3984" ], [ "0.0801", "0.401", "0.468", "0.4016", "0.468", "0.4167", "0.0801", "0.4153" ], [ "0.6396", "0.3933", "0.6846", "0.3933", "0.6846", "0.3965", "0.6396", "0.3965" ], [ "0.5034", "0.4092", "0.8906", "0.4092", "0.8906", "0.4224", "0.5034", "0.4224" ], [ "0.0809", "0.4172", "0.468", "0.4172", "0.468", "0.4312", "0.0809", "0.4312" ], [ "0.5034", "0.4248", "0.888", "0.4248", "0.888", "0.435", "0.5034", "0.435" ], [ "0.0825", "0.4343", "0.4673", "0.4343", "0.4673", "0.4443", "0.0825", "0.4443" ], [ "0.5024", "0.4375", "0.892", "0.4375", "0.892", "0.4507", "0.5024", "0.4507" ], [ "0.0809", "0.4475", "0.468", "0.4475", "0.468", "0.461", "0.0809", "0.461" ], [ "0.5024", "0.452", "0.5376", "0.452", "0.5376", "0.4634", "0.5024", "0.4634" ], [ "0.0809", "0.4634", "0.468", "0.4634", "0.468", "0.4766", "0.0809", "0.4766" ], [ "0.524", "0.4702", "0.8906", "0.4702", "0.8906", "0.4841", "0.524", "0.4841" ], [ "0.0801", "0.4778", "0.4673", "0.4778", "0.4673", "0.4912", "0.0801", "0.4912" ], [ "0.505", "0.486", "0.8906", "0.486", "0.8906", "0.5", "0.505", "0.5" ], [ "0.0809", "0.4937", "0.468", "0.4937", "0.468", "0.507", "0.0809", "0.507" ], [ "0.5024", "0.5", "0.892", "0.5005", "0.892", "0.515", "0.5024", "0.5146" ], [ "0.0809", "0.5083", "0.468", "0.5083", "0.468", "0.5215", "0.0809", "0.5215" ], [ "0.5034", "0.5166", "0.8906", "0.5166", "0.8906", "0.53", "0.5034", "0.53" ], [ "0.0801", "0.5234", "0.4673", "0.5234", "0.4673", "0.537", "0.0801", "0.537" ], [ "0.505", "0.531", "0.8906", "0.531", "0.8906", "0.5444", "0.505", "0.5444" ], [ "0.0784", "0.537", "0.468", "0.5386", "0.468", "0.553", "0.0784", "0.552" ], [ "0.505", "0.547", "0.8906", "0.547", "0.8906", "0.56", "0.505", "0.56" ], [ "0.0809", "0.553", "0.4673", "0.553", "0.4673", "0.5664", "0.0809", "0.5664" ], [ "0.5034", "0.561", "0.8906", "0.561", "0.8906", "0.5747", "0.5034", "0.5747" ], [ "0.0809", "0.5684", "0.468", "0.5684", "0.468", "0.582", "0.0809", "0.582" ], [ "0.505", "0.577", "0.8906", "0.577", "0.8906", "0.5903", "0.505", "0.5903" ], [ "0.0825", "0.5854", "0.4673", "0.5854", "0.4673", "0.595", "0.0825", "0.595" ], [ "0.0784", "0.5977", "0.464", "0.5986", "0.464", "0.6123", "0.0784", "0.6113" ], [ "0.505", "0.5923", "0.8906", "0.5923", "0.8906", "0.606", "0.505", "0.606" ], [ "0.0809", "0.6123", "0.4673", "0.6123", "0.4673", "0.627", "0.0809", "0.627" ], [ "0.5034", "0.607", "0.8906", "0.607", "0.8906", "0.6206", "0.5034", "0.6206" ], [ "0.5034", "0.622", "0.8906", "0.622", "0.8906", "0.6353", "0.5034", "0.6353" ], [ "0.0809", "0.629", "0.4697", "0.629", "0.4697", "0.6426", "0.0809", "0.6426" ], [ "0.505", "0.637", "0.892", "0.637", "0.892", "0.651", "0.505", "0.651" ], [ "0.0809", "0.646", "0.464", "0.646", "0.464", "0.656", "0.0809", "0.656" ], [ "0.5034", "0.651", "0.7974", "0.653", "0.7974", "0.6665", "0.5034", "0.664" ], [ "0.0809", "0.659", "0.468", "0.659", "0.468", "0.6724", "0.0809", "0.6724" ], [ "0.0809", "0.675", "0.3162", "0.674", "0.3162", "0.6875", "0.0809", "0.688" ], [ "0.5215", "0.6895", "0.8906", "0.69", "0.8906", "0.705", "0.5215", "0.704" ], [ "0.0784", "0.7085", "0.2173", "0.709", "0.2173", "0.7236", "0.0784", "0.7227" ], [ "0.5034", "0.7056", "0.8906", "0.7056", "0.8906", "0.7197", "0.5034", "0.7197" ], [ "0.505", "0.72", "0.8906", "0.72", "0.8906", "0.7344", "0.505", "0.7344" ], [ "0.1013", "0.7363", "0.4673", "0.7363", "0.4673", "0.75", "0.1013", "0.75" ], [ "0.505", "0.7363", "0.8896", "0.7363", "0.8896", "0.75", "0.505", "0.75" ], [ "0.0801", "0.75", "0.468", "0.7505", "0.468", "0.765", "0.0801", "0.7646" ], [ "0.5024", "0.7505", "0.8906", "0.75", "0.8906", "0.7646", "0.5024", "0.765" ], [ "0.0809", "0.7666", "0.468", "0.7666", "0.468", "0.781", "0.0809", "0.781" ], [ "0.5034", "0.7666", "0.8906", "0.7666", "0.8906", "0.78", "0.5034", "0.78" ], [ "0.0825", "0.7827", "0.4656", "0.7827", "0.4656", "0.7935", "0.0825", "0.7935" ], [ "0.505", "0.7827", "0.8896", "0.7827", "0.8896", "0.7935", "0.505", "0.7935" ], [ "0.0825", "0.797", "0.4673", "0.797", "0.4673", "0.81", "0.0825", "0.81" ], [ "0.5034", "0.797", "0.8906", "0.797", "0.8906", "0.81", "0.5034", "0.81" ], [ "0.0825", "0.811", "0.4673", "0.811", "0.4673", "0.8247", "0.0825", "0.8247" ], [ "0.5024", "0.811", "0.8896", "0.811", "0.8896", "0.8247", "0.5024", "0.8247" ], [ "0.0825", "0.8276", "0.4673", "0.8276", "0.4673", "0.8384", "0.0825", "0.8384" ], [ "0.505", "0.827", "0.892", "0.827", "0.892", "0.8403", "0.505", "0.8403" ], [ "0.0809", "0.8413", "0.468", "0.8413", "0.468", "0.855", "0.0809", "0.855" ], [ "0.505", "0.8413", "0.8896", "0.8413", "0.8896", "0.855", "0.505", "0.855" ], [ "0.0809", "0.857", "0.468", "0.857", "0.468", "0.8706", "0.0809", "0.8706" ], [ "0.5034", "0.857", "0.8906", "0.857", "0.8906", "0.8706", "0.5034", "0.8706" ], [ "0.0801", "0.8706", "0.468", "0.872", "0.468", "0.886", "0.0801", "0.8853" ], [ "0.505", "0.872", "0.8906", "0.872", "0.8906", "0.8853", "0.505", "0.8853" ], [ "0.0809", "0.887", "0.326", "0.887", "0.326", "0.901", "0.0809", "0.901" ], [ "0.5034", "0.886", "0.739", "0.887", "0.739", "0.9014", "0.5034", "0.901" ], [ "0.4805", "0.925", "0.4927", "0.925", "0.4927", "0.9375", "0.4805", "0.9375" ] ]

[ "Captioning Images with Diverse Objects", "Subhashini Venugopalan t", "Lisa Anne Hendricks*", "Marcus Rohrbach *", "Raymond Mooney t", "Trevor Darrell *", "Kate Saenko t", "t UT Austin", "*UC Berkeley", "+ Boston Univ.", "[ vsub,mooney @cs.utexas.edu", "[ lisa_anne, rohrbach, trevor }", "[email protected]", "@eecs.berkeley.edu", "Abstract", "NOC (ours): Jointly train on multiple", "eswithauxiliaryobjectives", "Aokapistandinginthe", "IMGENET+MSCOCO+", "middle of a field.", "Recent captioning models are limited in their ability", "Visual Classifiers.", "IMAGENET", "kani", "text corpora. We propose the Novel Object Captioner", "to scale and describe concepts unseen in paired image-", "MSCOCO", "orse standing in the dirt", "(NOC), a deep visual semantic captioning model", "aantnnar", "Figure 1. We propose a model that learns simultaneously", "that can describe a large number of object categories", "from multiple data sources with auxiliary objectives to de", "not present in existing image-caption datasets.Our", "scribe a variety of objects unseen in paired image-caption", "model takes advantage of external sources - labeled", "data.", "images from object recognition datasets, and semantic", "最近的工作「7]表明，为了在没有明确的成对字幕", "knowledge extracted from unannotated text. We pro-", "训练数据的情况下整合当前视觉识别网络的大量知识，", "pose minimizing a joint objective which can learn from", "字幕模型可以从外部来源学习，并学习撰写关于视觉", "these diverse data sources and leverage distributional", "概念的句子，这些视觉概念在图像描述语料库中不常", "semantic embeddings, enabling the model to general-", "见或不存在。但是，[7]的开创性DCC模型很笨拙，因", "ize and describe novel objects outside of image-caption", "为该模型需要将学习到的参数从以前看到的类别显式", "datasets. We demonstrate that our model exploits se-", "转移（“复制”）到新颖的类别。这不仅使其无法描述稀", "mantic information to generate captions for hundreds", "有类别，并限制了模型覆盖更多对象的能力，而且使其", "of object categories in the ImageNet object recogni-", "tion dataset that are not observed in MSCOCO image", "无法进行端到端训练。相反，我们提出了新颖的对象捕", "caption training data, as well as many categories that", "获器（NOC），该网络可以使用联合训练策略进行端", "到端训练，以整合来自外部视觉识别数据集的知识以", "are observed very rarely. Both automatic evaluations", "及来自独立的未注释文本语料库的语义信息，以生成", "and human judgements showthat our model consider", "各种稀有和新颖对象的字幕（如图。1）", "ably outperforms prior work in being able to describe", "many more categories of objects.", "具体来说，我们引入了辅助目标，使我们的网络能", "1.Introduction", "够同时学习图像-字幕对上的字幕模型以及未注释文本", "和标记图像上的深度语言模型和视觉识别系统。与以", "现代视觉分类器E[6,22]可以识别数千个对象类别，", "前的工作不同，辅助目标允许NOC模型以端到端的方", "其中一些是基本或人门级的（e。g。电视)，以及其他细", "式同时从多个数据源学习相关信息。此外，NOC隐式", "粒度和特定于任务的（例如。g。拨号电话、手机)。但", "地利用了预先训练的分布式单词嵌人，使其能够描述", "是，最近最先进的视觉字幕系统F2.3.8，10.15，26", "看不见的和稀有的对象类别。我们工作的主要贡献是", "可以直接从图像和描述中学习，仅依靠配对的图像字", "1）端到端模型来描述配对图像字幕数据中不存在的对", "幕数据进行监督，并且无法概括和描述大量可识别的", "象，2）在多个数据源上对视觉和语言模型进行辅助/联", "对象在上下文中。虽然可以通过构建更大的图像/视频", "合训练，以及3）结合了预先训练的任务语义嵌人。我", "描述数据集来扩展此类系统，但获取此类字幕数据将", "们通过对MSCOCO[13]中的对象以及标题数据集中", "是昂贵且费力的。此外，视觉描述具有挑战性，因为模", "看不见的ImageNet[21]中的数百个对象进行广泛的实", "型不仅必须正确识别图像中包含的视觉概念，而且还", "验，证明了模型的有效性。我们的模型在自动和人工评", "必须将这些概念组合成连贯的句子", "估方面都大大优于以前的工作[7]", "1" ]

{ "arxiv_id": "1606.07770", "language": "zh", "page_idx": 0 }

[ [ "0.0784", "0.09155", "0.2279", "0.089", "0.2279", "0.1042", "0.0784", "0.1061" ], [ "0.661", "0.09216", "0.7295", "0.09216", "0.7295", "0.1029", "0.661", "0.1029" ], [ "0.5146", "0.101", "0.5933", "0.1048", "0.5923", "0.1193", "0.513", "0.11554" ], [ "0.661", "0.1061", "0.744", "0.1061", "0.744", "0.11743", "0.661", "0.11743" ], [ "0.8013", "0.1061", "0.8447", "0.1061", "0.8447", "0.11743", "0.8013", "0.11743" ], [ "0.08496", "0.11554", "0.2336", "0.11615", "0.2336", "0.1307", "0.08496", "0.13" ], [ "0.2484", "0.11743", "0.468", "0.11743", "0.468", "0.1307", "0.2484", "0.1307" ], [ "0.0825", "0.1307", "0.4656", "0.1307", "0.4656", "0.1445", "0.0825", "0.1445" ], [ "0.699", "0.127", "0.709", "0.127", "0.709", "0.1307", "0.699", "0.1307" ], [ "0.0809", "0.1459", "0.468", "0.1465", "0.468", "0.161", "0.0809", "0.1604" ], [ "0.7256", "0.1483", "0.738", "0.1465", "0.739", "0.1509", "0.726", "0.1528" ], [ "0.822", "0.1459", "0.8486", "0.1477", "0.847", "0.1559", "0.8203", "0.154" ], [ "0.08417", "0.1622", "0.4673", "0.1622", "0.4673", "0.1755", "0.08417", "0.1755" ], [ "0.732", "0.1692", "0.7427", "0.1692", "0.7427", "0.1724", "0.732", "0.1724" ], [ "0.8325", "0.1654", "0.851", "0.1674", "0.8506", "0.1736", "0.831", "0.1718" ], [ "0.08417", "0.1787", "0.4673", "0.1787", "0.4673", "0.1895", "0.08417", "0.1895" ], [ "0.7183", "0.1863", "0.7563", "0.1863", "0.7563", "0.1963", "0.7183", "0.1963" ], [ "0.8286", "0.1849", "0.857", "0.1875", "0.8555", "0.1945", "0.8267", "0.1925" ], [ "0.0825", "0.1925", "0.468", "0.1925", "0.468", "0.2058", "0.0825", "0.2058" ], [ "0.5415", "0.1932", "0.5596", "0.1932", "0.5596", "0.1976", "0.5415", "0.1976" ], [ "0.7183", "0.2058", "0.7534", "0.2084", "0.7524", "0.2184", "0.7173", "0.2166" ], [ "0.822", "0.2058", "0.859", "0.2058", "0.859", "0.2172", "0.822", "0.2172" ], [ "0.0801", "0.2229", "0.468", "0.2216", "0.468", "0.2349", "0.0801", "0.2361" ], [ "0.0809", "0.2084", "0.4656", "0.2084", "0.4656", "0.2216", "0.0809", "0.2216" ], [ "0.513", "0.2229", "0.584", "0.2229", "0.584", "0.2329", "0.513", "0.2329" ], [ "0.6763", "0.2229", "0.746", "0.2229", "0.746", "0.2343", "0.6763", "0.2343" ], [ "0.839", "0.2235", "0.8613", "0.2235", "0.8613", "0.2317", "0.839", "0.2317" ], [ "0.0809", "0.238", "0.4656", "0.238", "0.4656", "0.252", "0.0809", "0.252" ], [ "0.5034", "0.2424", "0.888", "0.2424", "0.888", "0.2568", "0.5034", "0.2568" ], [ "0.08417", "0.2551", "0.4656", "0.2551", "0.4656", "0.2651", "0.08417", "0.2651" ], [ "0.5024", "0.2556", "0.8906", "0.2551", "0.8906", "0.2695", "0.5024", "0.2708" ], [ "0.0825", "0.2683", "0.4673", "0.2683", "0.4673", "0.2822", "0.0825", "0.2822" ], [ "0.505", "0.2727", "0.8896", "0.2727", "0.8896", "0.2827", "0.505", "0.2827" ], [ "0.0825", "0.2827", "0.468", "0.2827", "0.468", "0.2966", "0.0825", "0.2966" ], [ "0.5034", "0.2827", "0.8906", "0.2827", "0.8906", "0.2966", "0.5034", "0.2966" ], [ "0.0825", "0.2974", "0.468", "0.2974", "0.468", "0.3113", "0.0825", "0.3113" ], [ "0.501", "0.2974", "0.8896", "0.2986", "0.8896", "0.3125", "0.501", "0.3113" ], [ "0.0809", "0.3125", "0.4656", "0.3125", "0.4656", "0.3257", "0.0809", "0.3257" ], [ "0.5034", "0.3132", "0.6216", "0.3145", "0.6216", "0.3257", "0.5034", "0.3245" ], [ "0.08417", "0.3276", "0.468", "0.3276", "0.468", "0.3416", "0.08417", "0.3416" ], [ "0.5034", "0.3333", "0.8906", "0.3333", "0.8906", "0.3467", "0.5034", "0.3467" ], [ "0.0825", "0.3435", "0.326", "0.3435", "0.326", "0.3567", "0.0825", "0.3567" ], [ "0.505", "0.3484", "0.7812", "0.3484", "0.7812", "0.3623", "0.505", "0.3623" ], [ "0.08496", "0.3623", "0.273", "0.3655", "0.273", "0.379", "0.08496", "0.3757" ], [ "0.2656", "0.3643", "0.468", "0.3643", "0.468", "0.3774", "0.2656", "0.3774" ], [ "0.5215", "0.3718", "0.888", "0.3706", "0.888", "0.385", "0.5215", "0.3865" ], [ "0.0825", "0.379", "0.4673", "0.379", "0.4673", "0.3926", "0.0825", "0.3926" ], [ "0.5024", "0.385", "0.888", "0.3865", "0.888", "0.401", "0.5024", "0.3997" ], [ "0.0801", "0.3945", "0.468", "0.3933", "0.468", "0.4077", "0.0801", "0.4092" ], [ "0.5063", "0.4016", "0.8906", "0.4016", "0.8906", "0.4153", "0.5063", "0.4153" ], [ "0.0809", "0.4104", "0.4673", "0.4104", "0.4673", "0.4236", "0.0809", "0.4236" ], [ "0.5034", "0.4172", "0.8896", "0.4172", "0.8896", "0.4312", "0.5034", "0.4312" ], [ "0.0825", "0.4248", "0.4673", "0.4248", "0.4673", "0.4382", "0.0825", "0.4382" ], [ "0.5034", "0.4312", "0.8906", "0.432", "0.8906", "0.4458", "0.5034", "0.4443" ], [ "0.0809", "0.44", "0.468", "0.44", "0.468", "0.4539", "0.0809", "0.4539" ], [ "0.505", "0.4463", "0.888", "0.4463", "0.888", "0.4602", "0.505", "0.4602" ], [ "0.0809", "0.4539", "0.468", "0.455", "0.468", "0.4697", "0.0809", "0.4685" ], [ "0.5034", "0.4622", "0.9966", "0.4622", "0.9966", "0.476", "0.5034", "0.476" ], [ "0.0809", "0.4702", "0.468", "0.4702", "0.468", "0.4841", "0.0809", "0.4841" ], [ "0.505", "0.4792", "0.8896", "0.4792", "0.8896", "0.4893", "0.505", "0.4893" ], [ "0.08417", "0.4873", "0.4656", "0.4873", "0.4656", "0.4976", "0.08417", "0.4976" ], [ "0.505", "0.4924", "0.886", "0.4924", "0.886", "0.506", "0.505", "0.506" ], [ "0.08417", "0.502", "0.4673", "0.502", "0.4673", "0.512", "0.08417", "0.512" ], [ "0.0809", "0.513", "0.468", "0.5146", "0.468", "0.529", "0.0809", "0.528" ], [ "0.505", "0.509", "0.8896", "0.509", "0.8896", "0.5195", "0.505", "0.5195" ], [ "0.5034", "0.5215", "0.8906", "0.523", "0.8906", "0.536", "0.5034", "0.535" ], [ "0.08417", "0.532", "0.4673", "0.532", "0.4673", "0.5425", "0.08417", "0.5425" ], [ "0.5034", "0.5386", "0.8906", "0.5386", "0.8906", "0.552", "0.5034", "0.552" ], [ "0.0809", "0.5454", "0.4673", "0.5454", "0.4673", "0.5596", "0.0809", "0.5596" ], [ "0.505", "0.553", "0.8896", "0.553", "0.8896", "0.5664", "0.505", "0.5664" ], [ "0.08417", "0.5625", "0.4673", "0.5625", "0.4673", "0.5728", "0.08417", "0.5728" ], [ "0.5034", "0.5684", "0.8896", "0.5684", "0.8896", "0.582", "0.5034", "0.582" ], [ "0.0801", "0.5757", "0.468", "0.5747", "0.468", "0.589", "0.0801", "0.5903" ], [ "0.505", "0.584", "0.888", "0.584", "0.888", "0.5977", "0.505", "0.5977" ], [ "0.0809", "0.5913", "0.4673", "0.5913", "0.4673", "0.605", "0.0809", "0.605" ], [ "0.505", "0.5986", "0.8896", "0.5986", "0.8896", "0.6123", "0.505", "0.6123" ], [ "0.0809", "0.606", "0.326", "0.606", "0.326", "0.619", "0.0809", "0.619" ], [ "0.5063", "0.6157", "0.6504", "0.6157", "0.6504", "0.6255", "0.5063", "0.6255" ], [ "0.0866", "0.634", "0.468", "0.634", "0.468", "0.648", "0.0866", "0.648" ], [ "0.08417", "0.651", "0.4656", "0.651", "0.4656", "0.661", "0.08417", "0.661" ], [ "0.5034", "0.657", "0.7695", "0.657", "0.7695", "0.6704", "0.5034", "0.6704" ], [ "0.0809", "0.6636", "0.4673", "0.6636", "0.4673", "0.677", "0.0809", "0.677" ], [ "0.0809", "0.6787", "0.468", "0.6787", "0.468", "0.6924", "0.0809", "0.6924" ], [ "0.523", "0.6846", "0.8906", "0.6846", "0.8906", "0.6978", "0.523", "0.6978" ], [ "0.0809", "0.6934", "0.4673", "0.6934", "0.4673", "0.707", "0.0809", "0.707" ], [ "0.505", "0.6987", "0.8896", "0.6987", "0.8896", "0.712", "0.505", "0.712" ], [ "0.0825", "0.7104", "0.4673", "0.7104", "0.4673", "0.7236", "0.0825", "0.7236" ], [ "0.5063", "0.714", "0.892", "0.714", "0.892", "0.728", "0.5063", "0.728" ], [ "0.0825", "0.7246", "0.4673", "0.7246", "0.4673", "0.738", "0.0825", "0.738" ], [ "0.505", "0.729", "0.8906", "0.729", "0.8906", "0.742", "0.505", "0.742" ], [ "0.0801", "0.7393", "0.4673", "0.7407", "0.4673", "0.754", "0.0801", "0.7524" ], [ "0.505", "0.744", "0.8906", "0.744", "0.8906", "0.7583", "0.505", "0.7583" ], [ "0.0825", "0.755", "0.468", "0.755", "0.468", "0.768", "0.0825", "0.768" ], [ "0.5034", "0.7583", "0.8906", "0.759", "0.8906", "0.7734", "0.5034", "0.773" ], [ "0.0801", "0.768", "0.4673", "0.7695", "0.4673", "0.7827", "0.0801", "0.782" ], [ "0.5024", "0.7734", "0.8896", "0.7734", "0.8896", "0.787", "0.5024", "0.787" ], [ "0.0809", "0.784", "0.4673", "0.784", "0.4673", "0.7974", "0.0809", "0.7974" ], [ "0.5034", "0.789", "0.8906", "0.789", "0.8906", "0.803", "0.5034", "0.803" ], [ "0.0825", "0.8003", "0.1772", "0.8003", "0.1772", "0.8125", "0.0825", "0.8125" ], [ "0.5063", "0.8047", "0.8896", "0.8047", "0.8896", "0.8184", "0.5063", "0.8184" ], [ "0.0784", "0.8237", "0.3733", "0.8237", "0.3733", "0.8384", "0.0784", "0.8384" ], [ "0.5063", "0.8213", "0.8896", "0.8213", "0.8896", "0.832", "0.5063", "0.832" ], [ "0.505", "0.8354", "0.8906", "0.8354", "0.8906", "0.8486", "0.505", "0.8486" ], [ "0.10046", "0.8413", "0.468", "0.8413", "0.468", "0.855", "0.10046", "0.855" ], [ "0.0809", "0.856", "0.4673", "0.857", "0.4673", "0.8706", "0.0809", "0.869" ], [ "0.505", "0.851", "0.8896", "0.851", "0.8896", "0.8643", "0.505", "0.8643" ], [ "0.5024", "0.8643", "0.888", "0.866", "0.888", "0.88", "0.5024", "0.8774" ], [ "0.08496", "0.8735", "0.4673", "0.8735", "0.4673", "0.884", "0.08496", "0.884" ], [ "0.5034", "0.879", "0.8896", "0.88", "0.8896", "0.8945", "0.5034", "0.893" ], [ "0.0809", "0.887", "0.4673", "0.887", "0.4673", "0.901", "0.0809", "0.901" ], [ "0.5063", "0.8965", "0.585", "0.8965", "0.585", "0.9077", "0.5063", "0.9077" ] ]

[ "2.Related Work", "Joint-Objective L", "nape-specfe ousj", "mage-Text Loss", "(Text-Spel", "Visual Description.", "多年来，该领域已经看到了许", "多不同的方法[27，11，181，最近，深度模型因其性能", "", "和端到端培训的潜力而受到欢迎。深度视觉描述框架", "", "Waon", "首先将图像编码为固定长度的特征向量，然后通过在", "nhe", "Embe", "图像特征「2，8.26上调节文本生成或在预测下一个单", "LSTM", "ISTM", "词之前将图像特征和先前生成的单词嵌人到多模式空", "NN", "Wglove", "Wglove", "来自卷积神经网络的中间表示的图像（例如来自CNN", "间[9，10，15]中来生成描述。尽管大多数模型表示具有", "IM.LGENET", "?MSCOCO", "", "的fc7激活），但其他模型将图像表示为固定数量的视", "Figure 2. Our NOC image caption network. During train", "觉概念3，7的置信度向量。在儿乎所有情况下，可视", "ing, the visual recognition network (left), the LSTM-based", "化管道的参数都使用在ImageNet分类任务上训练的", "language model (right),and the caption model (center)are", "权重进行初始化。对于字幕生成，递归网络（RNN）是", "trained simultaneously on different sources with different", "建模语言的流行选择，但是对数双线性模型9和最大", "objectives but with shared parameters, thus enabling novel", "熵语言模型31也已被探索。。我们的模型类似于[715", "object captioning.", "中的CNN-RNN框架，但这些模型都不能端到端训练", "件，以创建一个可视化描述模型，该模型同时学习一个", "来描述图像字幕对中看不见的对象", "独立的对象识别模型以及一个语言模型", "Novel object captioning.", "「16]提出了一种方法，该", "我们首先训练基于LSTM的语言模型（LM）[24", "方法扩展了模型描述一小部分新颖概念的能力（例如", "以生成句子。我们的LM结合了在外部文本语料库", "g。quidditch，samisen）来自几个成对训练示例，同", "上预先训练的分布式嵌入（GloVe，[20]）中的单词", "时保留其描述以前学习的概念的能力。另一方面，「7]", "的密集表示。同时，我们还训练了最先进的视觉识别", "引人了一个模型，该模型可以描述英语语料库和对象", "网络，以在给定图像的情况下对词汇中的单词提供信", "识别数据集（ImageNet）中已经存在的许多对象，但", "心。Thisdecomposesourmodelintodiscretetextandvisu-", "不能描述字幕语料库（e。g。pheasant，otter）。我们", "alpipelineswhichcanbetrainedexclusivelyusingunpairedtextandunpaire", "的重点是后一种情况。[7]集成了来自外部文本和视觉", "2）。为了生成以图像内容为条件的描述，我们通过对", "源的信息，并将参数从图像字幕数据中看到的对象显", "单词词汇的文本和视觉置信度进行求和（元素方式）", "式传输（“复制”）到看不见的ImageNet对象，以对这", "些新颖的类别进行字幕。虽然这适用于许多ImageNet", "将语言和视觉识别网络的预测结合起来。在训练过程", "中，我们引入了辅助图像特定（CIM）和文本特定", "类：但它仍然限制了不同类别的覆盖范围，并且无法进", "（LcM）目标以及成对的图像字幕（LcM）损失。这", "行端到端的培训。此外，他们的模型不能为已经存在很", "些损失函数在联合训练时会影响我们的模型，不仅产", "少配对训练示例的对象加上标题。我们提出的框架隐", "生合理的图像描述，而且还预测视觉概念并生成内聚", "式地集成了分布式语义嵌人，避免了任何显式传输的", "文本（语言建模）。。我们首先讨论辅助目标和联合训", "需要，并使其端到端可训练。它还直接扩展到具有很少", "练，然后讨论如何利用用外部文本训练的嵌人来编写", "或没有描述的标题ImageNet对象", "关于新颖对象的描述", "Multi-modalandZero-ShotLearning.另一个密切", "相关的研究领域利用分布语义来学习使用视觉和文本", "3.1. Auxiliary Training Objectives", "信息的联合嵌入空间，以对新颖的对象类别「4，19]进", "行零拍标记，以及使用文本「12，23]检索图像。视觉描", "我们引人辅助目标的动机是学习如何描述图像而不", "述本身可以转换为一个多模式学习问题，其中在生成标", "会失去识别更多对象的能力。通常，图像字幕模型结合", "题中的下一个单词Wn之前，将标题单词Wo,.,wn-1", "了在源域上预先训练的视觉分类器（例如。g。ImageNet", "和图像投影到联合嵌人空间中[10，15]。尽管我们的方", "数据集），然后将其调整到目标域（图像字幕数据集）。", "法使用分布式单词嵌入，但我们的模型在以下方面有", "但是，如果微调时不存在类似信息，则可以抑制源数据", "所不同：它可以使用不成对的文本和视觉数据进行训", "集中的重要信息，从而导致网络忘记（覆盖权重）目标", "练，但仍可以在字幕生成的后期阶段结合语义信息。这", "域中不存在的对象。这在我们的场景中是有问题的，在", "在精神上类似于使用单语数据改进机器翻译5]的自然", "这个场景中，模型依赖于源数据集来学习目标数据集", "语言处理工作", "中不存在的大量视觉概念。但是，通过预训练以及互补", "3. Novel Object Captioner (NOC)", "的辅助目标，模型保持了识别更多对象的能力，并被鼓", "励在测试时描述目标数据集中不存在的对象。为了便", "我们的NOC模型如图所示。2。它由一个语言模", "型组成，该模型利用在未注释文本上训练的分布式语", "于说明，我们抽象了语言和视觉模型的细节，并首先描", "述了完整模型的联合训练目标，即。e。文本特定丢失", "义嵌入：并将其与视觉识别模型集成。我们引入辅助损", "图像特定丢失和图像字幕丢失。然后我们将描述语言", "失函数（目标），并在多个数据源上联合训练不同的组", "和视觉模型" ]

{ "arxiv_id": "1606.07770", "language": "zh", "page_idx": 1 }

[ [ "0.0801", "0.09344", "0.2737", "0.09344", "0.2737", "0.1067", "0.0801", "0.1067" ], [ "0.609", "0.09216", "0.7876", "0.09216", "0.7876", "0.1061", "0.609", "0.1061" ], [ "0.869", "0.09344", "0.8906", "0.09344", "0.8906", "0.1067", "0.869", "0.1067" ], [ "0.0809", "0.11237", "0.4673", "0.11237", "0.4673", "0.1256", "0.0809", "0.1256" ], [ "0.505", "0.1193", "0.8896", "0.1193", "0.8896", "0.13", "0.505", "0.13" ], [ "0.0809", "0.1276", "0.468", "0.1276", "0.468", "0.1414", "0.0809", "0.1414" ], [ "0.5024", "0.1332", "0.8906", "0.1332", "0.8906", "0.1465", "0.5024", "0.1465" ], [ "0.08417", "0.1445", "0.4673", "0.1445", "0.4673", "0.1547", "0.08417", "0.1547" ], [ "0.5024", "0.1477", "0.8906", "0.1483", "0.8906", "0.1622", "0.5024", "0.161" ], [ "0.0825", "0.1572", "0.4673", "0.1572", "0.4673", "0.1704", "0.0825", "0.1704" ], [ "0.505", "0.1635", "0.8906", "0.1635", "0.8906", "0.1768", "0.505", "0.1768" ], [ "0.0825", "0.1749", "0.4673", "0.1749", "0.4673", "0.1849", "0.0825", "0.1849" ], [ "0.5034", "0.1787", "0.6733", "0.1787", "0.6733", "0.1925", "0.5034", "0.1925" ], [ "0.0801", "0.1881", "0.468", "0.1875", "0.468", "0.202", "0.0801", "0.2026" ], [ "0.0809", "0.2052", "0.4673", "0.2052", "0.4673", "0.2153", "0.0809", "0.2153" ], [ "0.501", "0.1995", "0.8857", "0.202", "0.8857", "0.2166", "0.501", "0.214" ], [ "0.0801", "0.2184", "0.4673", "0.2172", "0.4673", "0.2311", "0.0801", "0.2317" ], [ "0.524", "0.2229", "0.8896", "0.2229", "0.8896", "0.2329", "0.524", "0.2329" ], [ "0.0825", "0.2349", "0.2712", "0.2349", "0.2712", "0.2456", "0.0825", "0.2456" ], [ "0.5034", "0.2361", "0.8896", "0.2361", "0.8896", "0.2507", "0.5034", "0.2507" ], [ "0.09235", "0.252", "0.4084", "0.252", "0.4084", "0.2727", "0.09235", "0.2727" ], [ "0.505", "0.252", "0.888", "0.252", "0.888", "0.2651", "0.505", "0.2651" ], [ "0.21", "0.2695", "0.2336", "0.2695", "0.2336", "0.281", "0.21", "0.281" ], [ "0.505", "0.2683", "0.8906", "0.2683", "0.8906", "0.2822", "0.505", "0.2822" ], [ "0.4495", "0.2747", "0.4656", "0.2747", "0.4656", "0.2842", "0.4495", "0.2842" ], [ "0.1765", "0.2886", "0.423", "0.2886", "0.423", "0.303", "0.1765", "0.303" ], [ "0.505", "0.2827", "0.888", "0.2827", "0.888", "0.2966", "0.505", "0.2966" ], [ "0.505", "0.2986", "0.8896", "0.2986", "0.8896", "0.3125", "0.505", "0.3125" ], [ "0.0809", "0.318", "0.4656", "0.318", "0.4656", "0.332", "0.0809", "0.332" ], [ "0.505", "0.3132", "0.8906", "0.3132", "0.8906", "0.327", "0.505", "0.327" ], [ "0.0809", "0.334", "0.2075", "0.334", "0.2075", "0.348", "0.0809", "0.348" ], [ "0.505", "0.3289", "0.8813", "0.3289", "0.8813", "0.3423", "0.505", "0.3423" ], [ "0.5034", "0.3435", "0.8906", "0.3435", "0.8906", "0.3567", "0.5034", "0.3567" ], [ "0.0801", "0.3547", "0.2646", "0.3547", "0.2646", "0.3687", "0.0801", "0.3687" ], [ "0.505", "0.3591", "0.8896", "0.3591", "0.8896", "0.3738", "0.505", "0.3738" ], [ "0.0809", "0.375", "0.4673", "0.375", "0.4673", "0.3882", "0.0809", "0.3882" ], [ "0.5073", "0.377", "0.888", "0.377", "0.888", "0.387", "0.5073", "0.387" ], [ "0.0801", "0.3882", "0.468", "0.3894", "0.468", "0.404", "0.0801", "0.4028" ], [ "0.505", "0.3914", "0.8896", "0.3914", "0.8896", "0.4016", "0.505", "0.4016" ], [ "0.0809", "0.4048", "0.468", "0.4048", "0.468", "0.4185", "0.0809", "0.4185" ], [ "0.5034", "0.4048", "0.8906", "0.4048", "0.8906", "0.4185", "0.5034", "0.4185" ], [ "0.0801", "0.4192", "0.468", "0.4185", "0.468", "0.433", "0.0801", "0.4343" ], [ "0.505", "0.4216", "0.888", "0.4216", "0.888", "0.432", "0.505", "0.432" ], [ "0.0825", "0.435", "0.4126", "0.435", "0.4126", "0.449", "0.0825", "0.449" ], [ "0.505", "0.435", "0.888", "0.435", "0.888", "0.449", "0.505", "0.449" ], [ "0.1323", "0.4558", "0.3064", "0.4558", "0.3064", "0.4702", "0.1323", "0.4702" ], [ "0.505", "0.4507", "0.8906", "0.4507", "0.8906", "0.464", "0.505", "0.464" ], [ "0.505", "0.4653", "0.7563", "0.4653", "0.7563", "0.4792", "0.505", "0.4792" ], [ "0.1478", "0.473", "0.4192", "0.4749", "0.4192", "0.4924", "0.1478", "0.4905" ], [ "0.4453", "0.4734", "0.4673", "0.4734", "0.4673", "0.486", "0.4453", "0.486" ], [ "0.1381", "0.4797", "0.1503", "0.476", "0.1544", "0.4841", "0.1422", "0.488" ], [ "0.5024", "0.4849", "0.665", "0.4849", "0.665", "0.4988", "0.5024", "0.4988" ], [ "0.0784", "0.5083", "0.2754", "0.507", "0.2754", "0.5215", "0.0784", "0.523" ], [ "0.523", "0.506", "0.8906", "0.506", "0.8906", "0.5195", "0.523", "0.5195" ], [ "0.5034", "0.5195", "0.8906", "0.5195", "0.8906", "0.5327", "0.5034", "0.5327" ], [ "0.0825", "0.528", "0.4673", "0.528", "0.4673", "0.5415", "0.0825", "0.5415" ], [ "0.5034", "0.535", "0.8896", "0.535", "0.8896", "0.549", "0.5034", "0.549" ], [ "0.0825", "0.5435", "0.468", "0.5435", "0.468", "0.5566", "0.0825", "0.5566" ], [ "0.5024", "0.549", "0.8906", "0.55", "0.8906", "0.5645", "0.5024", "0.563" ], [ "0.0809", "0.5566", "0.4656", "0.5566", "0.4656", "0.5713", "0.0809", "0.5713" ], [ "0.5034", "0.565", "0.7983", "0.565", "0.7983", "0.579", "0.5034", "0.579" ], [ "0.0801", "0.5737", "0.468", "0.5737", "0.468", "0.587", "0.0801", "0.587" ], [ "0.524", "0.5806", "0.8896", "0.5806", "0.8896", "0.5947", "0.524", "0.5947" ], [ "0.0809", "0.5884", "0.4673", "0.5884", "0.4673", "0.6016", "0.0809", "0.6016" ], [ "0.5034", "0.595", "0.8906", "0.595", "0.8906", "0.6094", "0.5034", "0.6094" ], [ "0.0784", "0.6016", "0.4697", "0.601", "0.4697", "0.6187", "0.0784", "0.619" ], [ "0.505", "0.61", "0.888", "0.61", "0.888", "0.6235", "0.505", "0.6235" ], [ "0.08417", "0.6206", "0.468", "0.6206", "0.468", "0.631", "0.08417", "0.631" ], [ "0.505", "0.6255", "0.7173", "0.6255", "0.7173", "0.6396", "0.505", "0.6396" ], [ "0.0801", "0.6333", "0.2263", "0.634", "0.2263", "0.6484", "0.0801", "0.648" ], [ "0.501", "0.646", "0.607", "0.648", "0.607", "0.661", "0.501", "0.659" ], [ "0.1095", "0.6577", "0.3325", "0.659", "0.3325", "0.674", "0.1095", "0.6724" ], [ "0.4468", "0.6685", "0.4673", "0.6685", "0.4673", "0.681", "0.4468", "0.681" ], [ "0.523", "0.6685", "0.8896", "0.6685", "0.8896", "0.6816", "0.523", "0.6816" ], [ "0.1095", "0.675", "0.4182", "0.6753", "0.4182", "0.69", "0.1095", "0.6895" ], [ "0.5034", "0.6816", "0.714", "0.683", "0.714", "0.6963", "0.5034", "0.696" ], [ "0.0825", "0.704", "0.4673", "0.704", "0.4673", "0.714", "0.0825", "0.714" ], [ "0.5024", "0.702", "0.8145", "0.7026", "0.8145", "0.7173", "0.5024", "0.7163" ], [ "0.0809", "0.7173", "0.4673", "0.7173", "0.4673", "0.731", "0.0809", "0.731" ], [ "0.524", "0.7197", "0.8906", "0.7197", "0.8906", "0.733", "0.524", "0.733" ], [ "0.0809", "0.733", "0.468", "0.733", "0.468", "0.747", "0.0809", "0.747" ], [ "0.5024", "0.733", "0.8896", "0.7344", "0.8896", "0.7476", "0.5024", "0.747" ], [ "0.0825", "0.7476", "0.1691", "0.7476", "0.1691", "0.759", "0.0825", "0.759" ], [ "0.505", "0.7485", "0.8906", "0.7485", "0.8906", "0.762", "0.505", "0.762" ], [ "0.0801", "0.762", "0.398", "0.762", "0.398", "0.776", "0.0801", "0.776" ], [ "0.505", "0.7646", "0.8906", "0.7646", "0.8906", "0.778", "0.505", "0.778" ], [ "0.0809", "0.782", "0.468", "0.782", "0.468", "0.7954", "0.0809", "0.7954" ], [ "0.505", "0.779", "0.8906", "0.779", "0.8906", "0.7925", "0.505", "0.7925" ], [ "0.0809", "0.797", "0.4673", "0.797", "0.4673", "0.81", "0.0809", "0.81" ], [ "0.505", "0.7944", "0.8906", "0.7944", "0.8906", "0.808", "0.505", "0.808" ], [ "0.0809", "0.811", "0.4673", "0.811", "0.4673", "0.8247", "0.0809", "0.8247" ], [ "0.5063", "0.811", "0.8896", "0.811", "0.8896", "0.8213", "0.5063", "0.8213" ], [ "0.0809", "0.827", "0.4673", "0.827", "0.4673", "0.8403", "0.0809", "0.8403" ], [ "0.5034", "0.8247", "0.8906", "0.8247", "0.8906", "0.8384", "0.5034", "0.8384" ], [ "0.08417", "0.8433", "0.4673", "0.8433", "0.4673", "0.854", "0.08417", "0.854" ], [ "0.5024", "0.8384", "0.6724", "0.8403", "0.6724", "0.854", "0.5024", "0.8516" ], [ "0.0809", "0.857", "0.468", "0.857", "0.468", "0.8706", "0.0809", "0.8706" ], [ "0.5024", "0.855", "0.6953", "0.856", "0.6953", "0.8706", "0.5024", "0.869" ], [ "0.0801", "0.8706", "0.4673", "0.872", "0.4673", "0.886", "0.0801", "0.8853" ], [ "0.524", "0.8726", "0.8906", "0.8726", "0.8906", "0.886", "0.524", "0.886" ], [ "0.0809", "0.888", "0.1054", "0.888", "0.1054", "0.901", "0.0809", "0.901" ], [ "0.505", "0.887", "0.8906", "0.887", "0.8906", "0.901", "0.505", "0.901" ] ]

[ "3.1.1 Image-specific Loss", "C=LCM+CIM+CCM", "(4)", "我们的视觉识别模型（图。E2，左）是由01参数化的", "通过与图像网络和语言网络共享字幕模型网络的权重", "神经网络，并在对象识别数据集上进行了训练。与在", "（如图。2(a))，该模型可以同时在独立的纯图像数据、", "分类任务上使用单个标签训练的典型视觉识别模型不", "未注释的文本数据以及成对图像-字幕数据上进行训练。", "同，对于图像字幕的任务，对同时出现在图像中的多个", "因此，共同优化不同的目标有助于模型识别成对的图", "视觉概念具有高置信度的图像模型将是优选的。因此，", "像-句子数据之外的类别", "我们选择使用多个标签来训练我们的模型（更多信息在", "Sec。5.1），具有多标签丢失。如果表示标签，而", "3.2. Language Model with Semantic Embeddings", "么表示标签的二进制地面真值，则视觉模型的目标由", "我们的语言模型由以下组件组成：单词的连续低维", "交叉摘损失（LIM）给出：", "嵌人空间（Wgloue），单个递归（LSTM）隐藏层以及", "CM(I;01) = -Zz log(S_(frM(;0r)", "两个线性变换层，其中第二层（Wtlone映射）。最后", "1", "在输出层上使用softmax激活函数来产生归一化概率", "(1)", "+(1 -z) log(1 - S_l(frM(I;0n)", "分布。将等价于最大似然的交叉摘损失作为训练目标", "除了我们的共同目标（Eqn。4），我们还在语言模型中", "其中S_i（α）是XMATHMAX层的最终索引", "使用语义嵌人来帮助在描述新颖对象时生成句子。具", "XMATH5的输出", "体来说，初始输人嵌人空间（Wgloue）用于将输人（", "热）单词表示为语义有意义的密集固定长度向量。虽然", "3.1.2Text-specific Loss", "最终的转换层（Wqloue）在softmax激活函数的帮助", "我们的语言模型（图。2，右）基于LSTM递归神经", "下将密集向量的映射15，25反转回完整词汇表。这些", "网络。我们用QL表示该网络的参数，用fLM表示该", "分布嵌入17，20共享语义相似的单词具有相似的向量", "网络最后一层的激活。训练语言模型以预测给定单词", "表示的属性。在输人和输出转换层中使用这些嵌入的", "序列wo,…,wt-1中的下一个单词wt。这是使用相当", "直观原因是为了帮助语言模型将图像文本语料库中看", "于最大似然的softmax损失LcM进行优化的：", "不见的单词处理为（在语义上）以前见过的相似单词", "CcM(wo,...Wt-i;OL) =", "从而鼓励合成句子生成i。e。根据视觉置信度，鼓励它", "在句子描述中使用新的/稀有的单词", "log(S_wt(fLM(wo, ., wt-1; OL)", "(2)", "", "3.3. Visual Classifier", "3.1.3Image-caption Loss", "我们模型的另一个主要组件是视觉分类器。与以前", "的工作「7相同，我们采用VGG-16E22]卷积网络作", "图像字幕模型的目标（图。2,center）是生成以图像", "为视觉识别网络。我们修改网络的最后一层，以合并多", "为条件的句子（I）。NOC预测序列中的下一个单词", "标签丢失（Eqn。1）预测完整词汇表中多个标签的视", "wt，该单词以先前生成的单词（wo,….,wt-1）和图像", "觉置信度。分类网络的其余部分保持不变", "（I）为条件，通过对来自深层语言模型的激活进行求", "最后，我们对视觉和语言输出进行元素求和，可以", "和，该模型对先前的单词进行操作，以及对图像进行", "将其视为语言模型在单词上产生平滑的概率分布（基", "操作的深层图像模型。我们用fcM表示这些最终（求", "于GloVe参数共享），然后根据视觉证据在其中“选择”", "和）激活。然后，预测下一个单词的概率由下式给出", "图像信号与语言模型信念相加", "P(wt|wo,.., wt-1, I)", "4. Datasets", "=S_wt(fcM(wo,.,wt-1,I;0))", "(3)", "在本节中，我们描述了图像描述数据集以及实验中", "=S_wt(fLM(wo,..,Wt-1;0L)+fIM(I;O))", "使用的外部文本和图像数据集", "给定图像和描述对，学幕模型通过最小化字幕模型", "4.1. External Text Corpus (WebCorpus)", "损失CcM来优化基础语言模型（OL）和图像模型", "我们从Gigaword，英国国家语料库（BNC），UkWaC", "(0）的参数：LcM(Wo,,Wt-1,I;0L,0)<NEXMATH-", "和Wikipedia中提取句子。斯坦福CoreNLP3。4。2[14]", "LINE>XXX", "用于提取标记。此数据集用于训练LSTM语言模型。", "3.1.4Joint Training with Auxiliary Losses", "对于网络中密集的单词表示，我们使用GloVe[20]在", "虽然许多以前的方法已经通过预训练图像和语言模型", "包括Gigaword和Wikipedia在内的外部语料库的6B", "并单独调整字幕模型而在图像字幕上取得了成功（Eqn。", "令牌上预先训练。为了创建我们的LM词汇表，我们", "？？），这不足以为图像字幕数据集之外的对象生成描述，", "从合并的外部语料库中确定了80.000个最常见的标记。", "因为模型往往会“忘记”（覆盖权重）仅在外部数据源", "我们将这个词汇进一步细化为一组72,700个单词，这", "中看到的对象。为了解决这个问题，我们建议在不同", "些单词也有GloVe嵌入", "的数据源上同时训练图像模型，语言模型和字幕模型。", "4.2. Image Caption data", "NOC模型的最终目标同时最小化了三个单独的互补目", "为了从经验上评估NOC描述新对象的能力，我们", "标：", "使用[7]的训练和测试集。此数据集是根据MSCOCO" ]

{ "arxiv_id": "1606.07770", "language": "zh", "page_idx": 2 }

[ [ "0.0809", "0.09216", "0.4673", "0.09344", "0.4673", "0.1067", "0.0809", "0.1061" ], [ "0.576", "0.0903", "0.6045", "0.0903", "0.6045", "0.09784", "0.576", "0.09784" ], [ "0.569", "0.1029", "0.576", "0.1029", "0.576", "0.1061", "0.569", "0.1061" ], [ "0.5767", "0.1042", "0.621", "0.1042", "0.621", "0.108", "0.5767", "0.108" ], [ "0.6357", "0.1029", "0.6665", "0.1029", "0.6665", "0.10986", "0.6357", "0.10986" ], [ "0.694", "0.1042", "0.721", "0.1042", "0.721", "0.108", "0.694", "0.108" ], [ "0.721", "0.1042", "0.738", "0.1042", "0.738", "0.108", "0.721", "0.108" ], [ "0.5767", "0.113", "0.615", "0.113", "0.615", "0.1206", "0.5767", "0.1206" ], [ "0.774", "0.113", "0.846", "0.113", "0.846", "0.1206", "0.774", "0.1206" ], [ "0.6284", "0.11554", "0.6406", "0.11554", "0.6406", "0.1187", "0.6284", "0.1187" ], [ "0.6406", "0.11554", "0.706", "0.11554", "0.706", "0.1193", "0.6406", "0.1193" ], [ "0.08417", "0.125", "0.4656", "0.125", "0.4656", "0.1351", "0.08417", "0.1351" ], [ "0.0809", "0.1383", "0.4673", "0.1383", "0.4673", "0.1515", "0.0809", "0.1515" ], [ "0.5786", "0.1395", "0.591", "0.1395", "0.591", "0.1477", "0.5786", "0.1477" ], [ "0.576", "0.1509", "0.615", "0.1528", "0.615", "0.1604", "0.576", "0.1578" ], [ "0.08417", "0.1547", "0.4673", "0.1547", "0.4673", "0.1654", "0.08417", "0.1654" ], [ "0.0825", "0.108", "0.4673", "0.108", "0.4673", "0.1218", "0.0825", "0.1218" ], [ "0.08417", "0.1692", "0.4656", "0.1692", "0.4656", "0.1799", "0.08417", "0.1799" ], [ "0.5786", "0.1635", "0.6094", "0.1635", "0.6094", "0.1704", "0.5786", "0.1704" ], [ "0.0825", "0.1831", "0.4673", "0.1831", "0.4673", "0.1963", "0.0825", "0.1963" ], [ "0.576", "0.1881", "0.603", "0.1881", "0.603", "0.1957", "0.576", "0.1957" ], [ "0.0809", "0.1989", "0.1936", "0.1989", "0.1936", "0.2122", "0.0809", "0.2122" ], [ "0.5728", "0.1989", "0.739", "0.1989", "0.739", "0.209", "0.5728", "0.209" ], [ "0.7363", "0.2008", "0.8545", "0.2008", "0.8545", "0.2084", "0.7363", "0.2084" ], [ "0.576", "0.2102", "0.876", "0.2102", "0.876", "0.2203", "0.576", "0.2203" ], [ "0.0809", "0.2166", "0.2075", "0.2166", "0.2075", "0.2299", "0.0809", "0.2299" ], [ "0.10046", "0.2317", "0.4673", "0.2317", "0.4673", "0.2456", "0.10046", "0.2456" ], [ "0.5728", "0.2349", "0.614", "0.2349", "0.614", "0.2456", "0.5728", "0.2456" ], [ "0.0801", "0.2468", "0.4673", "0.2456", "0.4673", "0.26", "0.0801", "0.2615" ], [ "0.572", "0.2443", "0.846", "0.2468", "0.846", "0.2615", "0.572", "0.2588" ], [ "0.572", "0.2568", "0.7295", "0.2583", "0.7295", "0.2683", "0.572", "0.2676" ], [ "0.7363", "0.26", "0.884", "0.26", "0.884", "0.2676", "0.7363", "0.2676" ], [ "0.0825", "0.2778", "0.468", "0.2778", "0.468", "0.2917", "0.0825", "0.2917" ], [ "0.0825", "0.262", "0.468", "0.262", "0.468", "0.276", "0.0825", "0.276" ], [ "0.5034", "0.2822", "0.584", "0.2822", "0.584", "0.2935", "0.5034", "0.2935" ], [ "0.576", "0.2798", "0.888", "0.2822", "0.888", "0.2954", "0.576", "0.2935" ], [ "0.0825", "0.2935", "0.468", "0.2935", "0.468", "0.307", "0.0825", "0.307" ], [ "0.5034", "0.2954", "0.8896", "0.2954", "0.8896", "0.3093", "0.5034", "0.3093" ], [ "0.0825", "0.308", "0.468", "0.308", "0.468", "0.3225", "0.0825", "0.3225" ], [ "0.505", "0.31", "0.5728", "0.31", "0.5728", "0.3208", "0.505", "0.3208" ], [ "0.0825", "0.3213", "0.4673", "0.3213", "0.4673", "0.336", "0.0825", "0.336" ], [ "0.505", "0.3289", "0.8896", "0.3289", "0.8896", "0.3423", "0.505", "0.3423" ], [ "0.0825", "0.3384", "0.468", "0.3384", "0.468", "0.3516", "0.0825", "0.3516" ], [ "0.5063", "0.3452", "0.876", "0.3452", "0.876", "0.356", "0.5063", "0.356" ], [ "0.0809", "0.3528", "0.468", "0.3528", "0.468", "0.3662", "0.0809", "0.3662" ], [ "0.0809", "0.3662", "0.4673", "0.3674", "0.4673", "0.3818", "0.0809", "0.3806" ], [ "0.5034", "0.3662", "0.7017", "0.3662", "0.7017", "0.3801", "0.5034", "0.3801" ], [ "0.0825", "0.3838", "0.3325", "0.3838", "0.3325", "0.3977", "0.0825", "0.3977" ], [ "0.524", "0.385", "0.8896", "0.385", "0.8896", "0.3984", "0.524", "0.3984" ], [ "0.0784", "0.4048", "0.3342", "0.404", "0.3342", "0.4185", "0.0784", "0.4192" ], [ "0.505", "0.401", "0.8906", "0.401", "0.8906", "0.414", "0.505", "0.414" ], [ "0.5063", "0.4172", "0.888", "0.4172", "0.888", "0.428", "0.5063", "0.428" ], [ "0.1013", "0.4268", "0.468", "0.4268", "0.468", "0.44", "0.1013", "0.44" ], [ "0.5034", "0.4312", "0.8896", "0.4312", "0.8896", "0.4443", "0.5034", "0.4443" ], [ "0.0825", "0.4426", "0.4656", "0.4426", "0.4656", "0.4558", "0.0825", "0.4558" ], [ "0.5034", "0.4458", "0.8906", "0.4458", "0.8906", "0.459", "0.5034", "0.459" ], [ "0.0801", "0.4558", "0.4656", "0.457", "0.4656", "0.4702", "0.0801", "0.4697" ], [ "0.5063", "0.4622", "0.886", "0.4622", "0.886", "0.473", "0.5063", "0.473" ], [ "0.0809", "0.473", "0.468", "0.473", "0.468", "0.486", "0.0809", "0.486" ], [ "0.5034", "0.476", "0.8906", "0.476", "0.8906", "0.4893", "0.5034", "0.4893" ], [ "0.0825", "0.4873", "0.4673", "0.4873", "0.4673", "0.5005", "0.0825", "0.5005" ], [ "0.505", "0.4912", "0.8906", "0.4912", "0.8906", "0.505", "0.505", "0.505" ], [ "0.0825", "0.504", "0.4673", "0.504", "0.4673", "0.5146", "0.0825", "0.5146" ], [ "0.5034", "0.506", "0.892", "0.506", "0.892", "0.5195", "0.5034", "0.5195" ], [ "0.0809", "0.5166", "0.4673", "0.518", "0.4673", "0.532", "0.0809", "0.53" ], [ "0.505", "0.521", "0.8906", "0.521", "0.8906", "0.535", "0.505", "0.535" ], [ "0.0825", "0.5327", "0.468", "0.5327", "0.468", "0.547", "0.0825", "0.547" ], [ "0.505", "0.536", "0.8906", "0.536", "0.8906", "0.55", "0.505", "0.55" ], [ "0.0809", "0.547", "0.468", "0.547", "0.468", "0.561", "0.0809", "0.561" ], [ "0.5024", "0.55", "0.892", "0.5503", "0.892", "0.565", "0.5024", "0.5645" ], [ "0.0809", "0.563", "0.3896", "0.563", "0.3896", "0.577", "0.0809", "0.577" ], [ "0.505", "0.565", "0.6396", "0.565", "0.6396", "0.579", "0.505", "0.579" ], [ "0.0809", "0.5835", "0.3906", "0.5835", "0.3906", "0.5967", "0.0809", "0.5967" ], [ "0.524", "0.582", "0.8906", "0.582", "0.8906", "0.595", "0.524", "0.595" ], [ "0.1013", "0.601", "0.468", "0.601", "0.468", "0.6143", "0.1013", "0.6143" ], [ "0.5024", "0.595", "0.8906", "0.5967", "0.8906", "0.6113", "0.5024", "0.61" ], [ "0.0825", "0.616", "0.4673", "0.616", "0.4673", "0.627", "0.0825", "0.627" ], [ "0.505", "0.6123", "0.8906", "0.6123", "0.8906", "0.6255", "0.505", "0.6255" ], [ "0.0809", "0.63", "0.4673", "0.63", "0.4673", "0.643", "0.0809", "0.643" ], [ "0.505", "0.627", "0.8906", "0.627", "0.8906", "0.64", "0.505", "0.64" ], [ "0.5034", "0.64", "0.888", "0.6416", "0.888", "0.656", "0.5034", "0.655" ], [ "0.0825", "0.6465", "0.4673", "0.6465", "0.4673", "0.657", "0.0825", "0.657" ], [ "0.08417", "0.662", "0.4673", "0.662", "0.4673", "0.6724", "0.08417", "0.6724" ], [ "0.5063", "0.659", "0.8896", "0.659", "0.8896", "0.6694", "0.5063", "0.6694" ], [ "0.0825", "0.6753", "0.4673", "0.6753", "0.4673", "0.6895", "0.0825", "0.6895" ], [ "0.5024", "0.674", "0.5215", "0.674", "0.5215", "0.686", "0.5024", "0.686" ], [ "0.0809", "0.6914", "0.4673", "0.6914", "0.4673", "0.705", "0.0809", "0.705" ], [ "0.5024", "0.6978", "0.6167", "0.6978", "0.6167", "0.711", "0.5024", "0.711" ], [ "0.08417", "0.7056", "0.468", "0.7056", "0.468", "0.7197", "0.08417", "0.7197" ], [ "0.0809", "0.72", "0.4673", "0.72", "0.4673", "0.7344", "0.0809", "0.7344" ], [ "0.523", "0.7217", "0.8906", "0.7217", "0.8906", "0.735", "0.523", "0.735" ], [ "0.0825", "0.7363", "0.468", "0.7363", "0.468", "0.75", "0.0825", "0.75" ], [ "0.5034", "0.7363", "0.8906", "0.7363", "0.8906", "0.75", "0.5034", "0.75" ], [ "0.0825", "0.7505", "0.4656", "0.7505", "0.4656", "0.7646", "0.0825", "0.7646" ], [ "0.5034", "0.752", "0.8906", "0.752", "0.8906", "0.765", "0.5034", "0.765" ], [ "0.0809", "0.765", "0.468", "0.7676", "0.468", "0.781", "0.0809", "0.779" ], [ "0.5034", "0.7666", "0.8896", "0.7676", "0.8896", "0.781", "0.5034", "0.78" ], [ "0.0809", "0.78", "0.468", "0.781", "0.468", "0.7944", "0.0809", "0.7935" ], [ "0.5034", "0.782", "0.8906", "0.782", "0.8906", "0.7954", "0.5034", "0.7954" ], [ "0.0825", "0.797", "0.468", "0.797", "0.468", "0.81", "0.0825", "0.81" ], [ "0.5063", "0.799", "0.8906", "0.799", "0.8906", "0.8096", "0.5063", "0.8096" ], [ "0.08417", "0.813", "0.4673", "0.813", "0.4673", "0.8237", "0.08417", "0.8237" ], [ "0.5034", "0.811", "0.8906", "0.811", "0.8906", "0.8247", "0.5034", "0.8247" ], [ "0.0825", "0.827", "0.4395", "0.827", "0.4395", "0.8403", "0.0825", "0.8403" ], [ "0.505", "0.827", "0.8906", "0.827", "0.8906", "0.8403", "0.505", "0.8403" ], [ "0.10046", "0.8413", "0.468", "0.8413", "0.468", "0.855", "0.10046", "0.855" ], [ "0.505", "0.8413", "0.8896", "0.8413", "0.8896", "0.855", "0.505", "0.855" ], [ "0.0825", "0.857", "0.4656", "0.857", "0.4656", "0.8706", "0.0825", "0.8706" ], [ "0.5034", "0.857", "0.8896", "0.857", "0.8896", "0.8706", "0.5034", "0.8706" ], [ "0.0809", "0.872", "0.468", "0.872", "0.468", "0.886", "0.0809", "0.886" ], [ "0.505", "0.872", "0.888", "0.872", "0.888", "0.886", "0.505", "0.886" ], [ "0.0809", "0.887", "0.468", "0.887", "0.468", "0.901", "0.0809", "0.901" ], [ "0.505", "0.887", "0.8896", "0.887", "0.8896", "0.901", "0.505", "0.901" ] ]

[ "[13]】创建的，方法是使用word2vec（对象标签）上的余", "Racker", "", ":Am", "playin", "ckef", "on:", "NOC(Ou", "racketataball", "", "Atennsnave", "类中选择一个对象以阻止训练。训练集包含8个对象", "（瓶子，公共汽车，沙发，微波炉，比萨饼，球拍，手提", "Bu", "DCC:A", "箱，斑马）的图像和句子，它们构成了MSCOCO数据", "弦距离对主要的80个对象类别进行聚类，并从每个聚", "集中大约10个%的训练图像和字幕对。我们的模型", "NOC", "是根据它能多好地生成关于包含八个伸出物体的图像", "Bottle", "的描述来评估的", "DCC: Aglass ofwine sitting onat", "tablewithaglassofwin", "NOC(Ours):Atablewithabottleofwineandaglass ofwine", "4.3. Image data", "我们还评估了NOC在大约700个不同的ImageNet", "Suitcase", "[21]对象上生成的句子，这些对象不存在于MSCOCO", "DCC: A close up ofa person sitting on a wooden bench.", "Noc (Ours):Abunch of suitcas", "sstackedontopofeachothe", "言语料库（词汇表）中但不存在于MSCOCO中的对", "数据集中。我们通过识别存在于ImageNet和我们的语", "Figure 3.", "COCO Captioning:Examples comparing cap", "象来选择此集合。选择的单词跨越各种类别，包括细粒", "tions by NOC (ours) and DCC [7] on held out objects from", "度类别（例如。g。，“猎犬”和“菊花”），形容词e。", "MSCOCO", "g。、“雪纺”、“羊毛”）和入门级单词(e。g。，“蟾蜍”）。", "更高的F1和METEOR表明NOC能够正确识别更多", "此外，为了研究我们的模型如何描述稀有对象，我们选", "看不见的对象的实例，并将单词整合到流畅的描述中", "择了一组单独的52个对象，这些对象在ImageNet中", "但在MSCOCO中很少提及（平均52次提及，在所有", "5.2. Training data source", "400k训练句子中平均有27次提及）", "为了研究不同数据源的影响，我们还在域外设置中", "5.Experiments onMSCOCO", "评估了我们的模型，其中使用来自ImageNet的图像", "训练了用于扩展对象的分类器，并在从外部语料库中", "我们执行以下实验以将NOC的性能与先前的工作", "挖掘的文本上训练了语言。表2报告了八个搁置对象", "[7]进行比较：1。我们评估了模型对在训练期间从", "的平均得分。我们将我们的NOC模型与7（DCC)", "MSCOCO伸出的对象进行字幕处理的能力（秒。5.1", "的结果以及在相同分割上训练的竞争性图像字幕模型", ")。2。为了研究数据源对训练的影响，我们报告了当域", "LRCNE21进行了比较。在域外设置（第2行）中，对", "内和域外源训练图像和语言网络时NOC的性能（Sec。", "于所选择的8个对象的集合，与DCC相比，NOC在", "5.2）。除此之外，为了更好地理解我们的模型：3。我", "F1上的性能略好，而在METEOR上的性能稍低。但", "们执行烧蚀来研究我们模型的每个组件有多少（比如词", "是，如前所述，DCC需要显式标识一组“可见”对象", "嵌入、辅助目标等。）有助于提高性能（秒。5.3）。4。", "类，以将权重传递给新类，而NOC可以直接用于推", "我们还研究了当从MSCOCO中提取不同的对象子集", "理。DCC的传输机制也导致了特殊的描述。E。g。，无", "时，模型的性能是否保持一致(Sec。5.4）", "花果中的球拍。3", "5.1. Empirical Evaluation on MSCOCO", "使用COCO图像训练（第3行），即使使用Web-", "我们通过遵循7的实验设置，对我们提出的模型描", "CorpusLM训练，NOC的F1分数也会大大提高。最", "述新颖对象的能力进行了经验评估。我们使用以下数", "后，在域内设置（第4行）中，NOC在F1上比DCC", "据集来优化模型中的每个损失：联合学习参数0L和01", "高出约10点，同时也略微提高了METEOR。这表明", "NOC能够通过域内训练更好地将对象与字幕相关联", "的字幕模型仅在MSCOCO的子集上进行训练，而没", "有8个对象（请参阅E4.2节），使用标记的图像来优", "并且辅助自标和嵌人有助于模型概括和描述新颖的对", "化更新参数的图像模型，以及使用相应的描述来训", "象", "练更新参数QL的语言模型。。当在来自COCO的图像", "5.3. Ablations", "上训练视觉网络时，我们通过在移除停止词后将关联", "字幕中的所有词视为标签来获得每个图像的多个标签。", "表3比较了训练的不同方面如何影响整体性能。", "我们首先对域内设置进行评估，其中使用所有COCO", "TunedVisioncontribution未包含手套或LM预训练", "训练图像训练图像分类器，并使用所有句子训练语言", "的模型性能较差（METEOR15。78，F114。41)；这种消", "模型。我们使用METEOR度量【1]来评估描述质量。", "融显示了视觉模型单独在识别和描述伸出的物体方面", "但是，METEOR仅捕获流利度，而不考虑提及（或缺", "的贡献。LM&Glove contribution：在没有辅助目标", "少）特定单词。因此，我们还使用F1来确定模型在包", "的情况下训练的模型，在F1为25的情况下表现更好。", "含对象的图像的描述中提到了对象名称。因此，指标衡", "38和19的流星。80；这种改进主要来自GloVe嵌人", "量模型是否能够识别对象并在句子中流利地使用它", "它有助于为新颖的对象类添加字幕。LM&Pre-trained", "COCOheldout对象。表1将NOC获得的F1分数", "Vision：有趣的是，当我们固定分类器的权重（在所", "与之前的最佳方法DCC[7]在8个被淘汰的 COCO", "有对象上预先训练）时，在调整图像字幕COCO子集", "对象上进行了比较。NOC在除“沙发”和“微波炉”之", "上的LM之前，F1大幅增加到39。70表明视觉模型", "外的所有对象上的性能均优于DCC（平均为10%F1)。", "可以识别许多对象，但可以“忘记”分类器在微调图像" ]

{ "arxiv_id": "1606.07770", "language": "zh", "page_idx": 3 }

[ [ "0.10046", "0.09344", "0.1478", "0.09344", "0.1478", "0.108", "0.10046", "0.108" ], [ "0.3096", "0.0947", "0.3545", "0.0947", "0.3545", "0.1092", "0.3096", "0.1092" ], [ "0.5073", "0.0947", "0.5557", "0.0947", "0.5557", "0.1092", "0.5073", "0.1092" ], [ "0.6396", "0.09216", "0.6807", "0.0947", "0.6787", "0.1092", "0.638", "0.1067" ], [ "0.6953", "0.0947", "0.755", "0.0947", "0.755", "0.1092", "0.6953", "0.1092" ], [ "0.7656", "0.0947", "0.871", "0.09344", "0.871", "0.108", "0.7656", "0.1092" ], [ "0.2632", "0.09534", "0.29", "0.09534", "0.29", "0.108", "0.2632", "0.108" ], [ "0.1993", "0.09534", "0.2418", "0.09534", "0.2418", "0.1067", "0.1993", "0.1067" ], [ "0.3677", "0.09534", "0.4412", "0.0966", "0.4412", "0.108", "0.3677", "0.1067" ], [ "0.4575", "0.0966", "0.4958", "0.0966", "0.4958", "0.108", "0.4575", "0.108" ], [ "0.569", "0.0966", "0.625", "0.0966", "0.625", "0.1067", "0.569", "0.1067" ], [ "0.2035", "0.1111", "0.2385", "0.11426", "0.2378", "0.1288", "0.2009", "0.127" ], [ "0.3105", "0.11237", "0.353", "0.11426", "0.3513", "0.13", "0.3096", "0.1276" ], [ "0.7017", "0.11237", "0.744", "0.11426", "0.7427", "0.1288", "0.701", "0.127" ], [ "0.384", "0.113", "0.4248", "0.113", "0.4248", "0.1276", "0.384", "0.1276" ], [ "0.2566", "0.113", "0.298", "0.113", "0.298", "0.1288", "0.2566", "0.1288" ], [ "0.455", "0.113", "0.4968", "0.113", "0.4968", "0.1276", "0.455", "0.1276" ], [ "0.511", "0.113", "0.5522", "0.113", "0.5522", "0.1276", "0.511", "0.1276" ], [ "0.5767", "0.113", "0.6177", "0.113", "0.6177", "0.1276", "0.5767", "0.1276" ], [ "0.638", "0.113", "0.6807", "0.113", "0.6807", "0.1276", "0.638", "0.1276" ], [ "0.8", "0.11426", "0.8384", "0.11426", "0.8384", "0.1276", "0.8", "0.1276" ], [ "0.1013", "0.1276", "0.183", "0.1276", "0.183", "0.1427", "0.1013", "0.1427" ], [ "0.1013", "0.11426", "0.1396", "0.11426", "0.1396", "0.1276", "0.1013", "0.1276" ], [ "0.2009", "0.1276", "0.2418", "0.1276", "0.2418", "0.1427", "0.2009", "0.1427" ], [ "0.2566", "0.1276", "0.2974", "0.1276", "0.2974", "0.1427", "0.2566", "0.1427" ], [ "0.312", "0.1276", "0.353", "0.1276", "0.353", "0.1427", "0.312", "0.1427" ], [ "0.384", "0.1276", "0.4248", "0.1276", "0.4248", "0.1427", "0.384", "0.1427" ], [ "0.4558", "0.1276", "0.4968", "0.1276", "0.4968", "0.1427", "0.4558", "0.1427" ], [ "0.576", "0.1276", "0.6177", "0.1276", "0.6177", "0.1427", "0.576", "0.1427" ], [ "0.6396", "0.1276", "0.6807", "0.1276", "0.6807", "0.1427", "0.6396", "0.1427" ], [ "0.7036", "0.1276", "0.744", "0.1276", "0.744", "0.1427", "0.7036", "0.1427" ], [ "0.8", "0.1288", "0.8384", "0.1288", "0.8384", "0.1427", "0.8", "0.1427" ], [ "0.1422", "0.1578", "0.8896", "0.1578", "0.8896", "0.1718", "0.1422", "0.1718" ], [ "0.513", "0.1288", "0.5513", "0.1288", "0.5513", "0.1427", "0.513", "0.1427" ], [ "0.0825", "0.159", "0.1422", "0.159", "0.1422", "0.1692", "0.0825", "0.1692" ], [ "0.1396", "0.1622", "0.1462", "0.1622", "0.1462", "0.1674", "0.1396", "0.1674" ], [ "0.0809", "0.1718", "0.8906", "0.1718", "0.8906", "0.1863", "0.0809", "0.1863" ], [ "0.0801", "0.1863", "0.2443", "0.1863", "0.2443", "0.1995", "0.0801", "0.1995" ], [ "0.1356", "0.2203", "0.1847", "0.2203", "0.1847", "0.2349", "0.1356", "0.2349" ], [ "0.4102", "0.2203", "0.433", "0.2203", "0.433", "0.2349", "0.4102", "0.2349" ], [ "0.5034", "0.2184", "0.888", "0.2184", "0.888", "0.2317", "0.5034", "0.2317" ], [ "0.259", "0.2216", "0.308", "0.2216", "0.308", "0.2329", "0.259", "0.2329" ], [ "0.2035", "0.2216", "0.2361", "0.2216", "0.2361", "0.2343", "0.2035", "0.2343" ], [ "0.3186", "0.2216", "0.3857", "0.2216", "0.3857", "0.2329", "0.3186", "0.2329" ], [ "0.5024", "0.2317", "0.8906", "0.2329", "0.8906", "0.2474", "0.5024", "0.2468" ], [ "0.1602", "0.2393", "0.2181", "0.2393", "0.2181", "0.254", "0.1602", "0.254" ], [ "0.2573", "0.2393", "0.3057", "0.2393", "0.3057", "0.254", "0.2573", "0.254" ], [ "0.3342", "0.2393", "0.3743", "0.2393", "0.3743", "0.254", "0.3342", "0.254" ], [ "0.4143", "0.2393", "0.4314", "0.2393", "0.4314", "0.254", "0.4143", "0.254" ], [ "0.10864", "0.2488", "0.11926", "0.2488", "0.11926", "0.2583", "0.10864", "0.2583" ], [ "0.2615", "0.2524", "0.3022", "0.2524", "0.3022", "0.2676", "0.2615", "0.2676" ], [ "0.3342", "0.2524", "0.3743", "0.2524", "0.3743", "0.2683", "0.3342", "0.2683" ], [ "0.5034", "0.2488", "0.8906", "0.2488", "0.8906", "0.262", "0.5034", "0.262" ], [ "0.4143", "0.254", "0.4307", "0.254", "0.4307", "0.2683", "0.4143", "0.2683" ], [ "0.1447", "0.254", "0.2336", "0.252", "0.2336", "0.2664", "0.1447", "0.2683" ], [ "0.5024", "0.262", "0.6357", "0.2632", "0.6357", "0.2778", "0.5024", "0.2766" ], [ "0.1356", "0.2734", "0.179", "0.2734", "0.179", "0.2886", "0.1356", "0.2886" ], [ "0.2051", "0.2734", "0.2345", "0.2734", "0.2345", "0.286", "0.2051", "0.286" ], [ "0.2646", "0.2734", "0.3022", "0.2734", "0.3022", "0.2874", "0.2646", "0.2874" ], [ "0.335", "0.2734", "0.3733", "0.2734", "0.3733", "0.2874", "0.335", "0.2874" ], [ "0.4019", "0.2727", "0.4429", "0.2727", "0.4429", "0.2874", "0.4019", "0.2874" ], [ "0.10864", "0.281", "0.1209", "0.281", "0.1209", "0.2922", "0.10864", "0.2922" ], [ "0.1438", "0.286", "0.1724", "0.2886", "0.1708", "0.3018", "0.1422", "0.2998" ], [ "0.1936", "0.2874", "0.2443", "0.2874", "0.2443", "0.3018", "0.1936", "0.3018" ], [ "0.2632", "0.2874", "0.3015", "0.2874", "0.3015", "0.3005", "0.2632", "0.3005" ], [ "0.335", "0.2886", "0.3733", "0.2886", "0.3733", "0.3018", "0.335", "0.3018" ], [ "0.4019", "0.2874", "0.4429", "0.2874", "0.4429", "0.3018", "0.4019", "0.3018" ], [ "0.5034", "0.2903", "0.7363", "0.2903", "0.7363", "0.305", "0.5034", "0.305" ], [ "0.2035", "0.305", "0.2385", "0.308", "0.2361", "0.3225", "0.2009", "0.3193" ], [ "0.3342", "0.3125", "0.3743", "0.3145", "0.3733", "0.3289", "0.3325", "0.327" ], [ "0.4019", "0.3113", "0.4429", "0.3132", "0.4412", "0.3289", "0.4004", "0.327" ], [ "0.524", "0.3132", "0.8896", "0.3132", "0.8896", "0.3276", "0.524", "0.3276" ], [ "0.2615", "0.3132", "0.3022", "0.3132", "0.3022", "0.3276", "0.2615", "0.3276" ], [ "0.10864", "0.315", "0.12335", "0.315", "0.12335", "0.3257", "0.10864", "0.3257" ], [ "0.1356", "0.315", "0.179", "0.315", "0.179", "0.327", "0.1356", "0.327" ], [ "0.1969", "0.3225", "0.2418", "0.3225", "0.2418", "0.334", "0.1969", "0.334" ], [ "0.505", "0.3289", "0.8896", "0.3289", "0.8896", "0.3423", "0.505", "0.3423" ], [ "0.2646", "0.3416", "0.2998", "0.3416", "0.2998", "0.3542", "0.2646", "0.3542" ], [ "0.3342", "0.3403", "0.3743", "0.3403", "0.3743", "0.3547", "0.3342", "0.3547" ], [ "0.4028", "0.3416", "0.4412", "0.3416", "0.4412", "0.3547", "0.4028", "0.3547" ], [ "0.5034", "0.3435", "0.8896", "0.3435", "0.8896", "0.3567", "0.5034", "0.3567" ], [ "0.1356", "0.3484", "0.1805", "0.3484", "0.1805", "0.3599", "0.1356", "0.3599" ], [ "0.1152", "0.3452", "0.1259", "0.3542", "0.1152", "0.3623", "0.10455", "0.3542" ], [ "0.1969", "0.3484", "0.2418", "0.3484", "0.2418", "0.3599", "0.1969", "0.3599" ], [ "0.2646", "0.3547", "0.3022", "0.3547", "0.3022", "0.3687", "0.2646", "0.3687" ], [ "0.335", "0.3547", "0.3733", "0.3547", "0.3733", "0.3687", "0.335", "0.3687" ], [ "0.4028", "0.3516", "0.4412", "0.3542", "0.4395", "0.3694", "0.4019", "0.3674" ], [ "0.5034", "0.3591", "0.892", "0.3591", "0.892", "0.3738", "0.5034", "0.3738" ], [ "0.505", "0.3738", "0.8896", "0.3738", "0.8896", "0.3882", "0.505", "0.3882" ], [ "0.0809", "0.385", "0.1487", "0.385", "0.1487", "0.3958", "0.0809", "0.3958" ], [ "0.1422", "0.3838", "0.468", "0.3838", "0.468", "0.3977", "0.1422", "0.3977" ], [ "0.505", "0.3894", "0.8906", "0.3894", "0.8906", "0.4028", "0.505", "0.4028" ], [ "0.0801", "0.3977", "0.468", "0.3977", "0.468", "0.411", "0.0801", "0.411" ], [ "0.5024", "0.404", "0.8896", "0.404", "0.8896", "0.4185", "0.5024", "0.4185" ], [ "0.0784", "0.4104", "0.4673", "0.411", "0.4673", "0.4255", "0.0784", "0.4248" ], [ "0.5034", "0.4192", "0.8906", "0.4192", "0.8906", "0.433", "0.5034", "0.433" ], [ "0.0784", "0.428", "0.1152", "0.428", "0.1152", "0.4395", "0.0784", "0.4395" ], [ "0.5034", "0.435", "0.8906", "0.435", "0.8906", "0.449", "0.5034", "0.449" ], [ "0.0809", "0.4475", "0.4673", "0.4475", "0.4673", "0.461", "0.0809", "0.461" ], [ "0.505", "0.4507", "0.8896", "0.4507", "0.8896", "0.464", "0.505", "0.464" ], [ "0.0825", "0.4634", "0.468", "0.4634", "0.468", "0.4766", "0.0825", "0.4766" ], [ "0.5034", "0.4653", "0.8906", "0.4653", "0.8906", "0.4792", "0.5034", "0.4792" ], [ "0.0825", "0.4778", "0.468", "0.4778", "0.468", "0.4912", "0.0825", "0.4912" ], [ "0.505", "0.4797", "0.8906", "0.4797", "0.8906", "0.4937", "0.505", "0.4937" ], [ "0.0809", "0.4937", "0.4673", "0.4937", "0.4673", "0.507", "0.0809", "0.507" ], [ "0.5034", "0.4944", "0.8906", "0.4944", "0.8906", "0.5083", "0.5034", "0.5083" ], [ "0.0809", "0.5083", "0.4656", "0.5083", "0.4656", "0.5215", "0.0809", "0.5215" ], [ "0.5034", "0.509", "0.892", "0.509", "0.892", "0.5234", "0.5034", "0.5234" ], [ "0.0809", "0.523", "0.4575", "0.523", "0.4575", "0.537", "0.0809", "0.537" ], [ "0.505", "0.526", "0.589", "0.526", "0.589", "0.537", "0.505", "0.537" ], [ "0.524", "0.5405", "0.8906", "0.5405", "0.8906", "0.5547", "0.524", "0.5547" ], [ "0.0809", "0.55", "0.4346", "0.55", "0.4346", "0.563", "0.0809", "0.563" ], [ "0.505", "0.556", "0.8906", "0.556", "0.8906", "0.5693", "0.505", "0.5693" ], [ "0.1013", "0.5737", "0.4673", "0.5737", "0.4673", "0.587", "0.1013", "0.587" ], [ "0.505", "0.571", "0.8906", "0.571", "0.8906", "0.5854", "0.505", "0.5854" ], [ "0.0801", "0.587", "0.468", "0.5884", "0.468", "0.603", "0.0801", "0.6016" ], [ "0.5034", "0.5854", "0.8896", "0.5854", "0.8896", "0.5986", "0.5034", "0.5986" ], [ "0.0825", "0.604", "0.468", "0.604", "0.468", "0.617", "0.0825", "0.617" ], [ "0.5024", "0.5996", "0.8906", "0.601", "0.8906", "0.6157", "0.5024", "0.6143" ], [ "0.0825", "0.619", "0.468", "0.619", "0.468", "0.6333", "0.0825", "0.6333" ], [ "0.5034", "0.616", "0.8906", "0.616", "0.8906", "0.63", "0.5034", "0.63" ], [ "0.0825", "0.634", "0.468", "0.634", "0.468", "0.648", "0.0825", "0.648" ], [ "0.5034", "0.632", "0.6055", "0.632", "0.6055", "0.643", "0.5034", "0.643" ], [ "0.0825", "0.6484", "0.468", "0.6484", "0.468", "0.662", "0.0825", "0.662" ], [ "0.523", "0.6465", "0.892", "0.6465", "0.892", "0.661", "0.523", "0.661" ], [ "0.0801", "0.662", "0.468", "0.6636", "0.468", "0.678", "0.0801", "0.677" ], [ "0.501", "0.661", "0.8906", "0.662", "0.8906", "0.677", "0.501", "0.6753" ], [ "0.0801", "0.678", "0.468", "0.6787", "0.468", "0.6934", "0.0801", "0.6924" ], [ "0.5034", "0.677", "0.8906", "0.677", "0.8906", "0.6914", "0.5034", "0.6914" ], [ "0.0801", "0.6924", "0.1487", "0.6943", "0.1487", "0.709", "0.0801", "0.707" ], [ "0.5034", "0.6924", "0.8896", "0.6924", "0.8896", "0.7056", "0.5034", "0.7056" ], [ "0.505", "0.7085", "0.876", "0.7085", "0.876", "0.7217", "0.505", "0.7217" ], [ "0.0801", "0.726", "0.3987", "0.726", "0.3987", "0.7407", "0.0801", "0.7407" ], [ "0.505", "0.7227", "0.8906", "0.7227", "0.8906", "0.7363", "0.505", "0.7363" ], [ "0.5034", "0.738", "0.8906", "0.738", "0.8906", "0.752", "0.5034", "0.752" ], [ "0.1013", "0.752", "0.468", "0.752", "0.468", "0.765", "0.1013", "0.765" ], [ "0.5034", "0.7524", "0.8906", "0.7524", "0.8906", "0.7666", "0.5034", "0.7666" ], [ "0.0825", "0.765", "0.4656", "0.765", "0.4656", "0.779", "0.0825", "0.779" ], [ "0.5034", "0.768", "0.7485", "0.768", "0.7485", "0.782", "0.5034", "0.782" ], [ "0.0809", "0.7827", "0.4673", "0.7827", "0.4673", "0.7935", "0.0809", "0.7935" ], [ "0.0809", "0.797", "0.4673", "0.797", "0.4673", "0.81", "0.0809", "0.81" ], [ "0.5034", "0.7915", "0.786", "0.7915", "0.786", "0.8047", "0.5034", "0.8047" ], [ "0.0825", "0.811", "0.4673", "0.811", "0.4673", "0.8247", "0.0825", "0.8247" ], [ "0.523", "0.81", "0.8896", "0.81", "0.8896", "0.8247", "0.523", "0.8247" ], [ "0.0809", "0.827", "0.468", "0.827", "0.468", "0.8403", "0.0809", "0.8403" ], [ "0.505", "0.8257", "0.8906", "0.8257", "0.8906", "0.8403", "0.505", "0.8403" ], [ "0.0809", "0.8413", "0.4673", "0.8413", "0.4673", "0.855", "0.0809", "0.855" ], [ "0.5034", "0.8413", "0.892", "0.8413", "0.892", "0.855", "0.5034", "0.855" ], [ "0.0809", "0.857", "0.4673", "0.857", "0.4673", "0.8706", "0.0809", "0.8706" ], [ "0.5024", "0.856", "0.8896", "0.856", "0.8896", "0.8706", "0.5024", "0.8706" ], [ "0.0825", "0.872", "0.468", "0.872", "0.468", "0.8853", "0.0825", "0.8853" ], [ "0.5034", "0.872", "0.888", "0.872", "0.888", "0.8853", "0.5034", "0.8853" ], [ "0.0825", "0.887", "0.4673", "0.887", "0.4673", "0.901", "0.0825", "0.901" ], [ "0.5034", "0.887", "0.892", "0.887", "0.892", "0.901", "0.5034", "0.901" ] ]

[ "Model", "couch", "racket", "zebra", "Avg. F1", "Avg.METEOR", "bus", "bottle", "microwave", "pizza", "suitcase", "4.63", "45.87", "39.78", "28.09", "29.79", "64.59", "52.24", "13.16", "79.88", "21.00", "NOC (ours)", "DCC", "17.78", "68.79", "25.55", "24.72", "69.33", "39.86", "89.02", "48.79", "21.32", "MSCOCO Captioning: F1 scores (in % ) of NOC (our model) and DCC [7] on held-out objects not seen jointly", "55.31", "Table 1.", "", "during image-caption training, along with the average F1 and METEOR scores of the generated captions across images", "containing these objects.", "Image", "F1", "能够描述的图像和对象。此外，我们获得了人类评估", "Model", "Text", "METEOR", "将我们的模型与先前的工作进行了比较，以确定该模", "Baseline", "LRCN", "19.33", "0", "1", "DCC", "19.90", "型是否能够将对象标签有意义地包含在描述中以及它", "0", "(no transfer)", "对图像的描述程度", "Image", "Web", "DCC", "20.66", "34.94", "2", "Net", "Corpus", "NOC", "17.56", "36.50", "6.1. Describing Novel Objects", "Web", "19.18", "41.74", "表E5使用以下指标比较了638个新颖对象类别（与", "NOC", "3", "COCO", "Corpus", "7]相同）上的模型：（i）描述新颖对象（%）是指描", "DCC", "21.00", "39.78", "述中提到的所选ImageNet对象的百分比，i。e。每个", "COCO", "M", "COCO", "NOC", "21.32", "48.79", "新单词的（e。g。，“otter”）模型应将单词（“otter\"）", "合并到有关对象（otter）的ImageNet图像的至少一个", "Table 2.", "Comparison with different training data sources", "描述中。虽然DCC能够识别和描述56。85描述中选", "on 8 held-out COCO objects. Having in-domain data helps", "定的ImageNet对象的%（363），NOC可以识别更多", "both the DCC [7] and our NOC model caption novel ob-", "对象，并且能够描述91。27个%（582个，共638个）", "jects.", "ImageNet对象。（i）准确性是指模型能够正确识别和", "字幕数据时学习的对象（没有8个对象）。Auxiliary", "描述类别的每个类别的图像百分比。我们报告所有类", "Objective：结合辅助目标，F1显着提高到47。02。我", "别的平均准确性。DCC正确合并了一个新词11。08", "们在这里注意到，由于包括辅助目标，视觉网络可以在", "当时的%，相比之下，NOC将其明显改进为24。74%。", "所有图像上进行调整，从而保留了对各种对象进行分", "(ii）基于描述中提及对象的精确度和召回率计算F1分", "类/识别的能力。最后，结合所有方面为NOC提供了", "数。同样，NOC的表现优于平均F133。76%到DCC", "最佳性能（F148。79，流星21。32)，显著优于DCC", "的14。47%", "尽管NOC和DCCE[7]使用相同的CNN，但NOC", "5.4. Validating on a different subset of COCO", "都能够描述更多类别，并且更频繁地将新词正确地集", "为了表明我们的模型在对象之间是一致的，我们通", "成到描述中。DCCE[7]可能无法找到在语义上和语法", "过从COCO中提取一组不同的八个对象来创建不同的", "上都与新颖对象相似的合适对象，或者在使用对象名", "训练/测试拆分。我们持有的对象是：床，书，胡萝卜，", "称组成句子的语言模型方面，在NOC中，前者从未发", "大象，勺子，厕所，卡车和雨伞。来自这八个对象的图", "生（i。e。我们不需要显式识别类似的对象），减少了错", "像和句子再次构成了MSCOCO训练数据集的大约10", "误的总体来源，", "个%。表4显示了该模型在该子集上的性能。我们观察", "图。4和Fig。6（第3列）显示了NOC描述", "到F1和METEOR分数虽然略低，但与表1中观察", "ImageNet中的各种对象的示例。Fig。4将我们的", "到的数字一致，证实了我们的模型能够推广到不同的", "模型与DCC进行了比较。Fig。5和图。6（右）", "对象子集", "概述了一些错误。无法描述新对象是NOC的一个常", "见错误。E。g。Fig。6（右上），NOC错误地将", "6. Experiments: Scaling to ImageNet", "个拿着“西塔”的人描述为一个拿着“棒球棒”的人。", "其他常见错误包括生成非语法或无意义的短语（例如", "为了演示NOC的可伸缩性，我们描述了ImageNet", "“角斗士”，“aardvark”）并重复特定对象（“梭子鱼。", "中不存在配对图像-句子数据的对象。我们的实验是在", "withabarracuda',triflecake’)。", "ImageNet的两个子集上进行的，（i）新颖对象：一组", "638个对象，它们存在于ImageNet以及模型的词汇表", "6.2. Describing Rare Objects/Words", "中，但在MSCOCO中未提及。（i）稀有对象：一组", "选定的稀有词在MSCOCO训练集中出现的频率各", "52个对象，它们在ImageNet和MSCOCO词汇表中，", "不相同，在所有训练句子中平均约有52次提及（中位", "但在MSCOCO标题中很少提及（中位数为27)。对于", "数为27）。例如，诸如“盆景”之类的单词仅出现5", "定量评估，（i）我们测量模型能够描述对象的至少一个", "次，“whisk”（11个注释），“茶壶”（30个注释），以", "图像（使用对象标签）的对象的百分比，（ii）我们还报", "及诸如南瓜之类的其他单词出现58次，“天鹅”（60个", "告准确性和F1分数，以便在整个集合中进行比较模型", "注释），而在较高的侧面，像围巾这样的对象出现144" ]

{ "arxiv_id": "1606.07770", "language": "zh", "page_idx": 4 }

[ [ "0.336", "0.0861", "0.665", "0.0861", "0.665", "0.10095", "0.336", "0.10095" ], [ "0.2988", "0.1401", "0.7153", "0.1395", "0.7153", "0.1567", "0.2988", "0.158" ], [ "0.3342", "0.1567", "0.6724", "0.158", "0.6724", "0.1758", "0.3342", "0.1746" ], [ "0.3167", "0.177", "0.6978", "0.177", "0.6978", "0.1912", "0.3167", "0.1912" ], [ "0.266", "0.2654", "0.3435", "0.2654", "0.3435", "0.2798", "0.266", "0.2798" ], [ "0.517", "0.2666", "0.8833", "0.2666", "0.8833", "0.2798", "0.517", "0.2798" ], [ "0.5146", "0.2832", "0.8403", "0.2832", "0.8403", "0.2964", "0.5146", "0.2964" ], [ "0.147", "0.2922", "0.4575", "0.2922", "0.4575", "0.306", "0.147", "0.306" ], [ "0.534", "0.2998", "0.882", "0.2998", "0.882", "0.3135", "0.534", "0.3135" ], [ "0.1486", "0.308", "0.4592", "0.306", "0.4592", "0.3196", "0.1486", "0.3218" ], [ "0.5146", "0.3142", "0.881", "0.3142", "0.881", "0.328", "0.5146", "0.328" ], [ "0.1486", "0.3218", "0.4592", "0.3218", "0.4592", "0.335", "0.1486", "0.335" ], [ "0.5156", "0.3308", "0.882", "0.3308", "0.882", "0.3445", "0.5156", "0.3445" ], [ "0.1486", "0.3374", "0.455", "0.3362", "0.455", "0.3474", "0.1486", "0.3486" ], [ "0.1503", "0.3516", "0.4575", "0.3516", "0.4575", "0.3623", "0.1503", "0.3623" ], [ "0.5156", "0.3474", "0.882", "0.3474", "0.882", "0.361", "0.5156", "0.361" ], [ "0.1453", "0.3635", "0.4592", "0.3652", "0.4592", "0.379", "0.1453", "0.3772" ], [ "0.517", "0.3635", "0.882", "0.3635", "0.882", "0.3772", "0.517", "0.3772" ], [ "0.147", "0.3777", "0.4568", "0.379", "0.4568", "0.3926", "0.147", "0.3914" ], [ "0.5156", "0.3801", "0.8833", "0.3801", "0.8833", "0.3938", "0.5156", "0.3938" ], [ "0.1486", "0.3926", "0.4592", "0.3926", "0.4592", "0.4062", "0.1486", "0.4062" ], [ "0.5156", "0.3955", "0.882", "0.3955", "0.882", "0.4092", "0.5156", "0.4092" ], [ "0.1486", "0.4067", "0.4575", "0.408", "0.4575", "0.4216", "0.1486", "0.4204" ], [ "0.517", "0.411", "0.882", "0.411", "0.882", "0.4246", "0.517", "0.4246" ], [ "0.147", "0.4216", "0.4575", "0.4216", "0.4575", "0.4353", "0.147", "0.4353" ], [ "0.517", "0.4275", "0.882", "0.4275", "0.882", "0.4412", "0.517", "0.4412" ], [ "0.147", "0.4353", "0.4592", "0.434", "0.4592", "0.449", "0.147", "0.4495" ], [ "0.517", "0.4448", "0.882", "0.4448", "0.882", "0.4578", "0.517", "0.4578" ], [ "0.1511", "0.4507", "0.2527", "0.4507", "0.2527", "0.4614", "0.1511", "0.4614" ], [ "0.2477", "0.4507", "0.4575", "0.4507", "0.4575", "0.4614", "0.2477", "0.4614" ], [ "0.1486", "0.4631", "0.4592", "0.4631", "0.4592", "0.4768", "0.1486", "0.4768" ], [ "0.5156", "0.4602", "0.881", "0.4602", "0.881", "0.4739", "0.5156", "0.4739" ], [ "0.1503", "0.4797", "0.4534", "0.4797", "0.4534", "0.4905", "0.1503", "0.4905" ], [ "0.517", "0.4768", "0.882", "0.4768", "0.882", "0.4905", "0.517", "0.4905" ], [ "0.147", "0.4922", "0.4575", "0.4922", "0.4575", "0.506", "0.147", "0.506" ], [ "0.5156", "0.4922", "0.881", "0.4922", "0.881", "0.506", "0.5156", "0.506" ], [ "0.1486", "0.5083", "0.455", "0.5083", "0.455", "0.5186", "0.1486", "0.5186" ], [ "0.5146", "0.5073", "0.882", "0.5083", "0.882", "0.5225", "0.5146", "0.5215" ], [ "0.1486", "0.5225", "0.2048", "0.5225", "0.2048", "0.5337", "0.1486", "0.5337" ], [ "0.5156", "0.525", "0.6826", "0.525", "0.6826", "0.538", "0.5156", "0.538" ], [ "0.11755", "0.549", "0.157", "0.549", "0.157", "0.56", "0.11755", "0.56" ], [ "0.1486", "0.5464", "0.266", "0.549", "0.266", "0.5635", "0.1486", "0.561" ], [ "0.511", "0.552", "0.6055", "0.552", "0.6055", "0.5635", "0.511", "0.5635" ], [ "0.12177", "0.5737", "0.4878", "0.5737", "0.4878", "0.5874", "0.12177", "0.5874" ], [ "0.523", "0.5767", "0.882", "0.579", "0.882", "0.5933", "0.523", "0.5913" ], [ "0.12177", "0.5884", "0.4878", "0.5884", "0.4878", "0.6016", "0.12177", "0.6016" ], [ "0.513", "0.594", "0.8833", "0.594", "0.8833", "0.608", "0.513", "0.608" ], [ "0.12006", "0.6055", "0.4895", "0.6055", "0.4895", "0.619", "0.12006", "0.619" ], [ "0.5156", "0.6094", "0.882", "0.61", "0.882", "0.6245", "0.5156", "0.6235" ], [ "0.12006", "0.6216", "0.4878", "0.6216", "0.4878", "0.6353", "0.12006", "0.6353" ], [ "0.517", "0.627", "0.882", "0.627", "0.882", "0.64", "0.517", "0.64" ], [ "0.1184", "0.637", "0.4878", "0.6357", "0.4878", "0.651", "0.1184", "0.652" ], [ "0.5156", "0.6436", "0.882", "0.6436", "0.882", "0.6567", "0.5156", "0.6567" ], [ "0.1184", "0.653", "0.486", "0.653", "0.486", "0.6665", "0.1184", "0.6665" ], [ "0.5146", "0.659", "0.882", "0.659", "0.882", "0.6733", "0.5146", "0.6733" ], [ "0.12006", "0.6694", "0.4878", "0.6694", "0.4878", "0.683", "0.12006", "0.683" ], [ "0.5156", "0.6733", "0.882", "0.6733", "0.882", "0.6885", "0.5156", "0.6885" ], [ "0.12177", "0.685", "0.486", "0.685", "0.486", "0.6987", "0.12177", "0.6987" ], [ "0.5146", "0.6914", "0.882", "0.6914", "0.882", "0.705", "0.5146", "0.705" ], [ "0.1234", "0.703", "0.4895", "0.703", "0.4895", "0.7163", "0.1234", "0.7163" ], [ "0.5156", "0.705", "0.8833", "0.706", "0.8833", "0.72", "0.5156", "0.7197" ], [ "0.12006", "0.7173", "0.378", "0.7173", "0.378", "0.731", "0.12006", "0.731" ], [ "0.5156", "0.7217", "0.771", "0.7217", "0.771", "0.7354", "0.5156", "0.7354" ], [ "0.1377", "0.7354", "0.4878", "0.7354", "0.4878", "0.749", "0.1377", "0.749" ], [ "0.532", "0.7393", "0.8833", "0.7393", "0.8833", "0.753", "0.532", "0.753" ], [ "0.1184", "0.75", "0.4844", "0.75", "0.4844", "0.763", "0.1184", "0.763" ], [ "0.5146", "0.7544", "0.8833", "0.755", "0.8833", "0.7695", "0.5146", "0.7686" ], [ "0.1184", "0.7656", "0.4878", "0.7666", "0.4878", "0.781", "0.1184", "0.78" ], [ "0.5146", "0.7715", "0.8833", "0.7715", "0.8833", "0.785", "0.5146", "0.785" ], [ "0.12177", "0.783", "0.4878", "0.783", "0.4878", "0.7964", "0.12177", "0.7964" ], [ "0.5156", "0.7876", "0.882", "0.7876", "0.882", "0.8003", "0.5156", "0.8003" ], [ "0.12006", "0.799", "0.4836", "0.799", "0.4836", "0.8135", "0.12006", "0.8135" ], [ "0.1142", "0.8057", "0.885", "0.8154", "0.885", "0.839", "0.1142", "0.83" ], [ "0.513", "0.802", "0.882", "0.8027", "0.882", "0.8164", "0.513", "0.8154" ], [ "0.126", "0.836", "0.1368", "0.836", "0.1368", "0.8413", "0.126", "0.8413" ], [ "0.1335", "0.8306", "0.4895", "0.832", "0.4895", "0.8457", "0.1335", "0.844" ], [ "0.511", "0.836", "0.58", "0.8345", "0.58", "0.8486", "0.511", "0.851" ], [ "0.1184", "0.847", "0.486", "0.847", "0.486", "0.861", "0.1184", "0.861" ], [ "0.532", "0.858", "0.8774", "0.8594", "0.8774", "0.8735", "0.532", "0.8726" ], [ "0.12006", "0.8643", "0.4895", "0.8643", "0.4895", "0.8774", "0.12006", "0.8774" ], [ "0.5146", "0.8716", "0.8833", "0.8716", "0.8833", "0.8853", "0.5146", "0.8853" ], [ "0.12006", "0.88", "0.4895", "0.88", "0.4895", "0.8936", "0.12006", "0.8936" ], [ "0.5146", "0.8843", "0.6245", "0.8843", "0.6245", "0.8955", "0.5146", "0.8955" ], [ "0.11584", "0.8945", "0.4895", "0.8926", "0.4895", "0.91", "0.11584", "0.912" ], [ "0.534", "0.8965", "0.882", "0.8965", "0.882", "0.91", "0.534", "0.91" ] ]

[ "BERT重新发现经典NLP管道", "Ian Tenneyl Dipanjan DaslEllie Pavlick1,2", "1Google Research2Brown University", "[iftenney,dipanjand,epavlick @google.com", "Abstract", "且在模型(Petersetal.,2018bBlevinsetal.,", "2018）的较高层中分层表示更复杂的结构", "Pre-trained text encoders have rapidly ad", "我们以后一种工作为基础，专注于BERT模", "vanced the state of the art on many NLP", "型（Devlinetal.,2019），并使用从传统NLP管", "tasks. We focus on one such model, BERT,", "道派生的一套探测任务（Tenneyetal，2019)", "and aim to quantify where linguistic infor", "mationis captured withinthenetwork.We", "来量化特定类型的语言信息的编码位置。基于", "find that the model represents the steps", "(Petersetal.，2018b）的观察结果，即语言模型", "of the traditional NLP pipeline in an in", "的较低层编码更多的本地语法，而较高层捕获", "terpretable and localizable way, and that", "更复杂的语义，我们提出了两个新颖的贡献。", "the regions responsible for each step ap-", "首先，我们提出了一个跨越传统NLP管道常", "pear in the expected sequence: POS tag-", "见组件的分析。我们证明了特定抽象的编码顺", "ging, parsing, NER, semantic roles, then", "序反映了这些任务的传统层次结构。其次，我", "coreference.", "Qualitative analysis reveals", "that the model can and often does adjust", "们定性分析BERT网络如何逐层处理单个句", "this pipelinedynamically,revising lower", "子。我们表明，虽然管道顺序保持在总体上", "level decisions on the basis of disambiguat", "但该模型可以允许各个决策以任意方式相互依", "inginformationfromhigher-levelrepresen", "赖，推迟模棱两可的决策或根据更高级别的信", "tations.", "息修改不正确的决策", "1", "Introduction", "2 Model", "诸如ELMo(Peters et al.,2018a）和 BERT", "EdgeProbing.我们的实验基于Tenney", "(Devlinetal.，2019）之类的预训练句子编码", "etal.（2019）的“边缘探测”方法，该方法", "器已迅速改善了许多NLP任务的最新技术水", "旨在测量如何从预先训练的编码器中提取有", "平，并且似乎准备取代静态单词嵌入（Mikolov", "关语言结构的信息。边缘探测将结构化预测任", "etal.，2013）和离散管道（Manningetal.,2014)", "务分解为通用格式，其中探测分类器接收跨度", "作为自然语言处理系统的基础。虽然这对性能", "S1=[31,J1）和（可选）S2=[2,j2），并且必须", "是一个福音，但它是以可解释性为代价的，目", "预测标签，例如成分或关系类型。1探测分类", "前还不清楚这些模型实际上是在学习我们直觉", "器只能访问目标跨度内的每令牌上下文向量，", "上认为对表示自然语言很重要的抽象，还是仅", "因此必须依赖编码器提供有关这些跨度与其在", "仅是对复杂的共现统计进行建模", "句子中的角色之间的关系的信息", "最近的工作浪潮已经开始“探索”最先进的", "我们使用边缘探测套件中的八个标记任务：", "模型，以了解它们是否以令人满意的方式表示", "词性（POS)，成分（Consts。）、依赖项（Deps。", "语言。这项工作大部分是基于行为的，设计受控", ")、实体、语义角色标签(SRL)、协指（Coref。）", "测试集并分析错误，以便对模型可能表示或不", "语义原角色（SPR；？）和关系分类（SemEval)。", "表示（?)抽象类型进行逆向工程[e。g。Jlcon", "neau2018cram,marvin2018targeted,poliak2018colt平均F1）进行评估，以便于跨任务进行", "这些任务源自标准基准数据集，并使用通用指", "", "并行工作直接检查网络的结构，以评估是否", "比较。2", "存在与不同类型的语言决策相关的可本地化区", "'For single-span tasks (POS, entities, and con", "域。这样的工作已经产生了证据，表明深层", "stituents), $2 is not used. For POS, s1 = [i,i + 1)", "语言模型可以编码一系列语法和语义信息（e。", "is a single token", "g。Shi et al., 2016; Belinkov, 2018;？)，并", "2We use the code from https://github.com/" ]

{ "arxiv_id": "1905.05950", "language": "zh", "page_idx": 0 }

[ [ "0.592", "0.07245", "0.6465", "0.07245", "0.6465", "0.08435", "0.592", "0.08435" ], [ "0.6826", "0.07184", "0.864", "0.07184", "0.864", "0.0849", "0.6826", "0.0849" ], [ "0.3267", "0.0766", "0.4878", "0.0766", "0.4878", "0.0903", "0.3267", "0.0903" ], [ "0.1394", "0.076", "0.336", "0.076", "0.336", "0.0903", "0.1394", "0.0903" ], [ "0.5903", "0.0885", "0.671", "0.0885", "0.671", "0.0998", "0.5903", "0.0998" ], [ "0.686", "0.0903", "0.6987", "0.0903", "0.6987", "0.0986", "0.686", "0.0986" ], [ "0.706", "0.0885", "0.8833", "0.0885", "0.8833", "0.10156", "0.706", "0.10156" ], [ "0.575", "0.1051", "0.6533", "0.1069", "0.6533", "0.1217", "0.575", "0.1194" ], [ "0.1184", "0.09143", "0.4895", "0.09265", "0.4895", "0.1069", "0.1184", "0.1057" ], [ "0.544", "0.1069", "0.581", "0.1069", "0.581", "0.1194", "0.544", "0.1194" ], [ "0.6753", "0.1057", "0.7095", "0.1057", "0.7095", "0.1217", "0.6753", "0.1217" ], [ "0.7793", "0.1081", "0.8154", "0.1081", "0.8154", "0.12054", "0.7793", "0.12054" ], [ "0.1184", "0.1247", "0.4878", "0.1247", "0.4878", "0.1383", "0.1184", "0.1383" ], [ "0.12006", "0.10925", "0.4878", "0.10925", "0.4878", "0.1223", "0.12006", "0.1223" ], [ "0.523", "0.1247", "0.581", "0.1247", "0.581", "0.1395", "0.523", "0.1395" ], [ "0.681", "0.1247", "0.7153", "0.1247", "0.7153", "0.1401", "0.681", "0.1401" ], [ "0.796", "0.1247", "0.837", "0.1247", "0.837", "0.1401", "0.796", "0.1401" ], [ "0.11755", "0.1383", "0.4265", "0.1395", "0.4265", "0.1567", "0.11755", "0.1561" ], [ "0.5845", "0.1271", "0.6533", "0.1271", "0.6533", "0.1383", "0.5845", "0.1383" ], [ "0.444", "0.1414", "0.486", "0.1414", "0.486", "0.1526", "0.444", "0.1526" ], [ "0.5356", "0.1414", "0.58", "0.1437", "0.5786", "0.1592", "0.534", "0.1567" ], [ "0.706", "0.1437", "0.7407", "0.1437", "0.7407", "0.1592", "0.706", "0.1592" ], [ "0.804", "0.1437", "0.8447", "0.1437", "0.8447", "0.1592", "0.804", "0.1592" ], [ "0.1184", "0.1567", "0.519", "0.1567", "0.519", "0.1716", "0.1184", "0.1716" ], [ "0.5845", "0.1455", "0.6533", "0.1455", "0.6533", "0.1567", "0.5845", "0.1567" ], [ "0.525", "0.1603", "0.581", "0.1633", "0.58", "0.1776", "0.523", "0.1746" ], [ "0.693", "0.1621", "0.729", "0.1621", "0.729", "0.1776", "0.693", "0.1776" ], [ "0.7983", "0.1621", "0.839", "0.1621", "0.839", "0.1787", "0.7983", "0.1787" ], [ "0.1351", "0.1746", "0.4836", "0.1746", "0.4836", "0.1853", "0.1351", "0.1853" ], [ "0.586", "0.1644", "0.6533", "0.1644", "0.6533", "0.1758", "0.586", "0.1758" ], [ "0.126", "0.1787", "0.1368", "0.1787", "0.1368", "0.1841", "0.126", "0.1841" ], [ "0.7173", "0.1812", "0.7524", "0.1812", "0.7524", "0.1965", "0.7173", "0.1965" ], [ "0.802", "0.1812", "0.843", "0.1812", "0.843", "0.1965", "0.802", "0.1965" ], [ "0.586", "0.1829", "0.6514", "0.1829", "0.6514", "0.1948", "0.586", "0.1948" ], [ "0.546", "0.1829", "0.58", "0.1829", "0.58", "0.1953", "0.546", "0.1953" ], [ "0.12006", "0.1895", "0.486", "0.1895", "0.486", "0.2031", "0.12006", "0.2031" ], [ "0.756", "0.1996", "0.79", "0.1996", "0.79", "0.2156", "0.756", "0.2156" ], [ "0.831", "0.199", "0.8716", "0.199", "0.8716", "0.2156", "0.831", "0.2156" ], [ "0.586", "0.2019", "0.6533", "0.2019", "0.6533", "0.2131", "0.586", "0.2131" ], [ "0.12177", "0.2048", "0.4878", "0.2048", "0.4878", "0.2185", "0.12177", "0.2185" ], [ "0.5366", "0.2019", "0.58", "0.2019", "0.58", "0.2131", "0.5366", "0.2131" ], [ "0.5767", "0.2185", "0.6533", "0.2185", "0.6533", "0.2328", "0.5767", "0.2328" ], [ "0.763", "0.2185", "0.831", "0.2185", "0.831", "0.234", "0.763", "0.234" ], [ "0.544", "0.2197", "0.58", "0.2197", "0.58", "0.2322", "0.544", "0.2322" ], [ "0.12006", "0.2214", "0.4878", "0.2214", "0.4878", "0.2351", "0.12006", "0.2351" ], [ "0.12006", "0.2382", "0.4878", "0.2382", "0.4878", "0.2517", "0.12006", "0.2517" ], [ "0.586", "0.2382", "0.6533", "0.2382", "0.6533", "0.2494", "0.586", "0.2494" ], [ "0.754", "0.2351", "0.8345", "0.2351", "0.8345", "0.2542", "0.754", "0.2542" ], [ "0.12006", "0.2542", "0.4878", "0.2542", "0.4878", "0.2673", "0.12006", "0.2673" ], [ "0.5156", "0.2393", "0.581", "0.2393", "0.581", "0.2505", "0.5156", "0.2505" ], [ "0.5146", "0.2683", "0.6035", "0.2683", "0.6035", "0.2798", "0.5146", "0.2798" ], [ "0.5996", "0.2673", "0.882", "0.2673", "0.882", "0.2808", "0.5996", "0.2808" ], [ "0.511", "0.2798", "0.882", "0.2798", "0.882", "0.2935", "0.511", "0.2935" ], [ "0.12006", "0.2695", "0.4878", "0.2695", "0.4878", "0.2832", "0.12006", "0.2832" ], [ "0.12006", "0.2852", "0.413", "0.2852", "0.413", "0.2986", "0.12006", "0.2986" ], [ "0.51", "0.2903", "0.817", "0.2903", "0.817", "0.3113", "0.51", "0.3113" ], [ "0.1377", "0.3018", "0.4878", "0.3018", "0.4878", "0.3164", "0.1377", "0.3164" ], [ "0.804", "0.2976", "0.881", "0.2976", "0.881", "0.308", "0.804", "0.308" ], [ "0.11755", "0.3135", "0.4902", "0.3135", "0.4902", "0.3342", "0.11755", "0.3342" ], [ "0.513", "0.31", "0.881", "0.31", "0.881", "0.3237", "0.513", "0.3237" ], [ "0.1335", "0.33", "0.492", "0.3289", "0.492", "0.3508", "0.1335", "0.3516" ], [ "0.513", "0.3247", "0.8833", "0.3247", "0.8833", "0.3384", "0.513", "0.3384" ], [ "0.51", "0.3374", "0.8623", "0.3384", "0.8623", "0.3528", "0.51", "0.3516" ], [ "0.12006", "0.3508", "0.4878", "0.3508", "0.4878", "0.3652", "0.12006", "0.3652" ], [ "0.12177", "0.3677", "0.4878", "0.3677", "0.4878", "0.3813", "0.12177", "0.3813" ], [ "0.5146", "0.3682", "0.5425", "0.3682", "0.5425", "0.3801", "0.5146", "0.3801" ], [ "0.559", "0.3665", "0.7446", "0.3665", "0.7446", "0.3801", "0.559", "0.3801" ], [ "0.1184", "0.3843", "0.4878", "0.3843", "0.4878", "0.398", "0.1184", "0.398" ], [ "0.5146", "0.3896", "0.8833", "0.3896", "0.8833", "0.4026", "0.5146", "0.4026" ], [ "0.1184", "0.401", "0.4878", "0.3984", "0.4878", "0.412", "0.1184", "0.4146" ], [ "0.5146", "0.4062", "0.881", "0.4062", "0.881", "0.4192", "0.5146", "0.4192" ], [ "0.12006", "0.4146", "0.4878", "0.415", "0.4878", "0.43", "0.12006", "0.4287" ], [ "0.5146", "0.4216", "0.8833", "0.4216", "0.8833", "0.4353", "0.5146", "0.4353" ], [ "0.12177", "0.4329", "0.2183", "0.4329", "0.2183", "0.4448", "0.12177", "0.4448" ], [ "0.509", "0.434", "0.665", "0.43", "0.6665", "0.4548", "0.51", "0.459" ], [ "0.6465", "0.4329", "0.7983", "0.4329", "0.7983", "0.4548", "0.6465", "0.4548" ], [ "0.723", "0.4424", "0.733", "0.4424", "0.733", "0.4465", "0.723", "0.4465" ], [ "0.1276", "0.4578", "0.2418", "0.4578", "0.2418", "0.4702", "0.1276", "0.4702" ], [ "0.2334", "0.4573", "0.4878", "0.4573", "0.4878", "0.4702", "0.2334", "0.4702" ], [ "0.1184", "0.4727", "0.486", "0.4727", "0.486", "0.4863", "0.1184", "0.4863" ], [ "0.6978", "0.4673", "0.7173", "0.4673", "0.7173", "0.481", "0.6978", "0.481" ], [ "0.617", "0.4785", "0.671", "0.485", "0.6675", "0.4995", "0.614", "0.4934" ], [ "0.692", "0.478", "0.781", "0.4727", "0.784", "0.4963", "0.6943", "0.502" ], [ "0.6675", "0.485", "0.6987", "0.485", "0.6987", "0.4963", "0.6675", "0.4963" ], [ "0.8555", "0.481", "0.885", "0.481", "0.885", "0.4976", "0.8555", "0.4976" ], [ "0.12006", "0.4893", "0.4878", "0.4893", "0.4878", "0.503", "0.12006", "0.503" ], [ "0.692", "0.4988", "0.721", "0.4988", "0.721", "0.511", "0.692", "0.511" ], [ "0.12006", "0.505", "0.4895", "0.505", "0.4895", "0.5186", "0.12006", "0.5186" ], [ "0.12006", "0.5215", "0.4895", "0.5215", "0.4895", "0.535", "0.12006", "0.535" ], [ "0.5156", "0.525", "0.8833", "0.525", "0.8833", "0.538", "0.5156", "0.538" ], [ "0.12006", "0.5376", "0.4878", "0.5376", "0.4878", "0.5503", "0.12006", "0.5503" ], [ "0.513", "0.5405", "0.882", "0.5405", "0.882", "0.554", "0.513", "0.554" ], [ "0.12006", "0.5527", "0.4878", "0.5527", "0.4878", "0.5664", "0.12006", "0.5664" ], [ "0.5146", "0.557", "0.882", "0.557", "0.882", "0.571", "0.5146", "0.571" ], [ "0.1184", "0.5684", "0.486", "0.5684", "0.486", "0.582", "0.1184", "0.582" ], [ "0.5146", "0.5737", "0.882", "0.5737", "0.882", "0.5874", "0.5146", "0.5874" ], [ "0.12006", "0.585", "0.4895", "0.585", "0.4895", "0.5986", "0.12006", "0.5986" ], [ "0.5156", "0.589", "0.8833", "0.589", "0.8833", "0.6025", "0.5156", "0.6025" ], [ "0.12006", "0.6016", "0.4878", "0.6016", "0.4878", "0.615", "0.12006", "0.615" ], [ "0.5156", "0.6055", "0.8833", "0.6055", "0.8833", "0.619", "0.5156", "0.619" ], [ "0.12006", "0.6177", "0.451", "0.6177", "0.451", "0.6313", "0.12006", "0.6313" ], [ "0.5156", "0.6216", "0.7188", "0.6216", "0.7188", "0.6353", "0.5156", "0.6353" ], [ "0.11755", "0.6455", "0.1587", "0.6455", "0.1587", "0.659", "0.11755", "0.659" ], [ "0.1511", "0.6465", "0.2225", "0.6465", "0.2225", "0.6577", "0.1511", "0.6577" ], [ "0.519", "0.6465", "0.6885", "0.6465", "0.6885", "0.66", "0.519", "0.66" ], [ "0.679", "0.648", "0.8833", "0.648", "0.8833", "0.661", "0.679", "0.661" ], [ "0.5156", "0.6646", "0.671", "0.6646", "0.671", "0.6777", "0.5156", "0.6777" ], [ "0.12006", "0.6704", "0.4878", "0.6704", "0.4878", "0.684", "0.12006", "0.684" ], [ "0.12006", "0.686", "0.4878", "0.686", "0.4878", "0.6997", "0.12006", "0.6997" ], [ "0.6987", "0.6934", "0.714", "0.6934", "0.714", "0.703", "0.6987", "0.703" ], [ "0.12006", "0.7017", "0.471", "0.7017", "0.471", "0.7153", "0.12006", "0.7153" ], [ "0.629", "0.705", "0.687", "0.705", "0.687", "0.72", "0.629", "0.72" ], [ "0.687", "0.7017", "0.769", "0.7017", "0.769", "0.7197", "0.687", "0.7197" ], [ "0.8555", "0.705", "0.885", "0.705", "0.885", "0.7217", "0.8555", "0.7217" ], [ "0.12177", "0.7188", "0.486", "0.7188", "0.486", "0.7324", "0.12177", "0.7324" ], [ "0.692", "0.7227", "0.7197", "0.7227", "0.7197", "0.7354", "0.692", "0.7354" ], [ "0.1351", "0.734", "0.4878", "0.734", "0.4878", "0.7476", "0.1351", "0.7476" ], [ "0.12177", "0.7505", "0.4878", "0.7505", "0.4878", "0.764", "0.12177", "0.764" ], [ "0.513", "0.7476", "0.882", "0.749", "0.882", "0.763", "0.513", "0.7627" ], [ "0.1184", "0.7656", "0.486", "0.7656", "0.486", "0.7793", "0.1184", "0.7793" ], [ "0.513", "0.764", "0.882", "0.764", "0.882", "0.778", "0.513", "0.778" ], [ "0.1184", "0.782", "0.4844", "0.782", "0.4844", "0.796", "0.1184", "0.796" ], [ "0.513", "0.78", "0.6987", "0.78", "0.6987", "0.7935", "0.513", "0.7935" ], [ "0.12006", "0.799", "0.486", "0.799", "0.486", "0.8135", "0.12006", "0.8135" ], [ "0.5156", "0.8086", "0.5425", "0.8086", "0.5425", "0.821", "0.5156", "0.821" ], [ "0.544", "0.811", "0.561", "0.811", "0.561", "0.8184", "0.544", "0.8184" ], [ "0.5557", "0.807", "0.7246", "0.8086", "0.7246", "0.8237", "0.5557", "0.822" ], [ "0.12006", "0.8154", "0.3418", "0.8154", "0.3418", "0.829", "0.12006", "0.829" ], [ "0.1184", "0.836", "0.4878", "0.836", "0.4878", "0.8496", "0.1184", "0.8496" ], [ "0.5146", "0.8306", "0.882", "0.8306", "0.882", "0.844", "0.5146", "0.844" ], [ "0.12006", "0.8486", "0.4902", "0.8486", "0.4902", "0.8623", "0.12006", "0.8623" ], [ "0.5156", "0.847", "0.882", "0.847", "0.882", "0.861", "0.5156", "0.861" ], [ "0.1184", "0.86", "0.4895", "0.86", "0.4895", "0.8735", "0.1184", "0.8735" ], [ "0.5146", "0.8633", "0.8833", "0.8633", "0.8833", "0.8774", "0.5146", "0.8774" ], [ "0.1184", "0.8716", "0.4895", "0.8716", "0.4895", "0.8853", "0.1184", "0.8853" ], [ "0.1184", "0.8843", "0.4895", "0.8843", "0.4895", "0.898", "0.1184", "0.898" ], [ "0.5146", "0.879", "0.882", "0.879", "0.882", "0.8926", "0.5146", "0.8926" ], [ "0.12006", "0.8965", "0.1646", "0.8965", "0.1646", "0.9077", "0.12006", "0.9077" ], [ "0.5156", "0.8955", "0.882", "0.8955", "0.882", "0.909", "0.5156", "0.909" ] ]

[ "F1Scores", "Expected layer &center-of-gravity", "(Devlin etal.,", "BERT.BERT模型", "1=0=240", "2", "46810121416", "88.596.7", "2019）在许多任务上都表现出了最先", "POS", "3.39", "11.68", "结构（Vaswanietal.2017）是许多最", "进的性能，其深层的Transformer体系", "Consts.", "3.79", "13.06", "新模型（?)典型代表[e。g。JD", "73.687.0", "rad-", "Deps.", "5.69", "13.75", "ford2018improving,radford2019language,liu2019m", "85.695.5", "Entities", "4.64", "13.16", "我们专注于股票BERT模型（基本和大型", "90.696.1", "", "6.54", "13.63", "81.391.4", "SRL", "无案例），这些模型通过3个任务的多任务", "9.47", "15.80", "80.591.9", "目标（掩蔽语言建模和下一句预测）进行训", "Coref.", "77.783.7", "9.9312.72", "SPR", "练。3B单词英语语料库。由于我们想了解网", "络如何表示语言作为预训练的结果，我们遵循", "60.784.2", "9.4012.83", "Tenneyetal.（2019）（偏离标准BERT用法）", "Relations", "Figure l:", "Summary statistics on BERT-large.", "Columns on left show F1 dev-set scores for the", "并冻结编码器权重。这可以防止编码器重新排", "列其内部表示以更好地适应探测任务", "baseline ( P(o)) and full-model ( P()", "给定输人令牌T=[to,ti.,tn]】，深度", ")probes.", "编码器产生一组层激活H(O),H(1)..,H(L)", "Dark (blue) are the mixing weight center of grav-", "其中H(e)=[h,h,..,h]是eth编", "ity (equation 2 ); light (purple) are the expected", "layer from the cumulative scores (equation 4).", "码器层的激活向量，H(O）对应于非上下文", "单词（片段）嵌入。我们使用跨层的加权和", "3.1", "Scalar Mixing Weights", "（S3.1）将它们合并为一组每个令牌表示向量", "要跨层合并，我们使用ELMo模型引人的标", "H=[ho,hi,...,hn]并使用Tenneyet al.", "量混合技术。根据Petersetal.（2018a）的方", "(2019）的体系结构和过程为每个任务训练探", "程式E（1），对于每个任务，我们引入标量参", "测分类器P", "数和", "a，并让", "", "Limitations", "这项工作旨在探索。我们专注", "于一个特定的编码器-BERT-探索如何在深度", "L", "hi,r=", "Es(h(9)", "T", "(1)", "语言模型中组织信息，并且需要进一步的工作", "l=0", "来确定趋势在多大程度上保持一般。此外，我", "们的工作带有所有基于检查的探测的局限性：", "其中sr=softmax(a）。我们与探测分类", "我们的探测分类器没有观察到语言模式这一事", "器P一起学习这些权重，以便允许它从编码", "实并不能保证它不存在，并且对模式的观察并", "器的许多层中提取信息，而无需添加大量参", "不能告诉我们如何使用它。因此，我们强调将", "数。在训练了探测模型之后，我们提取学习的", "结构分析与行为研究相结合的重要性（如SE", "系数，以估计不同层对该特定任务的贡献。我", "1中所述），以更全面地了解这些模型编码的", "们将较高的权重解释为相应层包含与该特定任", "信息以及该信息如何影响下游任务的性能", "务相关的更多信息的证据", "3", "Metrics", "Center-of-Gravity.", "作为汇总统计，我们将混", "合权重重心定义为：", "我们定义了两个互补的指标。首先，标量混合", "权重（S3.1）告诉我们，当探测分类器可以访", "L", "问整个BERT模型时，哪些层最相关。第", "E,[e] =", "Tl. s(e)", "(2)", "个，累积评分（S3.2）告诉我们，通过引入每", "l=0", "层，我们可以在探测任务上获得多少更高的", "分数。这些指标提供了关于模型内部正在发生", "这反映了每个任务所关注的平均层；直观地，", "的事情的补充观点。混合权重仅从训练数据中", "我们可以解释一个更高的值，表示该任务所需", "学习-它们告诉我们探测模型发现哪些层最有", "的信息被更高的层捕获", "用。相反，累积评分完全来自评估集，并告诉", "3.2", "", "Cumulative Scoring", "我们正确的预测需要多少层", "jsalt18-sentence-repl/jiant . Dependencies is the En-", "我们想估计在编码器中的哪一层可以正确预", "glish Web Treebank (Silveira et al., 2014) , SPR is the", "测目标（s1,S2,label）。混合权重不能直接告诉", "SPR1 dataset of (Teichert et al., 2017) , and relations", "我们这一点，因为它们是作为参数学习的，并", "is SemEval 2010 Task 8 (Hendrickx et al., 2009) : All", "other tasks are from OntoNotes 5.0 (Weischedel et al.,", "不对应于数据的分布。单层的朴素分类器也不", "2013)", "能，因为有关特定跨度的信息可能会散布在多" ]

{ "arxiv_id": "1905.05950", "language": "zh", "page_idx": 1 }

[ [ "0.409", "0.076", "0.4836", "0.0736", "0.4836", "0.0885", "0.409", "0.0903" ], [ "0.5156", "0.0766", "0.818", "0.0766", "0.818", "0.0903", "0.5156", "0.0903" ], [ "0.12177", "0.0873", "0.1377", "0.0873", "0.1377", "0.10986", "0.12177", "0.10986" ], [ "0.2141", "0.0778", "0.2854", "0.076", "0.2854", "0.0873", "0.2141", "0.08905" ], [ "0.293", "0.0932", "0.3315", "0.0932", "0.3315", "0.1051", "0.293", "0.1051" ], [ "0.5483", "0.1027", "0.8296", "0.1027", "0.8296", "0.12054", "0.5483", "0.12054" ], [ "0.8564", "0.1069", "0.881", "0.1069", "0.881", "0.1194", "0.8564", "0.1194" ], [ "0.2141", "0.1289", "0.287", "0.1289", "0.287", "0.1437", "0.2141", "0.1437" ], [ "0.4106", "0.1289", "0.4836", "0.1289", "0.4836", "0.1437", "0.4106", "0.1437" ], [ "0.12427", "0.1372", "0.1351", "0.1372", "0.1351", "0.155", "0.12427", "0.155" ], [ "0.5225", "0.136", "0.882", "0.136", "0.882", "0.1497", "0.5225", "0.1497" ], [ "0.1234", "0.1526", "0.1351", "0.1526", "0.1351", "0.1687", "0.1234", "0.1687" ], [ "0.517", "0.1509", "0.882", "0.152", "0.882", "0.1663", "0.517", "0.165" ], [ "0.2822", "0.1687", "0.3208", "0.1687", "0.3208", "0.1758", "0.2822", "0.1758" ], [ "0.517", "0.1675", "0.881", "0.1675", "0.881", "0.1812", "0.517", "0.1812" ], [ "0.215", "0.1829", "0.2854", "0.1829", "0.2854", "0.1948", "0.215", "0.1948" ], [ "0.4114", "0.1829", "0.4802", "0.1829", "0.4802", "0.1948", "0.4114", "0.1948" ], [ "0.513", "0.1829", "0.882", "0.1829", "0.882", "0.1965", "0.513", "0.1965" ], [ "0.12006", "0.1948", "0.1351", "0.1912", "0.1394", "0.2037", "0.12427", "0.206" ], [ "0.12006", "0.2031", "0.1377", "0.2031", "0.1377", "0.2197", "0.12006", "0.2197" ], [ "0.5156", "0.2019", "0.615", "0.2019", "0.615", "0.2131", "0.5156", "0.2131" ], [ "0.2747", "0.2173", "0.34", "0.2173", "0.34", "0.2286", "0.2747", "0.2286" ], [ "0.4106", "0.234", "0.482", "0.234", "0.482", "0.2488", "0.4106", "0.2488" ], [ "0.6055", "0.234", "0.785", "0.2185", "0.79", "0.2446", "0.6094", "0.26" ], [ "0.2124", "0.2351", "0.2854", "0.2328", "0.2854", "0.2477", "0.2124", "0.2494" ], [ "0.1184", "0.2393", "0.1394", "0.2393", "0.1394", "0.2766", "0.1184", "0.2766" ], [ "0.858", "0.2405", "0.8833", "0.2405", "0.8833", "0.2542", "0.858", "0.2542" ], [ "0.6885", "0.2488", "0.777", "0.2452", "0.7773", "0.2683", "0.69", "0.2715" ], [ "0.3728", "0.2559", "0.4434", "0.2559", "0.4434", "0.26", "0.3728", "0.26" ], [ "0.3376", "0.2778", "0.3687", "0.2778", "0.3687", "0.282", "0.3376", "0.282" ], [ "0.513", "0.2798", "0.882", "0.2798", "0.882", "0.2935", "0.513", "0.2935" ], [ "0.215", "0.2893", "0.2837", "0.2893", "0.2837", "0.3005", "0.215", "0.3005" ], [ "0.4114", "0.2893", "0.4802", "0.2893", "0.4802", "0.3005", "0.4114", "0.3005" ], [ "0.12006", "0.3005", "0.1377", "0.3005", "0.1377", "0.3218", "0.12006", "0.3218" ], [ "0.5156", "0.2976", "0.882", "0.2976", "0.882", "0.3113", "0.5156", "0.3113" ], [ "0.3914", "0.3088", "0.398", "0.3088", "0.398", "0.3123", "0.3914", "0.3123" ], [ "0.5146", "0.3135", "0.6206", "0.3135", "0.6206", "0.328", "0.5146", "0.328" ], [ "0.2334", "0.3206", "0.3535", "0.3206", "0.3535", "0.335", "0.2334", "0.335" ], [ "0.215", "0.3416", "0.2854", "0.3416", "0.2854", "0.3528", "0.215", "0.3528" ], [ "0.413", "0.3416", "0.482", "0.3416", "0.482", "0.3528", "0.413", "0.3528" ], [ "0.5146", "0.3433", "0.5283", "0.3433", "0.5283", "0.3516", "0.5146", "0.3516" ], [ "0.546", "0.3403", "0.617", "0.3403", "0.617", "0.3552", "0.546", "0.3552" ], [ "0.12177", "0.3486", "0.1377", "0.3486", "0.1377", "0.3682", "0.12177", "0.3682" ], [ "0.126", "0.3635", "0.1394", "0.3682", "0.1301", "0.379", "0.11755", "0.3735" ], [ "0.392", "0.361", "0.424", "0.361", "0.424", "0.3694", "0.392", "0.3694" ], [ "0.2284", "0.3718", "0.4106", "0.3723", "0.4106", "0.3901", "0.2284", "0.3896" ], [ "0.5156", "0.3665", "0.885", "0.3665", "0.885", "0.3801", "0.5156", "0.3801" ], [ "0.1973", "0.376", "0.2124", "0.376", "0.2124", "0.379", "0.1973", "0.379" ], [ "0.5146", "0.3818", "0.881", "0.3813", "0.881", "0.3955", "0.5146", "0.3967" ], [ "0.2141", "0.3926", "0.287", "0.3926", "0.287", "0.4067", "0.2141", "0.4067" ], [ "0.4106", "0.3938", "0.482", "0.3914", "0.482", "0.4062", "0.4106", "0.408" ], [ "0.12006", "0.4026", "0.1377", "0.4026", "0.1377", "0.4275", "0.12006", "0.4275" ], [ "0.513", "0.398", "0.86", "0.3984", "0.86", "0.4133", "0.513", "0.412" ], [ "0.513", "0.4133", "0.882", "0.4146", "0.882", "0.4287", "0.513", "0.4275" ], [ "0.1545", "0.4312", "0.2217", "0.4275", "0.2225", "0.4395", "0.1561", "0.4424" ], [ "0.2107", "0.4312", "0.318", "0.4287", "0.318", "0.4407", "0.2107", "0.4424" ], [ "0.5146", "0.4312", "0.7827", "0.4312", "0.7827", "0.4448", "0.5146", "0.4448" ], [ "0.12427", "0.4478", "0.1351", "0.4478", "0.1351", "0.4714", "0.12427", "0.4714" ], [ "0.215", "0.4478", "0.2854", "0.4478", "0.2854", "0.459", "0.215", "0.459" ], [ "0.4106", "0.4453", "0.482", "0.4453", "0.482", "0.4602", "0.4106", "0.4602" ], [ "0.5205", "0.456", "0.882", "0.456", "0.882", "0.4697", "0.5205", "0.4697" ], [ "0.3552", "0.4673", "0.4634", "0.4673", "0.4634", "0.4714", "0.3552", "0.4714" ], [ "0.1276", "0.4739", "0.1377", "0.4768", "0.1318", "0.4863", "0.12177", "0.4827" ], [ "0.5156", "0.4727", "0.882", "0.4727", "0.882", "0.4863", "0.5156", "0.4863" ], [ "0.12427", "0.4827", "0.1335", "0.4827", "0.1335", "0.4868", "0.12427", "0.4868" ], [ "0.2091", "0.4785", "0.2544", "0.4822", "0.2527", "0.4934", "0.2083", "0.4905" ], [ "0.3064", "0.484", "0.3225", "0.484", "0.3225", "0.4912", "0.3064", "0.4912" ], [ "0.346", "0.4827", "0.3972", "0.4827", "0.3972", "0.4934", "0.346", "0.4934" ], [ "0.5146", "0.4868", "0.8833", "0.488", "0.8833", "0.503", "0.5146", "0.502" ], [ "0.1436", "0.4954", "0.4878", "0.4963", "0.4878", "0.5083", "0.1436", "0.5073" ], [ "0.5146", "0.503", "0.882", "0.5044", "0.882", "0.5186", "0.5146", "0.517" ], [ "0.1184", "0.5195", "0.4895", "0.5186", "0.4895", "0.533", "0.1184", "0.5337" ], [ "0.5146", "0.5195", "0.881", "0.5195", "0.881", "0.5337", "0.5146", "0.5337" ], [ "0.12177", "0.535", "0.4895", "0.535", "0.4895", "0.55", "0.12177", "0.55" ], [ "0.5156", "0.536", "0.882", "0.536", "0.882", "0.55", "0.5156", "0.55" ], [ "0.1184", "0.5503", "0.486", "0.5503", "0.486", "0.5654", "0.1184", "0.5654" ], [ "0.5146", "0.5527", "0.882", "0.5527", "0.882", "0.5664", "0.5146", "0.5664" ], [ "0.12006", "0.5684", "0.4878", "0.5684", "0.4878", "0.579", "0.12006", "0.579" ], [ "0.5156", "0.568", "0.7715", "0.5684", "0.7715", "0.583", "0.5156", "0.582" ], [ "0.12006", "0.582", "0.1604", "0.582", "0.1604", "0.5933", "0.12006", "0.5933" ], [ "0.5356", "0.585", "0.882", "0.585", "0.882", "0.5986", "0.5356", "0.5986" ], [ "0.5156", "0.5996", "0.882", "0.601", "0.882", "0.615", "0.5156", "0.6143" ], [ "0.1184", "0.6143", "0.4895", "0.6143", "0.4895", "0.629", "0.1184", "0.629" ], [ "0.5156", "0.6177", "0.882", "0.6177", "0.882", "0.6313", "0.5156", "0.6313" ], [ "0.12006", "0.632", "0.413", "0.632", "0.413", "0.6455", "0.12006", "0.6455" ], [ "0.519", "0.633", "0.882", "0.633", "0.882", "0.6465", "0.519", "0.6465" ], [ "0.1411", "0.6494", "0.486", "0.6494", "0.486", "0.663", "0.1411", "0.663" ], [ "0.5156", "0.6494", "0.882", "0.6494", "0.882", "0.663", "0.5156", "0.663" ], [ "0.1184", "0.6665", "0.486", "0.6665", "0.486", "0.681", "0.1184", "0.681" ], [ "0.5156", "0.6665", "0.882", "0.6665", "0.882", "0.6797", "0.5156", "0.6797" ], [ "0.1184", "0.6816", "0.4902", "0.6816", "0.4902", "0.703", "0.1184", "0.703" ], [ "0.5156", "0.6816", "0.882", "0.6816", "0.882", "0.6953", "0.5156", "0.6953" ], [ "0.12177", "0.703", "0.4878", "0.703", "0.4878", "0.7173", "0.12177", "0.7173" ], [ "0.5156", "0.6978", "0.882", "0.6978", "0.882", "0.712", "0.5156", "0.712" ], [ "0.11755", "0.7144", "0.476", "0.7188", "0.476", "0.7393", "0.11755", "0.7354" ], [ "0.517", "0.7144", "0.882", "0.7144", "0.882", "0.728", "0.517", "0.728" ], [ "0.517", "0.73", "0.882", "0.73", "0.882", "0.7437", "0.517", "0.7437" ], [ "0.1377", "0.7407", "0.4895", "0.7407", "0.4895", "0.7544", "0.1377", "0.7544" ], [ "0.11755", "0.749", "0.492", "0.755", "0.492", "0.7793", "0.11755", "0.7725" ], [ "0.5156", "0.745", "0.882", "0.745", "0.882", "0.7603", "0.5156", "0.7603" ], [ "0.5156", "0.7627", "0.7", "0.7627", "0.7", "0.7754", "0.5156", "0.7754" ], [ "0.1184", "0.771", "0.486", "0.774", "0.486", "0.7915", "0.1184", "0.788" ], [ "0.519", "0.785", "0.8833", "0.786", "0.8833", "0.8003", "0.519", "0.8" ], [ "0.12177", "0.7915", "0.4878", "0.7915", "0.4878", "0.8047", "0.12177", "0.8047" ], [ "0.11755", "0.807", "0.4895", "0.804", "0.4895", "0.822", "0.11755", "0.8247" ], [ "0.5156", "0.8027", "0.8833", "0.8027", "0.8833", "0.818", "0.5156", "0.818" ], [ "0.12006", "0.8247", "0.486", "0.8247", "0.486", "0.8384", "0.12006", "0.8384" ], [ "0.5156", "0.8193", "0.882", "0.8193", "0.882", "0.833", "0.5156", "0.833" ], [ "0.5156", "0.8345", "0.882", "0.835", "0.882", "0.8496", "0.5156", "0.8486" ], [ "0.1377", "0.858", "0.4895", "0.8594", "0.4895", "0.8735", "0.1377", "0.8726" ], [ "0.532", "0.857", "0.885", "0.8594", "0.885", "0.8735", "0.532", "0.8716" ], [ "0.1184", "0.8716", "0.4895", "0.8716", "0.4895", "0.8853", "0.1184", "0.8853" ], [ "0.513", "0.8726", "0.763", "0.8726", "0.763", "0.883", "0.513", "0.883" ], [ "0.1184", "0.8843", "0.4902", "0.8843", "0.4902", "0.898", "0.1184", "0.898" ], [ "0.532", "0.882", "0.885", "0.883", "0.885", "0.898", "0.532", "0.8965" ], [ "0.1184", "0.898", "0.2032", "0.898", "0.2032", "0.9077", "0.1184", "0.9077" ], [ "0.5146", "0.8955", "0.7524", "0.8965", "0.7524", "0.91", "0.5146", "0.909" ] ]

[ "K(s)=0.19", "器层e，我们在探测任务上做得更好：", "SOd", "K(A)=1.60", "", "△(e) = Score(P(e)) - Score(P(e-1))", "(3)", "K(△) = 1.57", "K(s)=0.83", "", "ExpectedLayer同样，我们计算差分分数", "Poo", "上的（伪）4期望作为汇总统计量。为了关注", "", "上下文编码器层的行为，我们省略了在0层解", "K(A)=1.15", "K(s) = 0.87", "析的“琐碎”示例以及完整模型中剩余的净空", "0", "0", "的贡献。让：", "", "K(s)=0.06", "E[4] =- D1 L. 4(9]", "K(A)=1.61", "", "(4)", "E14()", "", "", "这可以被认为是，近似地，探测模型正确标记", "K(A)=1.31", "K(s)=0.46", "S", "示例的预期层，假设示例在编码器的某个层", "", "l≥1处解析", "", "K(A)=0.60", "K(s) =0.50", "4", "Results", "e", "3", "", "", "图1报告汇总统计信息和F1绝对得分，图2", "", "报告每层指标。两者都显示了24层BERT-", "K(△)= 0.33", "K(s)=0.01", "", "large模型的结果。我们还报告了K(+)", "KL(+l/Uniform），以估计每个任务的每个统", "", "O", "计量（=S,△）的非均匀性 5", "suo", "K(△)=0.50", "K(s) = 0.01", "LinguisticPatterns．我们在两个指标中都观", "", "0", "察到了一致的趋势，任务以自然的顺序编码：", "", "", "口", "", "POS标签处理最早，其次是成分，依赖关系，", "024681012141618202224", "语义角色和共同引用。也就是说，似乎基本的", "Figure 2: Layer-wise metrics on BERT-large. Solid", "语法信息在网络中出现得更早，而高级语义信", "(blue) are mixing weights s@) ( $3.1 ); outlined", "息出现在更高的层。我们注意到，这一发现与", "(purple) are differential scores () ( $3.2 ), nor-", "Petersetal.（2018b）的初步观察结果一致，后", "malized for each task. Horizontal axis is encoder", "者发现成分的表示早于共同引用", "layer.", "此外，我们观察到，一般而言，语法信息的", "本地化程度更高，与语法任务相关的权重往往", "个层上，并且如Petersetal.（2018b）中所示，", "集中在几层（高K(s）和K(△））上，而与语", "编码器可能会选择丢弃较高层的信息", "义任务相关的信息通常分布在整个网络中。例", "为了解决这个问题，我们训练了一系列分类", "如，我们发现对于语义关系和原始角色（SPR）", "器【P(e：它们使用标量混合（等式E1）来", "混合权重接近统一，并且这些任务的非平凡示", "处理层以及所有先前的层。P(O）对应于仅", "例在几乎所有层上都逐渐得到解决。对于实体", "使用一袋单词（片）嵌入的非上下文基线，而", "标记，在第1层中解析了许多示例，但此后", "P(C)=P对应于探测BERT模型的所有层", "有一个长尾，并且在高层中仅混合权重的浓度", "很低。需要进一步研究以确定这是因为BERT", "这些分类器是累积的，从某种意义上说，", "p(+1)具有相似数量的参数，但可以访问比", "难以表示这些任务的正确抽象，还是因为语义", "信息本质上难以本地化", "P(e）严格更多的信息，并且我们直观地看到", "ComparisonofMetrics.对于许多任务，我", "性能（F1分数）通常会随着添加更多层而增", "加。3然后，我们可以计算一个差分分数△(9)", "们发现差异得分在模型的前几层（BERT-large", "，它可以衡量如果我们观察到一个额外的编码", "的第1-7层）中最高，i。e。大多数例子可以", "很早就正确分类。我们将此归因于启发式快捷", "3Note that if a new layer provides distracting fea-", "4This is not a true expectation because the F1 score", "tures, the probing model can overfit and performance", "is not an expectation over examples.", "can drop. We see this in particular in the last 1-2 layers", "5KL(+|Uniform) = -H(+) + Constant , so higher", "(Figure 2).", "values correspond to lower entropy" ]

{ "arxiv_id": "1905.05950", "language": "zh", "page_idx": 2 }

[ [ "0.1462", "0.05176", "0.5186", "0.0505", "0.5186", "0.0625", "0.1462", "0.0631" ], [ "0.607", "0.04987", "0.851", "0.0505", "0.851", "0.0644", "0.607", "0.0631" ], [ "0.1626", "0.11615", "0.835", "0.11615", "0.835", "0.1307", "0.1626", "0.1307" ], [ "0.1447", "0.1445", "0.2893", "0.1459", "0.2893", "0.1604", "0.1447", "0.159" ], [ "0.614", "0.1465", "0.853", "0.1465", "0.853", "0.1604", "0.614", "0.1604" ], [ "0.1478", "0.1642", "0.2296", "0.1642", "0.2296", "0.1755", "0.1478", "0.1755" ], [ "0.1478", "0.1799", "0.3906", "0.1799", "0.3906", "0.1945", "0.1478", "0.1945" ], [ "0.1447", "0.1957", "0.5156", "0.1976", "0.5156", "0.2122", "0.1447", "0.2102" ], [ "0.4592", "0.276", "0.5415", "0.276", "0.5415", "0.2903", "0.4592", "0.2903" ], [ "0.179", "0.2966", "0.8203", "0.2966", "0.8203", "0.3113", "0.179", "0.3113" ], [ "0.1772", "0.3113", "0.822", "0.3125", "0.822", "0.327", "0.1772", "0.3257" ], [ "0.1772", "0.327", "0.8203", "0.3257", "0.8203", "0.3403", "0.1772", "0.3416" ], [ "0.179", "0.3423", "0.8203", "0.3423", "0.8203", "0.356", "0.179", "0.356" ], [ "0.1772", "0.356", "0.819", "0.3567", "0.819", "0.3718", "0.1772", "0.3706" ], [ "0.1772", "0.3718", "0.818", "0.3718", "0.818", "0.385", "0.1772", "0.385" ], [ "0.179", "0.387", "0.822", "0.387", "0.822", "0.401", "0.179", "0.401" ], [ "0.179", "0.4028", "0.818", "0.4028", "0.818", "0.4167", "0.179", "0.4167" ], [ "0.179", "0.4185", "0.822", "0.4185", "0.822", "0.432", "0.179", "0.432" ], [ "0.1772", "0.432", "0.8203", "0.433", "0.8203", "0.4463", "0.1772", "0.4458" ], [ "0.179", "0.4475", "0.8203", "0.4475", "0.8203", "0.4622", "0.179", "0.4622" ], [ "0.1772", "0.4634", "0.819", "0.4634", "0.819", "0.4766", "0.1772", "0.4766" ], [ "0.1772", "0.4792", "0.8203", "0.4792", "0.8203", "0.4924", "0.1772", "0.4924" ], [ "0.179", "0.4937", "0.818", "0.4937", "0.818", "0.507", "0.179", "0.507" ], [ "0.1772", "0.5083", "0.8203", "0.5083", "0.8203", "0.5215", "0.1772", "0.5215" ], [ "0.1772", "0.5215", "0.8203", "0.523", "0.8203", "0.537", "0.1772", "0.536" ], [ "0.1772", "0.5386", "0.486", "0.539", "0.486", "0.5537", "0.1772", "0.553" ], [ "0.1447", "0.5757", "0.3015", "0.5757", "0.3015", "0.5903", "0.1447", "0.5903" ], [ "0.1462", "0.603", "0.4656", "0.603", "0.4656", "0.616", "0.1462", "0.616" ], [ "0.4722", "0.603", "0.853", "0.603", "0.853", "0.616", "0.4722", "0.616" ], [ "0.1462", "0.619", "0.4656", "0.619", "0.4656", "0.6333", "0.1462", "0.6333" ], [ "0.6055", "0.622", "0.851", "0.622", "0.851", "0.632", "0.6055", "0.632" ], [ "0.1462", "0.637", "0.4673", "0.637", "0.4673", "0.651", "0.1462", "0.651" ], [ "0.5337", "0.638", "0.851", "0.638", "0.851", "0.6484", "0.5337", "0.6484" ], [ "0.1478", "0.655", "0.4412", "0.655", "0.4412", "0.6655", "0.1478", "0.6655" ], [ "0.451", "0.653", "0.851", "0.653", "0.851", "0.6665", "0.451", "0.6665" ], [ "0.1462", "0.672", "0.4495", "0.672", "0.4495", "0.685", "0.1462", "0.685" ], [ "0.4468", "0.672", "0.853", "0.672", "0.853", "0.685", "0.4468", "0.685" ], [ "0.1462", "0.6895", "0.8545", "0.6895", "0.8545", "0.7026", "0.1462", "0.7026" ], [ "0.1478", "0.7056", "0.853", "0.7056", "0.853", "0.7197", "0.1478", "0.7197" ], [ "0.1478", "0.7227", "0.853", "0.7227", "0.853", "0.7363", "0.1478", "0.7363" ], [ "0.1478", "0.7407", "0.853", "0.7407", "0.853", "0.754", "0.1478", "0.754" ], [ "0.1462", "0.757", "0.853", "0.757", "0.853", "0.771", "0.1462", "0.771" ], [ "0.1478", "0.776", "0.851", "0.776", "0.851", "0.786", "0.1478", "0.786" ], [ "0.1462", "0.7915", "0.853", "0.7915", "0.853", "0.8047", "0.1462", "0.8047" ], [ "0.1478", "0.808", "0.853", "0.808", "0.853", "0.8213", "0.1478", "0.8213" ], [ "0.1462", "0.8247", "0.853", "0.8247", "0.853", "0.839", "0.1462", "0.839" ], [ "0.1478", "0.8423", "0.853", "0.8423", "0.853", "0.857", "0.1478", "0.857" ], [ "0.1462", "0.86", "0.706", "0.86", "0.706", "0.8735", "0.1462", "0.8735" ], [ "0.1748", "0.8774", "0.3774", "0.8774", "0.3774", "0.8916", "0.1748", "0.8916" ], [ "0.1447", "0.9165", "0.2712", "0.9165", "0.2712", "0.9297", "0.1447", "0.9297" ] ]

[ "ProceedingsofMachineLearningResearch148:136-146,2021", "NeurIPS2020PreregistrationWorkshop", "SFTrack十十：一种用于时空一致跟踪的快速可学习光谱分割方法", "Elena Burceanu", "[email protected]", "Bitdefender", "University of Bucharest, Romania", "Institute of Mathematics of the Romanian Academy", "Abstract", "We propose an object tracking method, SFTrack++, that smoothly learns to preserve the", "tracked object consistency over space and time dimensions by taking a spectral clustering", "approach over the graph of pixels from the video, using a fast 3D filtering formulation", "for finding the principal eigenvector of this graph's adjacency matrix. To better capture", "complex aspects of the tracked object, we enrich our formulation to multi-channel inputs.", "which permit different points of view for the same input. The channel inputs are in our", "experiments, the output of multiple tracking methods. After combining them, instead of", "relying only on hidden layers representations to predict a good tracking bounding box", "we explicitly learn an intermediate, more refined one, namely the segmentation map of", "the tracked object. This prevents the rough common bounding box approach to introduce", "noise and distractors in the learning process. We test our method, SFTrack++, on five", "tracking benchmarks: OTB, UAV, NFS, GOT-10k, and TrackingNet, using five top trackers", "as input. Our experimental results validate the pre-registered hypothesis.We obtain", "consistent and robust results, competitive on the three traditional benchmarks (OTB, UAV", "NFS) and significantly on top of others (by over 1.1% on accuracy) on GOT-10k and", "TrackingNet, which are newer, larger, and more varied datasets. The code is available at", "https://github.com/bit-ml/sftrackpp", "1. Introduction", "在视觉任务中更好地使用视频的时间方面", "已经被积极讨论了相当长的时间，特别是在硬件", "方面取得了巨大而持续的进步。我们在方", "个方面是视觉对象跟踪中空间和", "时间维度的无缝混合当前的方法主要依赖", "小观和遂顿处理FlCaellesetal.(2017)：Li", "et al. (2019): Danellian et al. (2019)", "很少有人明确考虑时间一致性EBhatetal.（2020)：", "Jabrietal：（2020b）。在谱图方法中，", "节点是像素，边缘是它们在空间和时间上的局部关", "系，而该图中最强的簇由图的邻接矩阵的主特征向量给出，表示在空间和时间上一致的主", "要对象体积。第二个观察挑战用于跟踪的粗略边界框（bbox）形状。虽然它提供了一种方", "便的方式来注释数据集，但它是一个相当不完美的标签，因为它会导致随着时间的推移而", "累积，传播和放大的错误。对象很少看起来像盒子，而bboxes大多数时候都包含重要的背", "景信息或干扰因素。由于对感兴趣的对象进行良好的分割会直接影响跟踪性能，因此我们", "限制一种中间表示形式，即分割图，其自的是减少从一顿传递到下一顿的噪声量。我们", "将其集成到我们的端到端流程中，如图所示。1。我们强调的第三点是依赖于同一对象的", "多个独立特征或专用于不同方面的多个模块。这具有更好的理解复杂对象的能力，同时提", "高了鲁棒性Bhatetal.（2020）。因此，我们调整SFSeg频谱方法，增强其公式以支持在", "多个输人通道之上的学习。它们可能是来自不同方法的输入顿的一般特征，或者更具体地", "说，跟踪来自多个解决方案的输出，正如我们在Sec中测试的那样。4", "我们的方法的主要贡献是：", "@2021 E.Burceanu." ]

{ "arxiv_id": "2011.13843", "language": "zh", "page_idx": 0 }

[ [ "0.1478", "0.0505", "0.8613", "0.0505", "0.8613", "0.06125", "0.1478", "0.06125" ], [ "0.1666", "0.127", "0.4192", "0.1276", "0.4192", "0.1414", "0.1666", "0.1401" ], [ "0.1847", "0.1843", "0.2507", "0.1843", "0.2507", "0.1945", "0.1847", "0.1945" ], [ "0.6167", "0.207", "0.649", "0.209", "0.648", "0.2172", "0.615", "0.2153" ], [ "0.384", "0.2317", "0.407", "0.2317", "0.407", "0.2361", "0.384", "0.2361" ], [ "0.6177", "0.2405", "0.6465", "0.2405", "0.6465", "0.2488", "0.6177", "0.2488" ], [ "0.367", "0.2676", "0.4192", "0.2676", "0.4192", "0.2747", "0.367", "0.2747" ], [ "0.4697", "0.2708", "0.4958", "0.2708", "0.4958", "0.279", "0.4697", "0.279" ], [ "0.706", "0.2708", "0.7305", "0.2708", "0.7305", "0.279", "0.706", "0.279" ], [ "0.785", "0.2747", "0.819", "0.2747", "0.819", "0.2842", "0.785", "0.2842" ], [ "0.6855", "0.279", "0.751", "0.279", "0.751", "0.286", "0.6855", "0.286" ], [ "0.4534", "0.2798", "0.513", "0.2798", "0.513", "0.2874", "0.4534", "0.2874" ], [ "0.1887", "0.2715", "0.2361", "0.2715", "0.2361", "0.2822", "0.1887", "0.2822" ], [ "0.7812", "0.2822", "0.819", "0.2822", "0.819", "0.2922", "0.7812", "0.2922" ], [ "0.778", "0.2903", "0.818", "0.2886", "0.819", "0.2998", "0.7793", "0.3018" ], [ "0.5415", "0.2998", "0.716", "0.2998", "0.716", "0.31", "0.5415", "0.31" ], [ "0.7754", "0.2986", "0.818", "0.2966", "0.819", "0.308", "0.777", "0.31" ], [ "0.556", "0.31", "0.699", "0.31", "0.699", "0.3176", "0.556", "0.3176" ], [ "0.3218", "0.3547", "0.367", "0.3547", "0.367", "0.3662", "0.3218", "0.3662" ], [ "0.5596", "0.3901", "0.574", "0.3933", "0.5703", "0.4028", "0.5557", "0.3997" ], [ "0.636", "0.3882", "0.6665", "0.387", "0.6665", "0.3945", "0.636", "0.3958" ], [ "0.694", "0.3901", "0.732", "0.3901", "0.732", "0.401", "0.694", "0.401" ], [ "0.3962", "0.3933", "0.4192", "0.3933", "0.4192", "0.401", "0.3962", "0.401" ], [ "0.1772", "0.3914", "0.2467", "0.3914", "0.2467", "0.4016", "0.1772", "0.4016" ], [ "0.634", "0.3984", "0.6685", "0.3984", "0.6685", "0.4028", "0.634", "0.4028" ], [ "0.1544", "0.406", "0.312", "0.406", "0.312", "0.4167", "0.1544", "0.4167" ], [ "0.5825", "0.4072", "0.839", "0.4072", "0.839", "0.4204", "0.5825", "0.4204" ], [ "0.282", "0.4185", "0.3406", "0.4185", "0.3406", "0.4255", "0.282", "0.4255" ], [ "0.1462", "0.4634", "0.8545", "0.4634", "0.8545", "0.4778", "0.1462", "0.4778" ], [ "0.2239", "0.481", "0.8545", "0.481", "0.8545", "0.4956", "0.2239", "0.4956" ], [ "0.2222", "0.4956", "0.851", "0.4976", "0.851", "0.515", "0.2222", "0.513" ], [ "0.2255", "0.515", "0.8545", "0.515", "0.8545", "0.53", "0.2255", "0.53" ], [ "0.2255", "0.5327", "0.8545", "0.5327", "0.8545", "0.547", "0.2255", "0.547" ], [ "0.2255", "0.55", "0.853", "0.55", "0.853", "0.563", "0.2255", "0.563" ], [ "0.2239", "0.5664", "0.8506", "0.565", "0.8506", "0.58", "0.2239", "0.5806" ], [ "0.2222", "0.5835", "0.8555", "0.5835", "0.8555", "0.601", "0.2222", "0.601" ], [ "0.2222", "0.5996", "0.3865", "0.601", "0.3865", "0.6157", "0.2222", "0.6143" ], [ "0.1732", "0.6724", "0.8506", "0.6724", "0.8506", "0.6875", "0.1732", "0.6875" ], [ "0.1896", "0.69", "0.8545", "0.69", "0.8545", "0.705", "0.1896", "0.705" ], [ "0.1896", "0.707", "0.786", "0.707", "0.786", "0.72", "0.1896", "0.72" ], [ "0.1724", "0.765", "0.8506", "0.7646", "0.8506", "0.779", "0.1724", "0.78" ], [ "0.1896", "0.7827", "0.851", "0.7827", "0.851", "0.7974", "0.1896", "0.7974" ], [ "0.1691", "0.8423", "0.8486", "0.8423", "0.8486", "0.856", "0.1691", "0.856" ], [ "0.1896", "0.8594", "0.8545", "0.8594", "0.8545", "0.8735", "0.1896", "0.8735" ], [ "0.1896", "0.8755", "0.4307", "0.8755", "0.4307", "0.89", "0.1896", "0.89" ], [ "0.4836", "0.9136", "0.5146", "0.9136", "0.5146", "0.9287", "0.4836", "0.9287" ] ]

[ "SFTRACK++:ASPECTRALSEGMENTATIONAPPROACHFORSPACE-TIMECONSISTENTTRACKING", "Phase 1:Learn onmulti-instance segmentation", "1*tGTbbox", "NNtea", "racko", "NNa", "TrackerNte", "Nfear", "Nfeat", "earnto", "Seamentations", "InputChannels", "RGBvidec", "combine", "multipie", "Spectral clustering in space-time overa", "channels", "movingwindowofseveralframes", "NNesabo", "X", "spectra", "Sm,F,", "XNie", "Bbox Regression", "terations", "Phase3:Learnand Fine-tune", "Phase 2: Learn and Fine-tune on Segmentation", "on Tracking", "Figure l:SFTrack++: We start from video's RGB and ist frame GT bbox of the tracked", "object. We run state-of-the-art trackers, in an online manner, while fine-tuning", "NNfeat2seg network frame-by-frame (pretrained in Phase 1) to transform the ex-", "tracted feature maps (e.g. bboxes) to segmentation maps. Next, we learn to", "combine multiple segmentation inputs and refine the final mask using a spectral", "approach, applied also online over a moving window containing the previous N", "frames, for Niter spectral iterations (Phase 2).In Phase 3, we learn a bbox regres-", "sor from the final segmentation mask, NNseg2bor and fine-tune all our parameters", "on the tracking task", ". SFTrack++ brings to tracking a natural, contiguous, and efficient approach for inte-", "grating space and time components, using a fast 3D spectral clustering method over", "the graph of pixels from the video, to strengthen the tracked object's model", ". We explicitly learn intermediate fine-grained segmentation as opposed to rough bound", "ing boxes in our three phases end-to-end approach to a more robust tracking solution.", ". We integrate into our formulation a way of learning to combine multiple input chan", "nels, offering a wider view of the objects, harmonizing different perceptions, for a", "powerful and robust approach.", "137" ]

{ "arxiv_id": "2011.13843", "language": "zh", "page_idx": 1 }

[ [ "0.4575", "0.05176", "0.536", "0.05176", "0.536", "0.0631", "0.4575", "0.0631" ], [ "0.1438", "0.11743", "0.3105", "0.11615", "0.3105", "0.1307", "0.1438", "0.132" ], [ "0.1478", "0.1433", "0.853", "0.1433", "0.853", "0.1578", "0.1478", "0.1578" ], [ "0.1462", "0.161", "0.853", "0.161", "0.853", "0.1749", "0.1462", "0.1749" ], [ "0.1462", "0.178", "0.853", "0.178", "0.853", "0.1913", "0.1462", "0.1913" ], [ "0.1478", "0.1957", "0.853", "0.1957", "0.853", "0.209", "0.1478", "0.209" ], [ "0.1478", "0.2122", "0.851", "0.2122", "0.851", "0.2267", "0.1478", "0.2267" ], [ "0.1478", "0.2299", "0.8506", "0.2299", "0.8506", "0.2437", "0.1478", "0.2437" ], [ "0.1478", "0.2468", "0.851", "0.2468", "0.851", "0.2615", "0.1478", "0.2615" ], [ "0.1478", "0.2644", "0.853", "0.2644", "0.853", "0.2778", "0.1478", "0.2778" ], [ "0.1462", "0.281", "0.853", "0.281", "0.853", "0.2942", "0.1462", "0.2942" ], [ "0.1478", "0.2998", "0.4534", "0.2998", "0.4534", "0.31", "0.1478", "0.31" ], [ "0.492", "0.3005", "0.5293", "0.3005", "0.5293", "0.3113", "0.492", "0.3113" ], [ "0.544", "0.2986", "0.853", "0.2986", "0.853", "0.3125", "0.544", "0.3125" ], [ "0.1478", "0.315", "0.433", "0.315", "0.433", "0.3289", "0.1478", "0.3289" ], [ "0.584", "0.315", "0.8486", "0.315", "0.8486", "0.3289", "0.584", "0.3289" ], [ "0.536", "0.3225", "0.591", "0.3225", "0.591", "0.3257", "0.536", "0.3257" ], [ "0.1462", "0.332", "0.2467", "0.334", "0.2467", "0.3452", "0.1462", "0.3435" ], [ "0.6235", "0.3333", "0.705", "0.3333", "0.705", "0.3447", "0.6235", "0.3447" ], [ "0.7563", "0.3333", "0.8506", "0.3333", "0.8506", "0.3447", "0.7563", "0.3447" ], [ "0.3145", "0.334", "0.343", "0.334", "0.343", "0.3416", "0.3145", "0.3416" ], [ "0.246", "0.3372", "0.2607", "0.3372", "0.2607", "0.3403", "0.246", "0.3403" ], [ "0.7183", "0.3372", "0.739", "0.3372", "0.739", "0.3403", "0.7183", "0.3403" ], [ "0.1478", "0.3499", "0.4019", "0.351", "0.4019", "0.3623", "0.1478", "0.361" ], [ "0.6167", "0.3499", "0.853", "0.3499", "0.853", "0.363", "0.6167", "0.363" ], [ "0.1462", "0.3655", "0.3718", "0.3674", "0.3718", "0.3806", "0.1462", "0.379" ], [ "0.5884", "0.3687", "0.8486", "0.3687", "0.8486", "0.379", "0.5884", "0.379" ], [ "0.1487", "0.385", "0.2336", "0.385", "0.2336", "0.3958", "0.1487", "0.3958" ], [ "0.2303", "0.3882", "0.3162", "0.3882", "0.3162", "0.3958", "0.2303", "0.3958" ], [ "0.3186", "0.3901", "0.357", "0.3901", "0.357", "0.3933", "0.3186", "0.3933" ], [ "0.357", "0.3901", "0.3896", "0.3901", "0.3896", "0.3933", "0.357", "0.3933" ], [ "0.607", "0.3901", "0.634", "0.3901", "0.634", "0.3933", "0.607", "0.3933" ], [ "0.652", "0.387", "0.851", "0.385", "0.851", "0.3958", "0.652", "0.3977" ], [ "0.1487", "0.4028", "0.2893", "0.4028", "0.2893", "0.4136", "0.1487", "0.4136" ], [ "0.7417", "0.4048", "0.774", "0.4048", "0.774", "0.4124", "0.7417", "0.4124" ], [ "0.7817", "0.401", "0.8486", "0.4028", "0.8486", "0.414", "0.7817", "0.4124" ], [ "0.732", "0.4072", "0.7427", "0.4072", "0.7427", "0.4104", "0.732", "0.4104" ], [ "0.1503", "0.4192", "0.273", "0.4192", "0.273", "0.43", "0.1503", "0.43" ], [ "0.286", "0.406", "0.3022", "0.406", "0.3022", "0.4104", "0.286", "0.4104" ], [ "0.8013", "0.4216", "0.8506", "0.4216", "0.8506", "0.4287", "0.8013", "0.4287" ], [ "0.1478", "0.435", "0.2917", "0.435", "0.2917", "0.449", "0.1478", "0.449" ], [ "0.818", "0.4382", "0.8433", "0.4382", "0.8433", "0.4458", "0.818", "0.4458" ], [ "0.1487", "0.4539", "0.2197", "0.4539", "0.2197", "0.464", "0.1487", "0.464" ], [ "0.7837", "0.4539", "0.846", "0.4539", "0.846", "0.464", "0.7837", "0.464" ], [ "0.1487", "0.4702", "0.2418", "0.4702", "0.2418", "0.481", "0.1487", "0.481" ], [ "0.246", "0.473", "0.255", "0.473", "0.255", "0.4778", "0.246", "0.4778" ], [ "0.7812", "0.4702", "0.851", "0.4702", "0.851", "0.481", "0.7812", "0.481" ], [ "0.1478", "0.4873", "0.1978", "0.4849", "0.1978", "0.4956", "0.1478", "0.4976" ], [ "0.738", "0.4912", "0.749", "0.4912", "0.749", "0.4944", "0.738", "0.4944" ], [ "0.7686", "0.4873", "0.8447", "0.4873", "0.8447", "0.4988", "0.7686", "0.4988" ], [ "0.1487", "0.504", "0.2255", "0.504", "0.2255", "0.515", "0.1487", "0.515" ], [ "0.232", "0.506", "0.267", "0.506", "0.267", "0.513", "0.232", "0.513" ], [ "0.2769", "0.5083", "0.2876", "0.5083", "0.2876", "0.511", "0.2769", "0.511" ], [ "0.7183", "0.5083", "0.7607", "0.5083", "0.7607", "0.512", "0.7183", "0.512" ], [ "0.763", "0.505", "0.851", "0.505", "0.851", "0.515", "0.763", "0.515" ], [ "0.1487", "0.5215", "0.2852", "0.5215", "0.2852", "0.532", "0.1487", "0.532" ], [ "0.7197", "0.521", "0.853", "0.521", "0.853", "0.534", "0.7197", "0.534" ], [ "0.1462", "0.536", "0.3987", "0.5386", "0.3987", "0.552", "0.1462", "0.55" ], [ "0.6396", "0.5435", "0.6772", "0.5435", "0.6772", "0.547", "0.6396", "0.547" ], [ "0.675", "0.539", "0.851", "0.539", "0.851", "0.55", "0.675", "0.55" ], [ "0.1478", "0.5547", "0.4167", "0.5547", "0.4167", "0.5684", "0.1478", "0.5684" ], [ "0.652", "0.556", "0.853", "0.556", "0.853", "0.5693", "0.652", "0.5693" ], [ "0.1478", "0.5728", "0.4436", "0.5728", "0.4436", "0.586", "0.1478", "0.586" ], [ "0.4617", "0.577", "0.4885", "0.577", "0.4885", "0.58", "0.4617", "0.58" ], [ "0.5", "0.577", "0.513", "0.577", "0.513", "0.58", "0.5", "0.58" ], [ "0.5415", "0.577", "0.5522", "0.577", "0.5522", "0.58", "0.5415", "0.58" ], [ "0.5596", "0.5728", "0.853", "0.5728", "0.853", "0.586", "0.5596", "0.586" ], [ "0.1462", "0.589", "0.4795", "0.589", "0.4795", "0.603", "0.1462", "0.603" ], [ "0.5146", "0.5923", "0.5557", "0.5923", "0.5557", "0.5996", "0.5146", "0.5996" ], [ "0.5513", "0.5923", "0.5884", "0.5923", "0.5884", "0.5996", "0.5513", "0.5996" ], [ "0.614", "0.5923", "0.63", "0.5923", "0.63", "0.5996", "0.614", "0.5996" ], [ "0.648", "0.5903", "0.853", "0.5903", "0.853", "0.601", "0.648", "0.601" ], [ "0.1462", "0.605", "0.8545", "0.606", "0.8545", "0.6206", "0.1462", "0.619" ], [ "0.1462", "0.622", "0.853", "0.6226", "0.853", "0.637", "0.1462", "0.636" ], [ "0.1478", "0.64", "0.851", "0.64", "0.851", "0.654", "0.1478", "0.654" ], [ "0.1462", "0.656", "0.851", "0.657", "0.851", "0.672", "0.1462", "0.6704" ], [ "0.1478", "0.6724", "0.851", "0.6724", "0.851", "0.6875", "0.1478", "0.6875" ], [ "0.1478", "0.6914", "0.853", "0.6914", "0.853", "0.705", "0.1478", "0.705" ], [ "0.1447", "0.707", "0.831", "0.7085", "0.831", "0.7227", "0.1447", "0.7217" ], [ "0.1438", "0.747", "0.3145", "0.747", "0.3145", "0.7646", "0.1438", "0.7646" ], [ "0.1447", "0.773", "0.853", "0.7734", "0.853", "0.7886", "0.1447", "0.787" ], [ "0.1462", "0.7905", "0.853", "0.7905", "0.853", "0.808", "0.1462", "0.808" ], [ "0.1447", "0.808", "0.853", "0.8066", "0.853", "0.8213", "0.1447", "0.8228" ], [ "0.1478", "0.8257", "0.851", "0.8257", "0.851", "0.839", "0.1478", "0.839" ], [ "0.1462", "0.8433", "0.853", "0.8433", "0.853", "0.857", "0.1462", "0.857" ], [ "0.1447", "0.8594", "0.8545", "0.8594", "0.8545", "0.877", "0.1447", "0.877" ], [ "0.1462", "0.8755", "0.8545", "0.877", "0.8545", "0.8916", "0.1462", "0.89" ], [ "0.486", "0.9155", "0.5146", "0.9155", "0.5146", "0.9297", "0.486", "0.9297" ] ]

[ "BURCEANU", "2. Related work", "常规对象跟踪。在三个主要的跟踪器系列中，基于暹罗的跟踪器近年来因其高速和端到端", "功能Bertinetto et al.(2016b);Liet al.(2018,2019);Zhang and Peng(2020)而获得了很大", "的吸引力。大多数方法侧重于详尽的离线培训，无法监控变化。r。t。初始模板Taoetal.", "(2016)；Xuetal.（2020)；Chenetal.（2020），而其他人在线更新他们的模型Voigtlaender", "etal.（2020)；Daietal.（2020）。然而，训练时对看不见的物体和变换的鲁棒性仍然是暹", "罗跟踪器的基本问题。用于跟踪的元学习方法EPark andBerg（2018)；Bertinetto etal.", "(2016a)；Wangetal.（2020）通过提出与目标无关的跟踪模型，提供了一种有趣的方式来适", "应当前感兴趣的对象，同时保持较短的推理时间。这两种方法（Siamese和元学习跟踪器）", "的一个主要限制是，它们无法连续适应被跟踪对象的实时变化，而是使用了几个精心选择", "的补丁的历史记录，甚至只是最初的补", "相反", "我们的方法自然地集成了时间维度，通", "过不断地执行本地时间和空间对象的", "判别方法EDanelljanetal.（2019", "", "2020):Lukez", "宗对象Dar", "et al. (2017)", "202", "1+", "", "（背景，于扰物，硬底片）的变化", "hatetal.（2020）集成到方法流", "程中。它们被证明是强大的，", "的观察或不可训练的端到端模块", "SFTrack+-", "十提供一", "t-", "t", "ET", "限度地减少十扰和背景噪", "声。我们的方法将", "时日", "一致性认", "", "可作为算法的附", "F", "八多数", "跟踪解决方案都使", "2020", "2019):Xl", "割掩码", "此外：它们", "1", "提取器E", "Danell", "", "巢作。有厂", "个跟踪器昌", "然维", "", "", "性，通过显", "式建模不同的路名", "相比之下，我们", "的端到端多通道公式通过10输人", "", "。以前将图像和视频表", "示为图形，其中节点是像素，超像美", "此选择直接影响运行时间", "和性能。关于边，它们通常是无向的，", "", "", "", "函数建模，但是也有一些作品使用有向", "的 Torsello et al. (2006); Yu and Shi (2001", "晋聚车", "天万#", "Ng", "t al. (2001): Meila and Shi", "(2001)；LeordeanuandHebert（2005）搜索图的邻接矩阵的前导或最小特征向量，以解决聚", "类的分配。光谱聚类以前用于像素级图像分割EShiandMalik（2000），在运行时间和建立", "视频补丁之间的时空对应关系EJabrietal.（2020a）方面负担很重。图切割是谱聚类的常用", "方法，具有许多变体 Shi and Malik(2000)；Ding et al.(2001)；Sarkar and Soundararajan", "(2000）。SFSegBurceanuandLeordeanu（2020）提出了一种3D过滤技术，用于在不显式", "计算图的邻接矩阵的情况下有效地找到谱聚类解决方案。受这种方法的启发，我们将改进", "的版本与多通道输入的学习集成在一起，作为跟踪器中的中间组件，如Sec中所述。3", "3. Our approach", "SFTrack++算法有三个阶段，因为我们在图中直观地展示了它们。1。在阶段1中，我", "们学习了一个神经网络NNfeat2seg，该神经网络将RGB和使用跟踪器（来自跟踪器预测", "的e.g.bbox）提取的顿级特征图转换为分割掩码。仅使用RGB作为输入是不够的，因为", "顿可以包含多个对象和实例，并且我们还需要指向被跟踪对象的指针来预测其分割。接下", "来，在第2阶段，我们在输入上逐顿运行多个最先进的跟踪器作为在线过程，并从中提取", "输入通道（e.g.bboxes）。我们使用先前调用的模块NNfeat2seg将这些特征图转换为分割", "图。接下来，我们学习使用频谱解决方案来组合和优化当前顿的输出，以保持时空一致性，", "138" ]

{ "arxiv_id": "2011.13843", "language": "zh", "page_idx": 2 }

[ [ "0.1462", "0.0505", "0.8594", "0.0505", "0.8594", "0.0644", "0.1462", "0.0644" ], [ "0.1462", "0.1187", "0.853", "0.1187", "0.853", "0.132", "0.1462", "0.132" ], [ "0.1478", "0.1351", "0.853", "0.1351", "0.853", "0.1483", "0.1478", "0.1483" ], [ "0.1462", "0.1528", "0.853", "0.1528", "0.853", "0.1674", "0.1462", "0.1674" ], [ "0.1478", "0.1704", "0.3594", "0.1704", "0.3594", "0.1812", "0.1478", "0.1812" ], [ "0.1462", "0.1849", "0.853", "0.1875", "0.853", "0.202", "0.1462", "0.1995" ], [ "0.1478", "0.2039", "0.851", "0.2039", "0.851", "0.2172", "0.1478", "0.2172" ], [ "0.1462", "0.2216", "0.853", "0.2216", "0.853", "0.2349", "0.1462", "0.2349" ], [ "0.1478", "0.238", "0.8545", "0.238", "0.8545", "0.252", "0.1478", "0.252" ], [ "0.1405", "0.25", "0.857", "0.252", "0.857", "0.2766", "0.1405", "0.2747" ], [ "0.1447", "0.2727", "0.8545", "0.2734", "0.8545", "0.2886", "0.1447", "0.2874" ], [ "0.1478", "0.2917", "0.8545", "0.2917", "0.8545", "0.305", "0.1478", "0.305" ], [ "0.1462", "0.308", "0.4314", "0.308", "0.4314", "0.3213", "0.1462", "0.3213" ], [ "0.1732", "0.3257", "0.853", "0.3257", "0.853", "0.3403", "0.1732", "0.3403" ], [ "0.1447", "0.3416", "0.3105", "0.3423", "0.3105", "0.3567", "0.1447", "0.356" ], [ "0.1528", "0.3687", "0.824", "0.3706", "0.824", "0.3914", "0.1528", "0.3894" ], [ "0.83", "0.3865", "0.8486", "0.3865", "0.8486", "0.3984", "0.83", "0.3984" ], [ "0.581", "0.3945", "0.822", "0.3933", "0.822", "0.411", "0.581", "0.4124" ], [ "0.1462", "0.4268", "0.853", "0.4287", "0.853", "0.443", "0.1462", "0.4414" ], [ "0.1487", "0.4458", "0.468", "0.4458", "0.468", "0.459", "0.1487", "0.459" ], [ "0.1478", "0.4634", "0.851", "0.4634", "0.851", "0.4778", "0.1478", "0.4778" ], [ "0.1462", "0.4792", "0.853", "0.4792", "0.853", "0.4968", "0.1462", "0.4968" ], [ "0.1478", "0.5005", "0.8545", "0.5005", "0.8545", "0.515", "0.1478", "0.515" ], [ "0.1462", "0.518", "0.851", "0.518", "0.851", "0.532", "0.1462", "0.532" ], [ "0.1478", "0.535", "0.851", "0.535", "0.851", "0.55", "0.1478", "0.55" ], [ "0.1462", "0.5503", "0.5215", "0.552", "0.5215", "0.5664", "0.1462", "0.565" ], [ "0.1462", "0.5923", "0.3945", "0.5923", "0.3945", "0.607", "0.1462", "0.607" ], [ "0.1462", "0.6206", "0.851", "0.6206", "0.851", "0.634", "0.1462", "0.634" ], [ "0.1462", "0.637", "0.853", "0.637", "0.853", "0.651", "0.1462", "0.651" ], [ "0.1462", "0.654", "0.851", "0.654", "0.851", "0.6675", "0.1462", "0.6675" ], [ "0.1478", "0.672", "0.853", "0.672", "0.853", "0.685", "0.1478", "0.685" ], [ "0.1478", "0.688", "0.8545", "0.688", "0.8545", "0.702", "0.1478", "0.702" ], [ "0.1462", "0.7056", "0.853", "0.704", "0.853", "0.7183", "0.1462", "0.72" ], [ "0.1462", "0.7217", "0.8545", "0.72", "0.8545", "0.735", "0.1462", "0.7363" ], [ "0.1447", "0.7373", "0.8545", "0.7393", "0.8545", "0.754", "0.1447", "0.752" ], [ "0.1462", "0.756", "0.8545", "0.756", "0.8545", "0.771", "0.1462", "0.771" ], [ "0.1462", "0.7734", "0.853", "0.7734", "0.853", "0.7886", "0.1462", "0.7886" ], [ "0.1447", "0.7905", "0.8545", "0.7905", "0.8545", "0.808", "0.1447", "0.808" ], [ "0.1462", "0.808", "0.853", "0.808", "0.853", "0.8228", "0.1462", "0.8228" ], [ "0.1447", "0.8237", "0.853", "0.8247", "0.853", "0.839", "0.1447", "0.8384" ], [ "0.1462", "0.8433", "0.851", "0.8433", "0.851", "0.857", "0.1462", "0.857" ], [ "0.1462", "0.8594", "0.8545", "0.8594", "0.8545", "0.877", "0.1462", "0.877" ], [ "0.1462", "0.8774", "0.8506", "0.8774", "0.8506", "0.8916", "0.1462", "0.8916" ], [ "0.486", "0.9155", "0.511", "0.9155", "0.511", "0.9287", "0.486", "0.9287" ] ]

[ "SFTRACK++: A SPECTRAL SEGMENTATION APPROACH FOR SPACE-TIME CONSISTENT TRACKING", "适用于通过多个通道进行学习。请注意，当应用频谱选代时，我们在视频体积中的前N个", "顿上使用滑动窗口方法。为了监督这条路径，我们使用分段地面实况。第3阶段将神经网", "络作为bbox回归器学习到上一阶段的最终分割图NNseg2bor，同时使用跟踪GT微调模", "型中的所有其他可训练参数", "光谱分割方法。接下来我们通过以下几个方面，简要说明它们之间的联系：分段一→前", "导特征向量→幂送代???D滤波公式一多通道。先前将图像分割公式化为图分割问", "题，其中分割解决方案ShiandMalik（2000）是邻接矩阵的前导特征向量。它以类似", "的方式用于视频BurceanuandLeordeanu（2020）。幂迭代算法可以计算前导特征向量：", "a+1←jeN()Mij，其中M是N×N图的邻接矩阵，N是图中的节点数（在我们", "的例子中是视频时空体积中的像素），（i）是节点i的时空邻域，每个步骤k后跟归一化。", "幂送代中使用的邻接矩阵通常取决于两种类型的术语：一元的是关于单个节点属性的，成", "对的描述两个节点（对）之间的关系", "按照这种方法，SFSeg使用近似邻接矩阵的3D滤波重写功率迭代。解决方案在方程", "式E1中进行了描述：", "Xk+1 ← normalized(SP . (α-11 -F2) ·G3D * (SP .X) - SP .G3D * (F2 . SP.Xk)+", "(1", "2SP.F.G3D*(F.SP.Xk),", "其中*是在时空体积上具有高斯滤波器G3D的3D卷积，·是元素乘法，S和F是一元和", "成对矩阵，其重要性为MAXXXXXXXX", "多通道学习公式。我们扩展了SFSeg中的单通道公式，以便它可以学习如何分别组合多", "个输人通道S,和F，分别用于一元和成对项：Sm←一(Zi=iwsiS,+bs1），Fm←", "o（Z=1wfF；+bf1），其中Sm和Fm是多通道一元和成对映射，是符号XMATHX-6。", "我们分别用它们的多通道版本Sm和Fm替换公式1中的S和F。我们通过分段和跟踪", "任务学习ws,i,wf,i,bs，bf参数。更多，SFTrack++可以学习端到端，从原始输人顿一直到", "最终输出，在端到端可学习特征提取器的情况下", "4.Experimental protocol", "我们测试SFTrack十十是否通过具有中间的细粒度表示，在多个最先进的跟踪器输出上提", "取并在空间和时间上平滑，从而为跟踪带来了互补的维度。我们指导我们的实验，以便", "我们首先评估最便宜的途径。为了减少超参数搜索负担，我们使用AdamWELoshchilov", "andHutter（2019），并使用调度程序策略来降低平台上的学习率。为了提高效率和紧凑", "性，我们使用相同的通道来构造一元映射和成对映射：S，=F。他们学习的权重也被共享", "ws,i=wf,i。我们将使用顶级单对象跟踪器提取的bbox用作输人通道。我们选择了10个", "顶级跟踪器：SiamR-CNNEVoigtlaenderetal.（2020），LTMUEDaietal.（2020），KYSE", "Bhat et al.(2020),PrDiMPEDanelljan et al.(2020),ATOMEDanelljan et al.(2019),", "OceanEZhang and Peng (2020), D3SELukezic et al. (2020) , SiamFCEXMATHXESiM", "培训。在第1阶段中，我们针对每个对象在DAVIS-2017Pont-Tuset et al.（2017）和", "Youtube-VISYangetal.（2019）训练集上训练NNfeat2seg网络。它接收当前RGB和跟", "踪方法（bbox）的输出，在训练时随机抽样。我们使用U-Net架构，验证正确的参数数量", "（100K-1mil）和层数。我们使用DAVIS-2017和Youtube-VIS评估集来停止训练。遵循课", "程学习方法，在引跟踪方法之前，我们在跟踪GTbbox的开头使用（从分段GT中提取", "作为直bbox）。这允许NNfeat2seg组件在引人有故障的bbox提取器之前获得良好的初始", "化，即之前提到的前10种跟踪方法。对于第2阶段，我们还为分段任务进行训练。我们学", "139" ]

{ "arxiv_id": "2011.13843", "language": "zh", "page_idx": 3 }

[ [ "0.4575", "0.05176", "0.536", "0.05176", "0.536", "0.0631", "0.4575", "0.0631" ], [ "0.1478", "0.1187", "0.853", "0.1187", "0.853", "0.132", "0.1478", "0.132" ], [ "0.1462", "0.1338", "0.411", "0.1351", "0.411", "0.1497", "0.1462", "0.1483" ], [ "0.415", "0.1338", "0.851", "0.1338", "0.851", "0.1477", "0.415", "0.1477" ], [ "0.1462", "0.1515", "0.851", "0.1509", "0.851", "0.1654", "0.1462", "0.1666" ], [ "0.1503", "0.1692", "0.853", "0.1692", "0.853", "0.1843", "0.1503", "0.1843" ], [ "0.1462", "0.1843", "0.853", "0.1863", "0.853", "0.2008", "0.1462", "0.1989" ], [ "0.1478", "0.2026", "0.3708", "0.2026", "0.3708", "0.2166", "0.1478", "0.2166" ], [ "0.1478", "0.2203", "0.853", "0.2203", "0.853", "0.2343", "0.1478", "0.2343" ], [ "0.1478", "0.2374", "0.853", "0.2374", "0.853", "0.2507", "0.1478", "0.2507" ], [ "0.1462", "0.2551", "0.8545", "0.2551", "0.8545", "0.2683", "0.1462", "0.2683" ], [ "0.1478", "0.2715", "0.851", "0.2715", "0.851", "0.2854", "0.1478", "0.2854" ], [ "0.1478", "0.289", "0.853", "0.289", "0.853", "0.303", "0.1478", "0.303" ], [ "0.1447", "0.305", "0.853", "0.3062", "0.853", "0.3208", "0.1447", "0.3193" ], [ "0.1462", "0.3225", "0.8545", "0.3225", "0.8545", "0.3372", "0.1462", "0.3372" ], [ "0.1478", "0.3403", "0.853", "0.3403", "0.853", "0.3542", "0.1478", "0.3542" ], [ "0.1462", "0.3567", "0.853", "0.3567", "0.853", "0.3718", "0.1462", "0.3718" ], [ "0.1462", "0.375", "0.3064", "0.375", "0.3064", "0.3882", "0.1462", "0.3882" ], [ "0.1462", "0.3901", "0.851", "0.3901", "0.851", "0.4048", "0.1462", "0.4048" ], [ "0.1438", "0.4048", "0.8555", "0.406", "0.8555", "0.4268", "0.1438", "0.4255" ], [ "0.1462", "0.4255", "0.8545", "0.4255", "0.8545", "0.4395", "0.1462", "0.4395" ], [ "0.1462", "0.443", "0.853", "0.443", "0.853", "0.457", "0.1462", "0.457" ], [ "0.1478", "0.461", "0.8545", "0.461", "0.8545", "0.476", "0.1478", "0.476" ], [ "0.1487", "0.4778", "0.853", "0.4778", "0.853", "0.4912", "0.1487", "0.4912" ], [ "0.1462", "0.4944", "0.851", "0.4944", "0.851", "0.5083", "0.1462", "0.5083" ], [ "0.1462", "0.511", "0.853", "0.511", "0.853", "0.526", "0.1462", "0.526" ], [ "0.1447", "0.55", "0.3865", "0.55", "0.3865", "0.563", "0.1447", "0.563" ], [ "0.1528", "0.577", "0.851", "0.577", "0.851", "0.5913", "0.1528", "0.5913" ], [ "0.1462", "0.5923", "0.851", "0.5923", "0.851", "0.607", "0.1462", "0.607" ], [ "0.1478", "0.6113", "0.851", "0.6113", "0.851", "0.625", "0.1478", "0.625" ], [ "0.1478", "0.6274", "0.853", "0.6274", "0.853", "0.6426", "0.1478", "0.6426" ], [ "0.1478", "0.646", "0.851", "0.646", "0.851", "0.659", "0.1478", "0.659" ], [ "0.1478", "0.662", "0.853", "0.662", "0.853", "0.6753", "0.1478", "0.6753" ], [ "0.1478", "0.6797", "0.853", "0.6797", "0.853", "0.6934", "0.1478", "0.6934" ], [ "0.1478", "0.6963", "0.853", "0.6963", "0.853", "0.711", "0.1478", "0.711" ], [ "0.1462", "0.714", "0.8545", "0.714", "0.8545", "0.728", "0.1462", "0.728" ], [ "0.1447", "0.73", "0.8545", "0.731", "0.8545", "0.7456", "0.1447", "0.744" ], [ "0.1447", "0.747", "0.58", "0.7476", "0.58", "0.762", "0.1447", "0.761" ], [ "0.1528", "0.773", "0.853", "0.7734", "0.853", "0.7886", "0.1528", "0.787" ], [ "0.1478", "0.7915", "0.853", "0.7915", "0.853", "0.8047", "0.1478", "0.8047" ], [ "0.1438", "0.803", "0.8555", "0.806", "0.8555", "0.827", "0.1438", "0.8237" ], [ "0.1462", "0.8257", "0.8545", "0.8257", "0.8545", "0.839", "0.1462", "0.839" ], [ "0.1478", "0.8433", "0.853", "0.8433", "0.853", "0.857", "0.1478", "0.857" ], [ "0.1478", "0.86", "0.8545", "0.86", "0.8545", "0.8735", "0.1478", "0.8735" ], [ "0.1447", "0.8755", "0.853", "0.877", "0.853", "0.8916", "0.1447", "0.89" ], [ "0.486", "0.9155", "0.5146", "0.9155", "0.5146", "0.9297", "0.486", "0.9297" ] ]

[ "BURCEANU", "习了方法的第二部分，即具有中间的细粒度表示，通过多个通道提取，并在空间和时间上", "进行平滑。我们在这里验证Niters", "，频谱选代次数（1一5）。我们在DAVIS-2016Perazzi", "etal.（2016）和Youtube-VIS数据集上进行培训。在第3阶段，进行跟踪训练，我们学习", "了一个回归网络NNseg2bor（具有50K-500K参数），以将最终的分割转换为bbox。我", "们在TrackingNetEMuller et al.(2018)，LaSOTEFan et al.(2019)和GOT-10kEHuang", "etal.（2019）训练上进行训练", "基线比较。与其他方法进行比较的实验集中在改进上SFTrack_XmilXMAXXmil3可以带", "来最先进的方法和其他用于一般对象跟踪的竞争性方法：单一方法最先进的解决方案，跟", "踪器上的基本集成，SFTrack仅在最佳跟踪器上应用，SFTrack应用于基本集成和最佳", "学习的神经网络集成我们可以从几种配置中获得（2D和3D版本的具有U-Net的XTHX。", "如前所述，所有方法都从顶级10跟踪器接收相同的输人，并在TrackingNet，LaSOT", "和GOT-10k火车集上进行训练。我们针对七个跟踪基准的所有基线评估我们的解决方", "案:VOT2018EKristan et al.(2018)，LaSOT, TrackingNet,GOT-10k, NFSEGaloogahi", "et al.(2017)，OTB-100E/Wuet al.(2015)和UAV123EMueller et al.(2016)。对于本文", "的主要结论，我们将提供统计结果（几次运行的均值和方差），以更好地指示强烈的正/负", "结果或不确定的结果", "烧蚀研究。我们改变了端到端模型的几个组件，以更好地了解它们的作用和力量。我们训", "练我们的阶段1组件NNfeat2segnet，不仅用于bbox输人功能，还用于从每个跟踪器架", "构中提取的其他早期功能。我们测试这种情况下的整体跟踪性能。我们从管道中删除第2", "阶段的光谱细化并报告结果。我们通过用OpenCVEBradskiandKaehler（2008）的直盒和", "旋转盒提取器代替它来测试没有第3阶段神经网络NNseg2bor的跟踪器的性能。我们测试", "了几种损失，以优化分割和跟踪任务：加权diclossESudreetal.（2017）和二进制交叉熵，", "Focal-TverskyEAbraham and Khan (2019) 和 Focal-DiceE/Wang and Chung (2018)之间", "的线性组合。对于烧蚀实验，我们仅对OTB100、UAV123和NFS30跟踪数据集进行评估", "5. Experimental Results", "Comparisonwithothermethods．在选项卡中。1我们在五个跟踪基准上展示了", "我们的方法的结果：OTB100，UAV123，NFS30，GOT-10k和TrackingNet，将其与其他", "顶级单一方法和集成解决方案进行比较。对于单一方法，我们考虑了SFTrack十+中的每", "种输人法：D3S，SiamBAN，ATOM-18，SiamRPN++，PrDimp-18。我们为每个跟踪器", "只选择了一个轻量级配置，在所有基准测试中都很常见。对于集合，我们使用1）基本的", "每像素中值集合，然后是所有实验中使用的相同的边界框回归器（请参见Sec。E7），我", "们还训练了二个更复杂的神经网络：2）具有UNet架构（具有5缩减和5扩展层）和类", "似数量的参数，例如SFTrack十十（～4.3百万）。我们观察到，我们的方法在不同的运行", "（包括从头开始训练）之间的差异非常小，显示出稳健的结果和明确的结论。在GOT-10k", "和TrackingNet等更新，更大，更通用的数据集上，我们的方法大大超过了其他方法，而在", "OTB100，UAV123和NFS30上，它具有竞争力的结果", "Ablationstudies．为了验证我们方法的组件，我们在Tab中测试。2不同变体，在", "OTB100、UAV123和NFS30上报告结果。首先，我们从阶段2中删除频谱细化组件，去", "除时间依赖性并使每顿预测独立。在下一个实验中，我们从phase3NNsegm26boa中删除神", "经网络。在下一块中，我们调查输入法的数量。最后，我们改变第二阶段的频谱送代次数。", "从这种广泛的消融中得出的结论如下：1）频谱细化组件非常重要，强调了最初的直觉，即", "使用我们提出的频谱方法保持对象在空间和时间上的一致性可以提高跟踪的整体性能。2）", "我们的SFTrack+十方法中输人的质量很重要，但是我们使用的方法越多，效果越好。3）我", "140" ]

{ "arxiv_id": "2011.13843", "language": "zh", "page_idx": 4 }

[ [ "0.4324", "0.0956", "0.571", "0.0956", "0.571", "0.10986", "0.4324", "0.10986" ], [ "0.3914", "0.1134", "0.607", "0.11224", "0.607", "0.13", "0.3914", "0.1312" ], [ "0.215", "0.1567", "0.7866", "0.158", "0.7866", "0.1729", "0.215", "0.1716" ], [ "0.199", "0.1733", "0.804", "0.1746", "0.804", "0.1895", "0.199", "0.1882" ], [ "0.361", "0.1924", "0.6406", "0.1924", "0.6406", "0.206", "0.361", "0.206" ], [ "0.2627", "0.2542", "0.3418", "0.2542", "0.3418", "0.2683", "0.2627", "0.2683" ], [ "0.513", "0.2546", "0.881", "0.2546", "0.881", "0.2683", "0.513", "0.2683" ], [ "0.51", "0.2695", "0.879", "0.2695", "0.879", "0.2832", "0.51", "0.2832" ], [ "0.1453", "0.284", "0.4575", "0.284", "0.4575", "0.2976", "0.1453", "0.2976" ], [ "0.51", "0.2874", "0.875", "0.2861", "0.875", "0.2998", "0.51", "0.3005" ], [ "0.1453", "0.2976", "0.4575", "0.2976", "0.4575", "0.3113", "0.1453", "0.3113" ], [ "0.532", "0.304", "0.8774", "0.304", "0.8774", "0.3176", "0.532", "0.3176" ], [ "0.1453", "0.3123", "0.4592", "0.3113", "0.4592", "0.3247", "0.1453", "0.326" ], [ "0.513", "0.3196", "0.881", "0.3196", "0.881", "0.3333", "0.513", "0.3333" ], [ "0.1436", "0.326", "0.4575", "0.3267", "0.4575", "0.3403", "0.1436", "0.339" ], [ "0.1453", "0.3403", "0.4575", "0.3403", "0.4575", "0.354", "0.1453", "0.354" ], [ "0.511", "0.335", "0.879", "0.335", "0.879", "0.3486", "0.511", "0.3486" ], [ "0.147", "0.3557", "0.4568", "0.3557", "0.4568", "0.3665", "0.147", "0.3665" ], [ "0.513", "0.3516", "0.881", "0.3516", "0.881", "0.3652", "0.513", "0.3652" ], [ "0.1453", "0.3682", "0.4575", "0.3694", "0.4575", "0.383", "0.1453", "0.3818" ], [ "0.513", "0.3677", "0.876", "0.3677", "0.876", "0.3813", "0.513", "0.3813" ], [ "0.1453", "0.383", "0.4575", "0.383", "0.4575", "0.3967", "0.1453", "0.3967" ], [ "0.513", "0.3855", "0.881", "0.3855", "0.881", "0.3984", "0.513", "0.3984" ], [ "0.1453", "0.398", "0.4592", "0.398", "0.4592", "0.411", "0.1453", "0.411" ], [ "0.5146", "0.401", "0.881", "0.401", "0.881", "0.4146", "0.5146", "0.4146" ], [ "0.1453", "0.4104", "0.4592", "0.411", "0.4592", "0.4258", "0.1453", "0.4246" ], [ "0.513", "0.4163", "0.881", "0.415", "0.881", "0.43", "0.513", "0.4312" ], [ "0.1453", "0.4258", "0.4592", "0.4258", "0.4592", "0.4395", "0.1453", "0.4395" ], [ "0.51", "0.4316", "0.879", "0.4316", "0.879", "0.4453", "0.51", "0.4453" ], [ "0.1453", "0.4407", "0.4575", "0.4407", "0.4575", "0.4536", "0.1453", "0.4536" ], [ "0.511", "0.4478", "0.881", "0.449", "0.881", "0.4631", "0.511", "0.462" ], [ "0.1453", "0.4536", "0.4575", "0.4536", "0.4575", "0.4673", "0.1453", "0.4673" ], [ "0.1453", "0.4685", "0.4575", "0.4685", "0.4575", "0.4822", "0.1453", "0.4822" ], [ "0.513", "0.4656", "0.881", "0.4656", "0.881", "0.4785", "0.513", "0.4785" ], [ "0.1453", "0.4822", "0.4575", "0.4822", "0.4575", "0.4954", "0.1453", "0.4954" ], [ "0.511", "0.481", "0.879", "0.481", "0.879", "0.4946", "0.511", "0.4946" ], [ "0.147", "0.4963", "0.4592", "0.4963", "0.4592", "0.5103", "0.147", "0.5103" ], [ "0.513", "0.4976", "0.881", "0.4976", "0.881", "0.511", "0.513", "0.511" ], [ "0.1453", "0.511", "0.4575", "0.5103", "0.4575", "0.524", "0.1453", "0.525" ], [ "0.513", "0.513", "0.881", "0.513", "0.881", "0.527", "0.513", "0.527" ], [ "0.147", "0.527", "0.4568", "0.527", "0.4568", "0.5376", "0.147", "0.5376" ], [ "0.513", "0.5293", "0.882", "0.5293", "0.882", "0.5435", "0.513", "0.5435" ], [ "0.1453", "0.539", "0.4575", "0.539", "0.4575", "0.5527", "0.1453", "0.5527" ], [ "0.511", "0.5444", "0.881", "0.5435", "0.881", "0.558", "0.511", "0.5586" ], [ "0.1453", "0.5527", "0.4568", "0.5527", "0.4568", "0.5664", "0.1453", "0.5664" ], [ "0.513", "0.561", "0.882", "0.561", "0.882", "0.5747", "0.513", "0.5747" ], [ "0.1453", "0.5684", "0.3972", "0.5684", "0.3972", "0.582", "0.1453", "0.582" ], [ "0.3896", "0.5693", "0.4568", "0.5693", "0.4568", "0.58", "0.3896", "0.58" ], [ "0.1436", "0.582", "0.4575", "0.5806", "0.4575", "0.594", "0.1436", "0.5957" ], [ "0.513", "0.5767", "0.882", "0.5767", "0.882", "0.5903", "0.513", "0.5903" ], [ "0.1453", "0.5967", "0.4575", "0.5967", "0.4575", "0.61", "0.1453", "0.61" ], [ "0.511", "0.593", "0.881", "0.593", "0.881", "0.6055", "0.511", "0.6055" ], [ "0.1453", "0.61", "0.4575", "0.61", "0.4575", "0.6235", "0.1453", "0.6235" ], [ "0.511", "0.6094", "0.881", "0.6094", "0.881", "0.6226", "0.511", "0.6226" ], [ "0.1453", "0.6245", "0.318", "0.6245", "0.318", "0.638", "0.1453", "0.638" ], [ "0.511", "0.6245", "0.6616", "0.626", "0.6616", "0.64", "0.511", "0.6396" ], [ "0.531", "0.6426", "0.882", "0.6426", "0.882", "0.6562", "0.531", "0.6562" ], [ "0.11755", "0.6562", "0.1503", "0.6562", "0.1503", "0.6646", "0.11755", "0.6646" ], [ "0.1511", "0.655", "0.2612", "0.655", "0.2612", "0.6665", "0.1511", "0.6665" ], [ "0.513", "0.659", "0.882", "0.659", "0.882", "0.673", "0.513", "0.673" ], [ "0.1184", "0.6797", "0.486", "0.6797", "0.486", "0.6934", "0.1184", "0.6934" ], [ "0.5146", "0.675", "0.881", "0.675", "0.881", "0.6885", "0.5146", "0.6885" ], [ "0.511", "0.6904", "0.882", "0.6904", "0.882", "0.7036", "0.511", "0.7036" ], [ "0.1184", "0.697", "0.4844", "0.697", "0.4844", "0.7104", "0.1184", "0.7104" ], [ "0.513", "0.707", "0.881", "0.707", "0.881", "0.72", "0.513", "0.72" ], [ "0.12177", "0.7144", "0.4836", "0.7144", "0.4836", "0.7246", "0.12177", "0.7246" ], [ "0.511", "0.7217", "0.879", "0.7227", "0.879", "0.737", "0.511", "0.7363" ], [ "0.1184", "0.7285", "0.486", "0.7285", "0.486", "0.742", "0.1184", "0.742" ], [ "0.1184", "0.7446", "0.486", "0.7446", "0.486", "0.7583", "0.1184", "0.7583" ], [ "0.511", "0.7393", "0.882", "0.7393", "0.882", "0.753", "0.511", "0.753" ], [ "0.513", "0.755", "0.882", "0.755", "0.882", "0.7695", "0.513", "0.7695" ], [ "0.12006", "0.7627", "0.4836", "0.7627", "0.4836", "0.7725", "0.12006", "0.7725" ], [ "0.511", "0.771", "0.879", "0.771", "0.879", "0.7837", "0.511", "0.7837" ], [ "0.1184", "0.777", "0.4844", "0.777", "0.4844", "0.7905", "0.1184", "0.7905" ], [ "0.11584", "0.7915", "0.4836", "0.7915", "0.4836", "0.8047", "0.11584", "0.8047" ], [ "0.511", "0.786", "0.879", "0.786", "0.879", "0.8", "0.511", "0.8" ], [ "0.11755", "0.8086", "0.486", "0.8086", "0.486", "0.8223", "0.11755", "0.8223" ], [ "0.513", "0.8047", "0.8774", "0.8047", "0.8774", "0.8154", "0.513", "0.8154" ], [ "0.513", "0.8193", "0.881", "0.8193", "0.881", "0.833", "0.513", "0.833" ], [ "0.147", "0.832", "0.424", "0.833", "0.424", "0.8467", "0.147", "0.8457" ], [ "0.531", "0.835", "0.879", "0.835", "0.879", "0.8486", "0.531", "0.8486" ], [ "0.1411", "0.8457", "0.1797", "0.847", "0.178", "0.858", "0.1394", "0.8555" ], [ "0.1957", "0.847", "0.2568", "0.847", "0.2568", "0.858", "0.1957", "0.858" ], [ "0.2688", "0.8486", "0.2957", "0.8486", "0.2957", "0.857", "0.2688", "0.857" ], [ "0.293", "0.847", "0.4844", "0.847", "0.4844", "0.858", "0.293", "0.858" ], [ "0.511", "0.8516", "0.881", "0.8516", "0.881", "0.865", "0.511", "0.865" ], [ "0.11755", "0.8594", "0.4382", "0.8594", "0.4382", "0.869", "0.11755", "0.869" ], [ "0.1394", "0.8706", "0.4878", "0.8706", "0.4878", "0.8843", "0.1394", "0.8843" ], [ "0.511", "0.868", "0.634", "0.865", "0.634", "0.88", "0.511", "0.883" ], [ "0.1184", "0.8843", "0.4844", "0.8843", "0.4844", "0.8945", "0.1184", "0.8945" ], [ "0.11755", "0.8945", "0.486", "0.8945", "0.486", "0.9077", "0.11755", "0.9077" ], [ "0.531", "0.8936", "0.879", "0.8936", "0.879", "0.9067", "0.531", "0.9067" ], [ "0.11584", "0.9062", "0.4182", "0.9062", "0.4182", "0.9194", "0.11584", "0.9194" ], [ "0.511", "0.9062", "0.881", "0.9062", "0.881", "0.9194", "0.511", "0.9194" ] ]

[ "预培训什么？", "高效的中间任务选择", "Clifton Poth, Jonas Pfeiffer, Andreas Rickle, and Iryna Gurevych", "Ubiquitous Knowledge Processing Lab, Technical University of Darmstadt", "www.ukp.tu-darmstadt.de", "Abstract", "块化，可实现更广泛的迁移学习技术（Pfeiffer", "et al., 2020b, 2021a,b; Ustun et al., 2020; Vidoni", "Intermediate task fine-tuning has been shown", "et al., 2020; Rust et al., 2021; Ansell et al., 2021)", "to culminate in large transfer gains across", "在已建立的两步学习过程的基础上进行扩", "many NLP tasks. With an abundance of can-", "展，结合知识转移的中间阶段可以为完全微调", "didate datasets as well as pre-trained language", "models, it has become infeasible to experiment", "的模型带来进一步的收益。例如，Phang etal", "withall combinations to find thebesttransfer", "(2018)在目标任务微调之前，在兼容的中间任", "setting. In this work, we provide a compre-", "务上顺序微调预训练的语言模型。已经表明", "hensive comparison of different methods for", "这对于低资源目标任务是最有效的，但是，并", "efficiently identifying beneficial tasks for inter-", "非所有任务组合都是有益的，并且许多收益降", "mediate transfer learning. We focus on param-", "低了性能(Phang et al.,2018;Wang et al.,2019a;", "eter and computationally efficient adapter set-", "Pruksachatkunetal.，2020）。丰富的不同标记数", "tings, highlight different data-availability sce", "据集以及新的预训练LM的不断发展需要有效", "narios, and provide expense estimates for each", "method. We experiment with a diverse set", "地识别有利于目标任务的中间数据集的方法。", "of 42 intermediate and 11 target English clas", "到目前为止，尚不清楚基于适配器的方法如何", "sification , multiple choice, question answer-", "进行中间微调。因此，在这项工作的第一部分", "ing , and sequence tagging tasks. Our results", "中，我们确定此设置可为适配器带来类似的增", "demonstrate that efficient embedding based", "益，如完整模型微调（Phangetal.，2018；Pruk-", "methods, which rely solely on the respective", "sachatkun et al., 2020; Gururangan et al., 2020) 所", "datasets, outperform computational expensive", "示。专注于低资源目标任务设置，我们发现只", "few-shot fine-tuning approaches.", "Our best", "methods achieve an average Regret@3 of 1 %", "有一部分中间适配器产生正增益，而其他适配", "across all target tasks, demonstrating that we", "器则大大损害了性能（见表1和图2）。我们", "are able to efficiently identify the best datasets", "的结果表明，有必要获得有效识别有益的中间", "for intermediate training.", "训练适配器的方法", "在第二部分中，我们利用第一部分的转移", "1", "Introduction", "结果自动对有益的中间任务进行排名和识别。", "大型预训练语言模型（LM不断推动各种NLP", "随着NLP模型（Wolf etal.,2020;Pfeifferetal.，", "2020a）的大型可公开访问的存储库的兴起，找", "任务的发展。建立的过程在大型文本语料库", "到产生正转移收益的预训练模型的机会很高。", "上执行自监督预训练，然后在特定目标任务", "然而，强行确定最佳中间任务是不可行的。现有", "(Devlinetal.,2019;Liuetal.,2019b)上微调模", "型。相同的过程也已应用于基于适配器的训", "方法的重点是多任务学习（BingelandSogaard,", "2017）的有益任务选择，NLP任务（Vuetal.,", "练策略，该策略可实现与完整模型微调相同的", "2020）的基于中间和目标变压器的LM的完全", "任务性能，同时具有更高的参数效率（Houlsby", "etal.,2019)和更快的训练速度E(Ruickle etal..", "微调，视觉任务（Puigcerveretal.，2021）的基于", "2021）。2除了效率更高，适配器还具有高度模", "适配器的模型以及无监督的零拍方法社区问答", "(Rickléetal.,2020）的转移。这些工作中的每", "*Contributions made prior to joining Amazon", "个都需要不同类型的数据，例如中间任务数", "'Code", "released", "at", "https://github.com/", "据和/或中间模型权重，根据情况，这些数据可", "Adapter-Hub/efficient-task-transfer", "2Adapters are new weights at every layer of a pre-trained", "能无法访问。3", "transformer model.To fine-tune a model on a downstream", "task, all pre-trained transformer weights are frozen and only", "3Bingel and Sogaard (2017) and Vu et al. (2020) require", "the newly introduced adapter weights are trained.", "access to both intermediate task data and models, Puigcerver" ]

{ "arxiv_id": "2104.08247", "language": "zh", "page_idx": 1 }

[ [ "0.1377", "0.0885", "0.486", "0.0885", "0.486", "0.0986", "0.1377", "0.0986" ], [ "0.513", "0.0873", "0.876", "0.0873", "0.876", "0.0974", "0.513", "0.0974" ], [ "0.11755", "0.1027", "0.482", "0.1027", "0.482", "0.1164", "0.11755", "0.1164" ], [ "0.513", "0.1027", "0.879", "0.1027", "0.879", "0.1164", "0.513", "0.1164" ], [ "0.12006", "0.11816", "0.4844", "0.11816", "0.4844", "0.1318", "0.12006", "0.1318" ], [ "0.5156", "0.12054", "0.879", "0.12054", "0.879", "0.1312", "0.5156", "0.1312" ], [ "0.1184", "0.1342", "0.4844", "0.1342", "0.4844", "0.1478", "0.1184", "0.1478" ], [ "0.5146", "0.1372", "0.879", "0.1372", "0.879", "0.1478", "0.5146", "0.1478" ], [ "0.1184", "0.1497", "0.4844", "0.1509", "0.4844", "0.165", "0.1184", "0.1644" ], [ "0.511", "0.1497", "0.881", "0.1509", "0.881", "0.1644", "0.511", "0.1633" ], [ "0.1184", "0.1675", "0.486", "0.1675", "0.486", "0.1812", "0.1184", "0.1812" ], [ "0.513", "0.1675", "0.881", "0.1675", "0.881", "0.1812", "0.513", "0.1812" ], [ "0.1184", "0.1829", "0.486", "0.1818", "0.486", "0.1953", "0.1184", "0.1965" ], [ "0.5146", "0.1829", "0.879", "0.1829", "0.879", "0.1965", "0.5146", "0.1965" ], [ "0.1184", "0.1996", "0.4844", "0.1996", "0.4844", "0.2131", "0.1184", "0.2131" ], [ "0.51", "0.1978", "0.881", "0.199", "0.881", "0.2131", "0.51", "0.212" ], [ "0.12006", "0.2156", "0.4844", "0.2156", "0.4844", "0.2286", "0.12006", "0.2286" ], [ "0.513", "0.2156", "0.879", "0.2156", "0.879", "0.2286", "0.513", "0.2286" ], [ "0.1184", "0.2322", "0.4878", "0.2322", "0.4878", "0.2452", "0.1184", "0.2452" ], [ "0.511", "0.2322", "0.881", "0.2322", "0.881", "0.2452", "0.511", "0.2452" ], [ "0.1184", "0.2477", "0.4878", "0.2477", "0.4878", "0.2612", "0.1184", "0.2612" ], [ "0.51", "0.2477", "0.879", "0.2477", "0.879", "0.2612", "0.51", "0.2612" ], [ "0.1184", "0.2642", "0.4844", "0.2642", "0.4844", "0.2778", "0.1184", "0.2778" ], [ "0.51", "0.2625", "0.882", "0.2632", "0.882", "0.2778", "0.51", "0.2766" ], [ "0.1184", "0.2798", "0.486", "0.2798", "0.486", "0.2935", "0.1184", "0.2935" ], [ "0.513", "0.2808", "0.6094", "0.2808", "0.6094", "0.2922", "0.513", "0.2922" ], [ "0.1184", "0.2957", "0.4844", "0.2957", "0.4844", "0.3088", "0.1184", "0.3088" ], [ "0.532", "0.2957", "0.879", "0.2957", "0.879", "0.3088", "0.532", "0.3088" ], [ "0.12006", "0.3135", "0.486", "0.3135", "0.486", "0.3237", "0.12006", "0.3237" ], [ "0.513", "0.3123", "0.881", "0.3123", "0.881", "0.326", "0.513", "0.326" ], [ "0.1184", "0.328", "0.4844", "0.328", "0.4844", "0.3416", "0.1184", "0.3416" ], [ "0.513", "0.328", "0.881", "0.328", "0.881", "0.3416", "0.513", "0.3416" ], [ "0.1184", "0.3445", "0.3652", "0.3445", "0.3652", "0.3582", "0.1184", "0.3582" ], [ "0.511", "0.3425", "0.882", "0.3433", "0.882", "0.3582", "0.511", "0.357" ], [ "0.513", "0.3599", "0.881", "0.3599", "0.881", "0.3735", "0.513", "0.3735" ], [ "0.1142", "0.3718", "0.1587", "0.3718", "0.1587", "0.386", "0.1142", "0.386" ], [ "0.1503", "0.3723", "0.2747", "0.3723", "0.2747", "0.386", "0.1503", "0.386" ], [ "0.5156", "0.3777", "0.879", "0.3777", "0.879", "0.3884", "0.5156", "0.3884" ], [ "0.11584", "0.3984", "0.346", "0.3984", "0.346", "0.412", "0.11584", "0.412" ], [ "0.513", "0.3938", "0.54", "0.3938", "0.54", "0.4062", "0.513", "0.4062" ], [ "0.1184", "0.4216", "0.486", "0.4216", "0.486", "0.4316", "0.1184", "0.4316" ], [ "0.4978", "0.4233", "0.5054", "0.4233", "0.5054", "0.4287", "0.4978", "0.4287" ], [ "0.5073", "0.4204", "0.8506", "0.4204", "0.8506", "0.434", "0.5073", "0.434" ], [ "0.1184", "0.4358", "0.4844", "0.4358", "0.4844", "0.4495", "0.1184", "0.4495" ], [ "0.511", "0.4453", "0.879", "0.4453", "0.879", "0.459", "0.511", "0.459" ], [ "0.11755", "0.452", "0.4844", "0.4507", "0.4844", "0.4644", "0.11755", "0.4656" ], [ "0.511", "0.4614", "0.879", "0.4614", "0.879", "0.4744", "0.511", "0.4744" ], [ "0.1184", "0.4685", "0.486", "0.4685", "0.486", "0.4822", "0.1184", "0.4822" ], [ "0.513", "0.478", "0.881", "0.478", "0.881", "0.4912", "0.513", "0.4912" ], [ "0.11755", "0.484", "0.4878", "0.4827", "0.4878", "0.4963", "0.11755", "0.4976" ], [ "0.51", "0.4922", "0.6245", "0.4934", "0.6245", "0.5083", "0.51", "0.5073" ], [ "0.12006", "0.502", "0.4844", "0.502", "0.4844", "0.5127", "0.12006", "0.5127" ], [ "0.11755", "0.5166", "0.486", "0.5166", "0.486", "0.53", "0.11755", "0.53" ], [ "0.509", "0.5186", "0.6055", "0.521", "0.6055", "0.535", "0.509", "0.533" ], [ "0.12006", "0.532", "0.4844", "0.532", "0.4844", "0.546", "0.12006", "0.546" ], [ "0.513", "0.5415", "0.879", "0.5415", "0.879", "0.5547", "0.513", "0.5547" ], [ "0.12006", "0.5503", "0.4844", "0.5503", "0.4844", "0.561", "0.12006", "0.561" ], [ "0.11755", "0.5625", "0.1411", "0.5664", "0.1368", "0.58", "0.1125", "0.5767" ], [ "0.513", "0.557", "0.879", "0.557", "0.879", "0.571", "0.513", "0.571" ], [ "0.5146", "0.5737", "0.8774", "0.5737", "0.8774", "0.5874", "0.5146", "0.5874" ], [ "0.1368", "0.5806", "0.4878", "0.5806", "0.4878", "0.594", "0.1368", "0.594" ], [ "0.511", "0.589", "0.879", "0.589", "0.879", "0.6025", "0.511", "0.6025" ], [ "0.1184", "0.5967", "0.4878", "0.5967", "0.4878", "0.61", "0.1184", "0.61" ], [ "0.5146", "0.6055", "0.881", "0.6055", "0.881", "0.619", "0.5146", "0.619" ], [ "0.1184", "0.6133", "0.3762", "0.6133", "0.3762", "0.627", "0.1184", "0.627" ], [ "0.513", "0.6206", "0.881", "0.6226", "0.881", "0.6357", "0.513", "0.6343" ], [ "0.11755", "0.641", "0.4634", "0.641", "0.4634", "0.655", "0.11755", "0.655" ], [ "0.513", "0.638", "0.881", "0.638", "0.881", "0.652", "0.513", "0.652" ], [ "0.51", "0.653", "0.6543", "0.652", "0.6543", "0.6665", "0.51", "0.6675" ], [ "0.1184", "0.662", "0.482", "0.662", "0.482", "0.673", "0.1184", "0.673" ], [ "0.1184", "0.6777", "0.4475", "0.6777", "0.4475", "0.6914", "0.1184", "0.6914" ], [ "0.511", "0.6797", "0.7153", "0.681", "0.7153", "0.6953", "0.511", "0.695" ], [ "0.1377", "0.695", "0.486", "0.695", "0.486", "0.708", "0.1377", "0.708" ], [ "0.513", "0.7017", "0.881", "0.7017", "0.881", "0.7153", "0.513", "0.7153" ], [ "0.11755", "0.7104", "0.4844", "0.709", "0.4844", "0.7236", "0.11755", "0.7246" ], [ "0.511", "0.7163", "0.876", "0.7163", "0.876", "0.73", "0.511", "0.73" ], [ "0.12006", "0.728", "0.4836", "0.728", "0.4836", "0.7383", "0.12006", "0.7383" ], [ "0.5146", "0.7354", "0.879", "0.7354", "0.879", "0.745", "0.5146", "0.745" ], [ "0.12006", "0.7446", "0.4844", "0.7446", "0.4844", "0.755", "0.12006", "0.755" ], [ "0.513", "0.75", "0.881", "0.75", "0.881", "0.763", "0.513", "0.763" ], [ "0.1184", "0.759", "0.486", "0.759", "0.486", "0.7725", "0.1184", "0.7725" ], [ "0.5146", "0.767", "0.879", "0.767", "0.879", "0.778", "0.5146", "0.778" ], [ "0.1184", "0.775", "0.4844", "0.775", "0.4844", "0.788", "0.1184", "0.788" ], [ "0.513", "0.782", "0.881", "0.782", "0.881", "0.796", "0.513", "0.796" ], [ "0.1184", "0.7915", "0.4844", "0.7915", "0.4844", "0.8047", "0.1184", "0.8047" ], [ "0.513", "0.799", "0.881", "0.799", "0.881", "0.8125", "0.513", "0.8125" ], [ "0.12006", "0.807", "0.486", "0.807", "0.486", "0.821", "0.12006", "0.821" ], [ "0.511", "0.814", "0.8774", "0.814", "0.8774", "0.8276", "0.511", "0.8276" ], [ "0.1184", "0.8237", "0.486", "0.8237", "0.486", "0.8374", "0.1184", "0.8374" ], [ "0.513", "0.8306", "0.881", "0.8306", "0.881", "0.844", "0.513", "0.844" ], [ "0.1184", "0.839", "0.4492", "0.839", "0.4492", "0.8525", "0.1184", "0.8525" ], [ "0.511", "0.8457", "0.6885", "0.844", "0.6885", "0.8594", "0.511", "0.86" ], [ "0.1351", "0.855", "0.4844", "0.855", "0.4844", "0.868", "0.1351", "0.868" ], [ "0.1184", "0.8716", "0.486", "0.8716", "0.486", "0.8853", "0.1184", "0.8853" ], [ "0.532", "0.868", "0.881", "0.868", "0.881", "0.882", "0.532", "0.882" ], [ "0.5146", "0.8813", "0.879", "0.8813", "0.879", "0.8945", "0.5146", "0.8945" ], [ "0.1184", "0.8955", "0.4844", "0.8955", "0.4844", "0.9062", "0.1184", "0.9062" ], [ "0.511", "0.8936", "0.7246", "0.8936", "0.7246", "0.9067", "0.511", "0.9067" ], [ "0.1184", "0.9062", "0.471", "0.9062", "0.471", "0.9194", "0.1184", "0.9194" ], [ "0.534", "0.9062", "0.817", "0.9062", "0.817", "0.9194", "0.534", "0.9194" ] ]

[ "因此，在这项工作中，我们旨在从多个不同", "构建视觉任务的分类，深入了解任务之间的", "的角度解决NLP中迁移学习的效率问题，从", "非平凡转移关系。多个作品建议使用嵌入来捕", "而做出以下贡献：1)E我们专注于基于适配器", "获统计数据，特征或数据集的域。Edwardsand", "的迁移学习，它的参数(Houlsbyetal.,2019)和", "Storkey（2017）利用变分自动编码器（Kingma", "计算效率比全模型微调(Rickleetal.,2021）高", "andWelling，2014）对数据集的所有样本进行", "得多，同时实现了同等的性能；2)E我们评估", "编码。Jomaaetal.(2019）训练一个可以成功捕", "了基于适配器的方法的顺序微调，并针对42", "获数据集域的数据集元特征提取器。Vuetal", "个不同。e。分类，多项选择，问题回答和序列", "(2020）通过对最后一层的BERT表示进行平均", "标记）；3)E我们确定了迁移学习的最佳中间", "来编码数据集的每个训练示例。Rickleetal.", "任务，而无需对所有潜在候选人进行计算昂贵", "(2020）通过使用句子嵌人模型嵌入数据集示", "的显式训练。我们比较了不同的选择技术，合", "例来捕获域相似性。Achilleetal.(2019）和Vu", "并了以前提出的和新的方法；4)E我们对不同", "etal.(2020)基于探测网络的Fisher信息矩阵计", "的技术，可用的数据方案以及任务和模型类型", "算任务嵌入。", "进行了全面的分析，从而为每种设置的最佳方", "虽然已经提出了许多不同的方法，但它们之", "法提供了更深入的见解；5E我们提供计算成", "间缺乏直接的比较。此外，以前的工作只关注", "本估算，从而可以做出明智的决策，以在费用", "BERT，我们发现它的行为与其他模型类型有", "和下游任务性能之间进行权衡。", "很大的不同，例如某些方法的RoBERTa。在", "这项工作中，我们的目标是整合所有方法，并", "21", "Related Work", "尝试使用较新的模型类型，以提供更全面的视", "2.1Transfer between tasks", "角。", "Phangetal.（2018）显示，对中间任务进行培训", "", "3Adapter-Based Sequential Transfer", "可提高许多目标任务的性能。后续工作进一步", "我们对基于适配器的顺序微调进行了大规模", "探讨了对更多不同任务集（Wangetal.2019a", "研究，发现大约一半的任务组合不会产生正收", "Talmor and Berant, 2019; Liu et al., 2019a; Sap", "益。这表明了找到有效识别合适的中间任务的", "et al., 2019; Pruksachatkun et al., 2020; Vu et al.,", "方法的重要性", "2020）的影响。Wangetal.（2019a），Yogatama", "et al.(2019)和Pruksachatkun et al.(2020)强调", "3.1Tasks", "了灾难性遗忘和负转移结果的风险，发现在考", "我们从MultiQA存储库（TalmorandBerant，", "虑不同的中间任务时，顺序转移的成功差异很", "大", "2019）中选择QA任务，并从Liuetal.（2019a）", "中顺序标记任务。我们的大多数分类任务都可", "虽然以前的工作表明，中间任务训练可以在", "以在 (Super）GLUE(Wang et al.,2018,2019b)", "完全微调设置中提高目标任务的性能，但我们", "基准测试中使用。我们尝试多项选择常识推理", "确定基于适配器的训练也是如此", "任务，以涵盖更广泛的不同类型和领域。我们", "2.2 Predicting Beneficial Transfer Sources", "总共实验了53个任务，分为42个中间任务和", "11个目标任务。4", "在考虑到任务和模型的可用性不断提高时，自", "动选择产生传输增益的中间任务至关重要", "3.2ExperimentalSetup", "已提出代理估计器来评估预训练模型对目标", "我们使用基于BERT的(Devlinetal.,2019）和", "任务的可传递性。Nguyenetal.（2020），Lietal.", "基于RoBERTA的（Liuetal,2019b）进行实验", "（2021）和Deshpandeetal.（2021）通过从源和日", "训练具有Pfeifferetal.（202la）建议的配置的适", "标任务标签分布构建经验分类器来估计分类任", "配器。我们采用Phangetal.（2018）的两阶段顺", "务之间的可传递性。具有多种模型选择方法的", "序微调设置，将任务分为两个不相交的子集S", "Puigcerveretal.（2021）实验，包括用于估计目", "和T，分别表示为中间任务和目标任务。对于", "标任务性能的kNN代理模型。在类似的方向", "具有sES和tET的每一对（s,t），我们首", "上，Rengglietal.2020）研究了基于kNN和线", "先在s上训练一个随机初始化的适配器（保持", "性分类器的代理模型，发现任务感知和任务不", "基本模型的参数固定）。然后，我们在t上微调", "可知策略的混合方法组合产生了最佳结果", "训练有素的适配器。5", "BingelandSogaard（2017）发现学习曲线的梯", "度与多任务学习成功相关。Zamiretal.(2018)", "4The choice for our intermediate and target task split was", "motivated by previous work (Sap et al., 2019;?, inter alia)", "et al. (2021)require access to only the intermediate model,", "For more details see Appendix A .", "and Rickle et al. (2020) only to the intermediate task data.", "°For more details please refer to Appendix B ." ]

{ "arxiv_id": "2104.08247", "language": "zh", "page_idx": 2 }

[ [ "0.5146", "0.0885", "0.5483", "0.0885", "0.5483", "0.0968", "0.5146", "0.0968" ], [ "0.5425", "0.0861", "0.869", "0.0861", "0.869", "0.0998", "0.5425", "0.0998" ], [ "0.544", "0.1051", "0.7095", "0.1051", "0.7095", "0.1152", "0.544", "0.1152" ], [ "0.1301", "0.08905", "0.147", "0.08905", "0.147", "0.0974", "0.1301", "0.0974" ], [ "0.1301", "0.10986", "0.1453", "0.10986", "0.1453", "0.11816", "0.1301", "0.11816" ], [ "0.4375", "0.1069", "0.4517", "0.1051", "0.4534", "0.10986", "0.4382", "0.11224" ], [ "0.3896", "0.1223", "0.4382", "0.1134", "0.444", "0.1277", "0.3955", "0.1372" ], [ "0.1368", "0.133", "0.1453", "0.1372", "0.1394", "0.1426", "0.1318", "0.1383" ], [ "0.513", "0.13", "0.8774", "0.13", "0.8774", "0.1401", "0.513", "0.1401" ], [ "0.1066", "0.1395", "0.1503", "0.136", "0.1545", "0.1776", "0.11084", "0.1799" ], [ "0.5146", "0.1443", "0.881", "0.1443", "0.881", "0.158", "0.5146", "0.158" ], [ "0.513", "0.1609", "0.881", "0.1609", "0.881", "0.1746", "0.513", "0.1746" ], [ "0.315", "0.1716", "0.3525", "0.1644", "0.3552", "0.1733", "0.318", "0.1812" ], [ "0.4148", "0.1733", "0.455", "0.1733", "0.455", "0.1953", "0.4148", "0.1953" ], [ "0.5146", "0.177", "0.881", "0.177", "0.881", "0.1907", "0.5146", "0.1907" ], [ "0.294", "0.1776", "0.3123", "0.1776", "0.3123", "0.1818", "0.294", "0.1818" ], [ "0.12177", "0.1758", "0.1436", "0.1758", "0.1436", "0.1882", "0.12177", "0.1882" ], [ "0.4375", "0.1787", "0.444", "0.1841", "0.4375", "0.1895", "0.43", "0.1841" ], [ "0.409", "0.1853", "0.4165", "0.1859", "0.413", "0.1948", "0.4055", "0.1936" ], [ "0.4055", "0.1936", "0.4182", "0.1936", "0.4182", "0.2037", "0.4055", "0.2037" ], [ "0.4072", "0.1924", "0.476", "0.1733", "0.492", "0.2031", "0.4224", "0.2214" ], [ "0.513", "0.1936", "0.8774", "0.1936", "0.8774", "0.2037", "0.513", "0.2037" ], [ "0.1066", "0.2031", "0.1394", "0.1907", "0.1528", "0.2079", "0.12006", "0.2203" ], [ "0.403", "0.2031", "0.4148", "0.2007", "0.419", "0.2102", "0.4072", "0.212" ], [ "0.4248", "0.2114", "0.482", "0.2114", "0.482", "0.2227", "0.4248", "0.2227" ], [ "0.513", "0.209", "0.742", "0.209", "0.742", "0.2227", "0.513", "0.2227" ], [ "0.4224", "0.2173", "0.455", "0.2197", "0.4534", "0.2322", "0.4207", "0.2299" ], [ "0.4114", "0.2203", "0.419", "0.2257", "0.4114", "0.2299", "0.4048", "0.2245" ], [ "0.1301", "0.2227", "0.1453", "0.2227", "0.1453", "0.2322", "0.1301", "0.2322" ], [ "0.409", "0.2131", "0.4148", "0.2131", "0.4148", "0.2197", "0.409", "0.2197" ], [ "0.409", "0.2299", "0.4182", "0.2328", "0.413", "0.2393", "0.4048", "0.2363" ], [ "0.4248", "0.228", "0.4534", "0.228", "0.4534", "0.2405", "0.4248", "0.2405" ], [ "0.4048", "0.2363", "0.4207", "0.2363", "0.4207", "0.2588", "0.4048", "0.2588" ], [ "0.4265", "0.2328", "0.455", "0.2382", "0.451", "0.2517", "0.4224", "0.2465" ], [ "0.513", "0.2369", "0.5557", "0.2369", "0.5557", "0.2488", "0.513", "0.2488" ], [ "0.5474", "0.2363", "0.7715", "0.2363", "0.7715", "0.2494", "0.5474", "0.2494" ], [ "0.419", "0.2465", "0.4517", "0.2465", "0.4517", "0.2588", "0.419", "0.2588" ], [ "0.4072", "0.2559", "0.4182", "0.2559", "0.4182", "0.2737", "0.4072", "0.2737" ], [ "0.4248", "0.2546", "0.4534", "0.2546", "0.4534", "0.2683", "0.4248", "0.2683" ], [ "0.529", "0.2583", "0.879", "0.2583", "0.879", "0.2715", "0.529", "0.2715" ], [ "0.4248", "0.2642", "0.4802", "0.2642", "0.4802", "0.2756", "0.4248", "0.2756" ], [ "0.1335", "0.2666", "0.1503", "0.2683", "0.147", "0.2778", "0.1301", "0.2756" ], [ "0.513", "0.2737", "0.8774", "0.2737", "0.8774", "0.2874", "0.513", "0.2874" ], [ "0.1763", "0.279", "0.2024", "0.279", "0.2024", "0.2915", "0.1763", "0.2915" ], [ "0.2568", "0.26", "0.288", "0.26", "0.288", "0.2715", "0.2568", "0.2715" ], [ "0.2167", "0.2808", "0.2317", "0.2808", "0.2317", "0.2903", "0.2167", "0.2903" ], [ "0.3333", "0.2808", "0.351", "0.2808", "0.351", "0.288", "0.3333", "0.288" ], [ "0.3745", "0.2808", "0.3914", "0.2808", "0.3914", "0.2893", "0.3745", "0.2893" ], [ "0.4182", "0.2798", "0.4324", "0.2798", "0.4324", "0.2893", "0.4182", "0.2893" ], [ "0.4568", "0.2808", "0.471", "0.2808", "0.471", "0.2893", "0.4568", "0.2893" ], [ "0.293", "0.282", "0.3132", "0.282", "0.3132", "0.2893", "0.293", "0.2893" ], [ "0.2612", "0.2832", "0.2703", "0.2832", "0.2703", "0.288", "0.2612", "0.288" ], [ "0.2612", "0.2915", "0.378", "0.2915", "0.378", "0.2986", "0.2612", "0.2986" ], [ "0.517", "0.2915", "0.881", "0.2915", "0.881", "0.3018", "0.517", "0.3018" ], [ "0.11755", "0.3088", "0.4836", "0.3088", "0.4836", "0.3196", "0.11755", "0.3196" ], [ "0.513", "0.306", "0.8447", "0.306", "0.8447", "0.3196", "0.513", "0.3196" ], [ "0.11584", "0.3218", "0.4895", "0.3237", "0.4895", "0.3374", "0.11584", "0.335" ], [ "0.5146", "0.326", "0.881", "0.326", "0.881", "0.339", "0.5146", "0.339" ], [ "0.11755", "0.3374", "0.3594", "0.3384", "0.3594", "0.3499", "0.11755", "0.3486" ], [ "0.5146", "0.3416", "0.881", "0.3416", "0.881", "0.3552", "0.5146", "0.3552" ], [ "0.5146", "0.3594", "0.7573", "0.3594", "0.7573", "0.3694", "0.5146", "0.3694" ], [ "0.1351", "0.3682", "0.486", "0.3682", "0.486", "0.3818", "0.1351", "0.3818" ], [ "0.5146", "0.3777", "0.881", "0.3777", "0.881", "0.3914", "0.5146", "0.3914" ], [ "0.12006", "0.386", "0.4836", "0.386", "0.4836", "0.3967", "0.12006", "0.3967" ], [ "0.5156", "0.3938", "0.879", "0.3938", "0.879", "0.4067", "0.5156", "0.4067" ], [ "0.12006", "0.4026", "0.4836", "0.4026", "0.4836", "0.4133", "0.12006", "0.4133" ], [ "0.513", "0.4104", "0.881", "0.4104", "0.881", "0.4233", "0.513", "0.4233" ], [ "0.1184", "0.4163", "0.4844", "0.415", "0.4844", "0.43", "0.1184", "0.4312" ], [ "0.511", "0.4246", "0.7153", "0.4258", "0.7153", "0.4407", "0.511", "0.4395" ], [ "0.11755", "0.4329", "0.4844", "0.4329", "0.4844", "0.4465", "0.11755", "0.4465" ], [ "0.1184", "0.449", "0.4836", "0.449", "0.4836", "0.462", "0.1184", "0.462" ], [ "0.51", "0.4531", "0.7446", "0.4531", "0.7446", "0.466", "0.51", "0.466" ], [ "0.11755", "0.4656", "0.4836", "0.4656", "0.4836", "0.4785", "0.11755", "0.4785" ], [ "0.513", "0.4739", "0.8774", "0.4739", "0.8774", "0.488", "0.513", "0.488" ], [ "0.11584", "0.4797", "0.2024", "0.4822", "0.2024", "0.4963", "0.11584", "0.4946" ], [ "0.5146", "0.4905", "0.879", "0.4905", "0.879", "0.5044", "0.5146", "0.5044" ], [ "0.513", "0.5073", "0.885", "0.5073", "0.885", "0.521", "0.513", "0.521" ], [ "0.11584", "0.514", "0.2225", "0.514", "0.2225", "0.5293", "0.11584", "0.5293" ], [ "0.513", "0.5225", "0.75", "0.5225", "0.75", "0.536", "0.513", "0.536" ], [ "0.1184", "0.539", "0.4844", "0.539", "0.4844", "0.5527", "0.1184", "0.5527" ], [ "0.5146", "0.546", "0.881", "0.546", "0.881", "0.5586", "0.5146", "0.5586" ], [ "0.11755", "0.5547", "0.486", "0.5547", "0.486", "0.5684", "0.11755", "0.5684" ], [ "0.5146", "0.5635", "0.8774", "0.5635", "0.8774", "0.5737", "0.5146", "0.5737" ], [ "0.11755", "0.5693", "0.486", "0.571", "0.486", "0.585", "0.11755", "0.5845" ], [ "0.513", "0.5776", "0.881", "0.5776", "0.881", "0.5913", "0.513", "0.5913" ], [ "0.1184", "0.5874", "0.486", "0.5874", "0.486", "0.601", "0.1184", "0.601" ], [ "0.5146", "0.5957", "0.879", "0.5957", "0.879", "0.6055", "0.5146", "0.6055" ], [ "0.1184", "0.6025", "0.486", "0.6025", "0.486", "0.616", "0.1184", "0.616" ], [ "0.513", "0.6094", "0.86", "0.6094", "0.86", "0.6226", "0.513", "0.6226" ], [ "0.11755", "0.618", "0.4836", "0.618", "0.4836", "0.632", "0.11755", "0.632" ], [ "0.1184", "0.6357", "0.486", "0.6357", "0.486", "0.6494", "0.1184", "0.6494" ], [ "0.513", "0.633", "0.881", "0.633", "0.881", "0.6465", "0.513", "0.6465" ], [ "0.11755", "0.651", "0.4844", "0.651", "0.4844", "0.6646", "0.11755", "0.6646" ], [ "0.513", "0.651", "0.879", "0.651", "0.879", "0.661", "0.513", "0.661" ], [ "0.12006", "0.6694", "0.4844", "0.6694", "0.4844", "0.6797", "0.12006", "0.6797" ], [ "0.511", "0.6646", "0.879", "0.6646", "0.879", "0.6777", "0.511", "0.6777" ], [ "0.1184", "0.684", "0.4878", "0.684", "0.4878", "0.6978", "0.1184", "0.6978" ], [ "0.513", "0.6816", "0.843", "0.6816", "0.843", "0.6953", "0.513", "0.6953" ], [ "0.11755", "0.6997", "0.4844", "0.6987", "0.4844", "0.712", "0.11755", "0.7134" ], [ "0.513", "0.708", "0.7173", "0.708", "0.7173", "0.7217", "0.513", "0.7217" ], [ "0.1184", "0.7163", "0.486", "0.7163", "0.486", "0.73", "0.1184", "0.73" ], [ "0.1184", "0.731", "0.486", "0.731", "0.486", "0.7446", "0.1184", "0.7446" ], [ "0.511", "0.7285", "0.881", "0.7285", "0.881", "0.742", "0.511", "0.742" ], [ "0.11755", "0.7476", "0.486", "0.7466", "0.486", "0.761", "0.11755", "0.7627" ], [ "0.511", "0.7466", "0.881", "0.7466", "0.881", "0.7603", "0.511", "0.7603" ], [ "0.12006", "0.764", "0.486", "0.764", "0.486", "0.778", "0.12006", "0.778" ], [ "0.511", "0.761", "0.8774", "0.761", "0.8774", "0.775", "0.511", "0.775" ], [ "0.1184", "0.781", "0.4878", "0.781", "0.4878", "0.7944", "0.1184", "0.7944" ], [ "0.511", "0.7793", "0.876", "0.7793", "0.876", "0.789", "0.511", "0.789" ], [ "0.12006", "0.799", "0.465", "0.799", "0.465", "0.8086", "0.12006", "0.8086" ], [ "0.513", "0.7944", "0.879", "0.7944", "0.879", "0.808", "0.513", "0.808" ], [ "0.1368", "0.814", "0.4878", "0.814", "0.4878", "0.8276", "0.1368", "0.8276" ], [ "0.511", "0.8125", "0.8774", "0.8125", "0.8774", "0.8223", "0.511", "0.8223" ], [ "0.11755", "0.8306", "0.4844", "0.8306", "0.4844", "0.844", "0.11755", "0.844" ], [ "0.511", "0.8267", "0.879", "0.8267", "0.879", "0.8403", "0.511", "0.8403" ], [ "0.11755", "0.8467", "0.4878", "0.8467", "0.4878", "0.86", "0.11755", "0.86" ], [ "0.5156", "0.844", "0.7153", "0.844", "0.7153", "0.855", "0.5156", "0.855" ], [ "0.1184", "0.8633", "0.3997", "0.8633", "0.3997", "0.877", "0.1184", "0.877" ], [ "0.5146", "0.866", "0.879", "0.866", "0.879", "0.88", "0.5146", "0.88" ], [ "0.5146", "0.882", "0.879", "0.882", "0.879", "0.8955", "0.5146", "0.8955" ], [ "0.1368", "0.891", "0.4878", "0.8936", "0.4878", "0.9077", "0.1368", "0.9062" ], [ "0.11584", "0.904", "0.2795", "0.905", "0.2795", "0.9194", "0.11584", "0.9185" ], [ "0.5356", "0.9062", "0.781", "0.9062", "0.781", "0.9194", "0.5356", "0.9194" ] ]

[ "4", "Methods for the Efficient Selection of", "lIntermediateTasks", "90", "80", "", "", "10", "我们现在介绍不同的模型选择方法，稍后在85", "", "中，研究它们在上述设置中的有效性。我们根", "据他们对中间任务数据Ds和中间模型Ms的", "福", "CTPA", "可用性所做的假设对不同的方法进行分组。在", "", "50", "V", "8", ".", "eose", "考虑具有数百个任务的大型预训练模型存储库", "40", "+", "CoNLL2003", "时，访问两者可能会很昂贵。", "CSQA", "M", "30", "4", "?", "STS-B", "", "DROP", "4.1", "Metadata: Educated Guess", "Quail", "丰电", "Boolo", "种设置，其中既不存在对中间任务数据Ds", "DepRel-EWT", "10", "的访问，也不存在对数据Ms进行训练的模型", "20", "", "30", "60", "70", "80", "90", "50", "40", "PerformancewithRoBERTa", "可以将其视为有根据的猜测方案。选择标准只", "Figurel:Comparison oftransferperformancebetween", "能依赖于可用于中间任务数据集的元数据", "BERT and RoBERTaforthe respective target tasks, pre-", "数据集大小。在更多数据意味着更好的传输性", "trainedonthe42intermediatetasks", "能的假设下，表示为Size的选择标准按训练数", "据大小降序排列所有中间任务。", "对于目标任务微调，我们通过将t上的最", "任务类型。在相似目标函数传递良好的假设", "大训练示例数限制为1000来模拟低资源设置。", "下，我们预先选择相同类型的任务子集。这种", "此选择的动机是观察到较小的目标任务从顺序", "方法可以与剩余任务的随机选择相结合，或者", "微调中受益最大，同时揭示了最大的性能差异", "按照大小对它们进行排名", "(Phangetal.,2018;Vuetal.,2020)。因此，低资", "源设置反映了我们迁移学习策略最有利的应用", "4.2IntermediateTaskData", "程序设置，也使我们能够更彻底地研究不同的", "随着大量可用数据集7和新LM的不断开发", "迁移关系。", "每个任务-模型组合的微调版本（立即）不可", "用。因此，以下方法利用了中间任务数据Ds，", "3.3Results", "而不需要相应的微调模型Ms", "图E2显示了相对转移收益，表E1列出了", "文本嵌人（TextEmb)。Vuetal.(2020）通过LM", "RoBERTa的所有中间和目标任务组合的绝对", "传递每个示例，并对最终层的输出表示进行平", "得分。6我们观察到不同组合的转移收益（和损", "均（跨所有示例和所有输人令牌）。假设相似", "失)有很大的变化。即使较大的差异可以通过", "的嵌人意味着正的转移增益，他们根据嵌人与", "较高的任务难度来解释（请参见表1中的“无", "目标任务的余弦相似度对中间任务进行排序", "转移”），它们也说明了基于适配器的设置中顺", "序微调的异构性和潜力。同时，我们发现了几", "SBERT嵌人(SEmb)。诸如Sentence-BERT", "种传输损失的情况---多达60个%的性能较低", "（SBERT：？之类的句子入模型可能更适合", "（请参见图E2）---可能是由于灾难性遗忘而发", "表示数据集示例。与TextEmb类似，我们根据", "生的。总体而言，对于RoBERTa，243（53%）", "中间任务的嵌入余弦相似性对其进行排序", "转移组合产生正转移收益，而203（44%）产", "4.3IntermediateModel", "生损失。所有转移增益的平均值为2.3%。然", "而，从我们的11个目标任务中，平均只有5个", "场景中，我们只能访问训练的中间模型（Ms）", "受益（见“平均。表1中的“传输”）。这说明", "时发生的训练数据是专有的，或者如果实现所", "了选择错误的中间任务的高风险。避免这种有", "有数据集太乏味。随着模型存储库（Wolfetal.", "害的组合并有效地确定最佳组合是必要的；评", "2020Pfeifferetal2020a）的可用性，这样的方", "估所有组合是低效的，而且往往是不可行的", "法可以在模型上传期间不需要额外的数据的情", "我们进一步发现，BERT和RoBERTa的最佳", "况下实现（i。e。与TaskEmbs相反，TaskEmbs", "执行中间任务如图E1所示，其中传输性能与", "需要提供训练数据集信息）。下面介绍只需要", "Spearman相关性为0相关。所有设置的平均值", "访问中间模型Ms的方法", "为94，0。每个目标任务平均为68。", "少量微调(FSFT)。在整个目标任务上微调所有", "可用的中间任务模型是不可行的。作为替代方", "6we list the corresponding transfer results for BERT in", "Table ?? of the Appendix.", "'e.g. via https:/huggingface.co/datasets" ]

{ "arxiv_id": "2104.08247", "language": "zh", "page_idx": 3 }

[ [ "0.3506", "0.1256", "0.6504", "0.1256", "0.6504", "0.1401", "0.3506", "0.1401" ], [ "0.4534", "0.1787", "0.5474", "0.1787", "0.5474", "0.1963", "0.4534", "0.1963" ], [ "0.4714", "0.1957", "0.5312", "0.1957", "0.5312", "0.2102", "0.4714", "0.2102" ], [ "0.4167", "0.2102", "0.5825", "0.2102", "0.5825", "0.2235", "0.4167", "0.2235" ], [ "0.2484", "0.2954", "0.3162", "0.2974", "0.3162", "0.3125", "0.2484", "0.31" ], [ "0.5195", "0.2954", "0.9087", "0.2966", "0.9087", "0.3113", "0.5195", "0.31" ], [ "0.523", "0.3113", "0.911", "0.3113", "0.911", "0.3245", "0.523", "0.3245" ], [ "0.10297", "0.318", "0.4631", "0.318", "0.4631", "0.3333", "0.10297", "0.3333" ], [ "0.10297", "0.33", "0.4592", "0.3308", "0.4592", "0.3452", "0.10297", "0.3447" ], [ "0.5215", "0.3257", "0.5537", "0.3257", "0.5537", "0.3384", "0.5215", "0.3384" ], [ "0.10297", "0.3435", "0.4631", "0.3435", "0.4631", "0.3567", "0.10297", "0.3567" ], [ "0.5376", "0.3384", "0.91", "0.3384", "0.91", "0.3528", "0.5376", "0.3528" ], [ "0.1013", "0.3567", "0.4631", "0.3567", "0.4631", "0.3706", "0.1013", "0.3706" ], [ "0.5215", "0.3528", "0.911", "0.3528", "0.911", "0.3662", "0.5215", "0.3662" ], [ "0.10297", "0.3694", "0.4631", "0.3694", "0.4631", "0.3833", "0.10297", "0.3833" ], [ "0.5215", "0.3662", "0.911", "0.3662", "0.911", "0.3801", "0.5215", "0.3801" ], [ "0.1013", "0.3806", "0.4592", "0.3806", "0.4592", "0.3945", "0.1013", "0.3945" ], [ "0.5215", "0.3806", "0.9126", "0.3806", "0.9126", "0.3945", "0.5215", "0.3945" ], [ "0.1013", "0.3933", "0.4592", "0.3933", "0.4592", "0.4072", "0.1013", "0.4072" ], [ "0.5215", "0.3945", "0.9126", "0.3945", "0.9126", "0.4077", "0.5215", "0.4077" ], [ "0.10297", "0.4092", "0.4575", "0.4092", "0.4575", "0.4192", "0.10297", "0.4192" ], [ "0.5215", "0.4072", "0.9126", "0.4077", "0.9126", "0.4224", "0.5215", "0.4216" ], [ "0.10455", "0.4216", "0.46", "0.4216", "0.46", "0.432", "0.10455", "0.432" ], [ "0.5215", "0.4204", "0.911", "0.4216", "0.911", "0.4363", "0.5215", "0.435" ], [ "0.10455", "0.4343", "0.4592", "0.4343", "0.4592", "0.4443", "0.10455", "0.4443" ], [ "0.5195", "0.4343", "0.911", "0.435", "0.911", "0.4495", "0.5195", "0.449" ], [ "0.10455", "0.4463", "0.4575", "0.4463", "0.4575", "0.457", "0.10455", "0.457" ], [ "0.5215", "0.449", "0.911", "0.4495", "0.911", "0.464", "0.5215", "0.4634" ], [ "0.10455", "0.459", "0.4617", "0.459", "0.4617", "0.4697", "0.10455", "0.4697" ], [ "0.1013", "0.4685", "0.46", "0.4697", "0.46", "0.483", "0.1013", "0.4817" ], [ "0.523", "0.4634", "0.911", "0.4634", "0.911", "0.4766", "0.523", "0.4766" ], [ "0.5215", "0.4766", "0.9126", "0.476", "0.9126", "0.4905", "0.5215", "0.4912" ], [ "0.1013", "0.483", "0.4617", "0.4817", "0.4617", "0.4956", "0.1013", "0.4968" ], [ "0.10455", "0.4968", "0.4592", "0.4968", "0.4592", "0.507", "0.10455", "0.507" ], [ "0.5215", "0.4912", "0.9126", "0.4912", "0.9126", "0.505", "0.5215", "0.505" ], [ "0.10455", "0.509", "0.4592", "0.509", "0.4592", "0.5195", "0.10455", "0.5195" ], [ "0.5215", "0.505", "0.9126", "0.505", "0.9126", "0.518", "0.5215", "0.518" ], [ "0.10455", "0.5215", "0.46", "0.5215", "0.46", "0.532", "0.10455", "0.532" ], [ "0.5215", "0.518", "0.9087", "0.518", "0.9087", "0.532", "0.5215", "0.532" ], [ "0.10297", "0.5327", "0.4631", "0.5327", "0.4631", "0.547", "0.10297", "0.547" ], [ "0.524", "0.534", "0.911", "0.534", "0.911", "0.5444", "0.524", "0.5444" ], [ "0.10297", "0.5454", "0.4617", "0.5454", "0.4617", "0.5596", "0.10297", "0.5596" ], [ "0.5215", "0.547", "0.9126", "0.547", "0.9126", "0.56", "0.5215", "0.56" ], [ "0.1013", "0.5566", "0.46", "0.5566", "0.46", "0.571", "0.1013", "0.571" ], [ "0.5215", "0.56", "0.9126", "0.56", "0.9126", "0.5737", "0.5215", "0.5737" ], [ "0.10297", "0.5713", "0.4592", "0.5713", "0.4592", "0.582", "0.10297", "0.582" ], [ "0.5215", "0.5747", "0.9126", "0.5747", "0.9126", "0.5884", "0.5215", "0.5884" ], [ "0.10297", "0.584", "0.4617", "0.584", "0.4617", "0.5977", "0.10297", "0.5977" ], [ "0.523", "0.5884", "0.9126", "0.5884", "0.9126", "0.6016", "0.523", "0.6016" ], [ "0.10297", "0.5967", "0.4575", "0.5967", "0.4575", "0.607", "0.10297", "0.607" ], [ "0.523", "0.6016", "0.9126", "0.6016", "0.9126", "0.6157", "0.523", "0.6157" ], [ "0.10455", "0.61", "0.46", "0.61", "0.46", "0.6206", "0.10455", "0.6206" ], [ "0.5215", "0.6157", "0.911", "0.6157", "0.911", "0.629", "0.5215", "0.629" ], [ "0.10455", "0.6226", "0.4617", "0.6226", "0.4617", "0.6333", "0.10455", "0.6333" ], [ "0.1013", "0.6333", "0.46", "0.6333", "0.46", "0.6465", "0.1013", "0.6465" ], [ "0.523", "0.63", "0.8", "0.63", "0.8", "0.643", "0.523", "0.643" ], [ "0.10297", "0.6465", "0.4617", "0.6465", "0.4617", "0.66", "0.10297", "0.66" ], [ "0.5415", "0.6426", "0.911", "0.6426", "0.911", "0.656", "0.5415", "0.656" ], [ "0.5254", "0.657", "0.9126", "0.657", "0.9126", "0.6704", "0.5254", "0.6704" ], [ "0.5215", "0.6694", "0.911", "0.6704", "0.911", "0.685", "0.5215", "0.6846" ], [ "0.2115", "0.6846", "0.2467", "0.6846", "0.2467", "0.696", "0.2115", "0.696" ], [ "0.2402", "0.6846", "0.353", "0.6846", "0.353", "0.6978", "0.2402", "0.6978" ], [ "0.523", "0.6846", "0.9126", "0.6846", "0.9126", "0.6978", "0.523", "0.6978" ], [ "0.08826", "0.7026", "0.478", "0.7026", "0.478", "0.7163", "0.08826", "0.7163" ], [ "0.5195", "0.6987", "0.569", "0.6963", "0.5703", "0.711", "0.5215", "0.7134" ], [ "0.08826", "0.7163", "0.478", "0.7163", "0.478", "0.73", "0.08826", "0.73" ], [ "0.535", "0.712", "0.9126", "0.711", "0.9126", "0.726", "0.535", "0.7266" ], [ "0.0866", "0.73", "0.478", "0.729", "0.478", "0.742", "0.0866", "0.7437" ], [ "0.5215", "0.7266", "0.91", "0.7266", "0.91", "0.7407", "0.5215", "0.7407" ], [ "0.523", "0.742", "0.614", "0.742", "0.614", "0.7524", "0.523", "0.7524" ], [ "0.6274", "0.7407", "0.91", "0.7407", "0.91", "0.754", "0.6274", "0.754" ], [ "0.08826", "0.756", "0.478", "0.757", "0.478", "0.7715", "0.08826", "0.771" ], [ "0.08905", "0.7437", "0.4763", "0.7437", "0.4763", "0.757", "0.08905", "0.757" ], [ "0.523", "0.756", "0.911", "0.756", "0.911", "0.7666", "0.523", "0.7666" ], [ "0.08826", "0.7715", "0.4763", "0.7715", "0.4763", "0.785", "0.08826", "0.785" ], [ "0.5195", "0.7666", "0.9087", "0.7676", "0.9087", "0.782", "0.5195", "0.781" ], [ "0.08826", "0.785", "0.478", "0.785", "0.478", "0.799", "0.08826", "0.799" ], [ "0.523", "0.782", "0.911", "0.782", "0.911", "0.7954", "0.523", "0.7954" ], [ "0.0866", "0.8", "0.478", "0.8", "0.478", "0.813", "0.0866", "0.813" ], [ "0.5215", "0.797", "0.9126", "0.797", "0.9126", "0.81", "0.5215", "0.81" ], [ "0.08826", "0.813", "0.478", "0.813", "0.478", "0.827", "0.08826", "0.827" ], [ "0.5215", "0.81", "0.911", "0.81", "0.911", "0.8237", "0.5215", "0.8237" ], [ "0.0866", "0.827", "0.478", "0.827", "0.478", "0.8403", "0.0866", "0.8403" ], [ "0.5254", "0.8247", "0.911", "0.8247", "0.911", "0.8384", "0.5254", "0.8384" ], [ "0.0866", "0.8403", "0.478", "0.8403", "0.478", "0.855", "0.0866", "0.855" ], [ "0.5215", "0.8384", "0.911", "0.8384", "0.911", "0.8516", "0.5215", "0.8516" ], [ "0.0866", "0.863", "0.4756", "0.863", "0.4756", "0.877", "0.0866", "0.877" ], [ "0.5435", "0.8623", "0.83", "0.8623", "0.83", "0.8755", "0.5435", "0.8755" ], [ "0.08496", "0.875", "0.4478", "0.8755", "0.4478", "0.89", "0.08496", "0.8896" ], [ "0.54", "0.8755", "0.9062", "0.8755", "0.9062", "0.8896", "0.54", "0.8896" ] ]

[ "用于在线学习的投影二次回归", "Wenye Ma", "Tencent", "[email protected]", "Abstract", "的理论分析都是基于OCO的凸需求。凸度是两个褶皱", "首先，损失函数是凸的。第二，学习变量的可行集也是", "This paper considers online convex optimization", "(OCO) problems - the paramount framework for on-", "凸的", "line learning algorithm design: The loss function of", "同时，在实际应用中，特征尺寸可能非常高。因此", "learning task in OCO setting is based on streaming", "在这种情况下，传统的线性模型仍然是最受欢迎的模", "data so that OCO is a powerful tool to model large", "型。然而，线性模型无法利用特征之间的相互作用，因", "scale applications such as online recommender sys-", "此需要大量的人工特征工程工作。二次回归捕获每个", "tems. Meanwhile, real-world data are usually of ex", "输入特征的一阶信息以及二阶成对特征交互。但是，在", "treme high-dimensional due to modern feature engi", "处理高维数据时，就空间复杂性而言，通常是不可接", "neering techniques so that the quadratic regressionis", "受的，因为模型大小是具有d-dim特征的O(d²）阶。", "impractical.Factorization Machine as well as its vari-", "因子分解机（FM）（Rendle2010）是一种有效的机制", "ants areefficient models for capturing feature interac", "用于以因子分解形式捕获具有低秩矩阵的二阶特征信", "tions with low-rank matrix model but they can't fulfill", "the OCO setting due to their non-convexity. In this", "息。它只需要O(md）空间与秩m。FM在各种应用", "(Juan, Lefortier, and Chapelle 2017; Li et al. 2016; Lu", "paper, We propose a projective quadratic regression", "(PQR)model.First,itcan capturetheimport second-", "et al. 2017; Nguyen, Karatzoglou, and Baltrunas 2014;", "order featureinformation. Second, it is a convex model", "Rendle et al. 2011; Zhong et al. 2016; Chen et al. 2019b;", "sotherequirementsofOCOarefulfilledandtheglobal", "2019a）中实现了最先进的性能，并且最近重新获得了", "optimal solution can be achieved. Moreover, existing", "研究人员（(Blondeletal.2016a:2016b:Heetal.2017:", "modern online optimization methods such as Online", "Xuetal.2016）的关注。不幸的是，香草FM（Rendle", "Gradient Descent (OGD) or Follow-The-Regularized-", "2010)及其变体(Blondel et al.2016a;Cheng etal.2014;", "Leader (FTRL)canbeapplied directly.In addition,by", "Juan, Lefortier, and Chapelle 2017; Lu et al. 2017) 是", "choosing a proper hyper-parameter, we show that it has", "非凸配方，无法满足OCO要求的基本要求。提出了几", "the same order of spaceand timecomplexityasthelin", "种凸调频公式（Blondel，Fujino，andUeda2015；Linet", "ear model and thus can handle high-dimensional data.", "al.2018；Yamadaetal.2017）来解决这一弱点，但计算", "Experimental results demonstrate the performance of", "the proposed PQR model in terms of accuracy and effi-", "成本很高，因此在高维问题中不切实际", "ciency by comparing with the state-of-the-art methods.", "上述公式均基于特征交互矩阵的低秩假设，从而减少", "了学习变量的数量。但是，问题是我们是否真的需要低", "秩假设？还有其他保存参数的方法吗？我们的答案是我", "1", "Introduction", "们不需要低秩假设，我们可以通过共享参数的直觉来保", "在在线学习的设置中，训练数据以流的方式到达，系", "存参数", "统应立即提供响应。具体来说，在线学习是在以时间t", "这种直觉基于我们的观察，即现实世界数据", "为索引的一系列连续轮次中执行的。在t回合中，学", "集中发生的不同特征的频率差异很大。Forin-", "stance,figure", "1 showsthenumbersofoccurrencesofdif-", "变量。在线学习是一个强大的工具，适用于广泛的应", "习者使用通常仅基于单个样本的简单计算来更新学习", "ferentfeaturesintwopublicdataset,i。e。MovieLens1和", "用，如在线评级，新闻馈送和广告点击预测。在线学", "Avazu2。结果表明，数据集中几乎没有出现大量特征", "习的最重要框架是在线凸优化（OCO）（Hazan2016;", "而只有少数特征经常出现。因此，我们将功能分为两", "Shalev-Shwartz2012），可以通过许多最先进的在线优", "类：高频繁功能和低频繁功能。我们假设高频特征更重", "化算法来解决，例如在线梯度校正（OGD）（Zinkevich", "要，因此它们的交互更有价值。在此背景下，我们提出", "2003）或遵循正则化领导者（FTRL）（McMahanetal.", "了一个投影二次回归（PQR）模型。具体来说，对于高", "2013）。值得注意的是，对OCO的研究和对上述算法", "频特征，我们采用全二次形式。在低频率特征中，我们", "Copyright 2020, Association for the Advancement of Ar", "https://grouplens.org/datasets/movielens/", "tificial Intelligence (www.aaai.org). All rights reserved.", "?https://www.kaggle.com/c/avazu-ctr-prediction/data" ]

{ "arxiv_id": "1911.10658", "language": "zh", "page_idx": 1 }

[ [ "0.576", "0.089", "0.6177", "0.089", "0.6177", "0.1042", "0.576", "0.1042" ], [ "0.156", "0.09534", "0.1978", "0.09534", "0.1978", "0.1067", "0.156", "0.1067" ], [ "0.156", "0.125", "0.1978", "0.125", "0.1978", "0.1364", "0.156", "0.1364" ], [ "0.576", "0.1206", "0.615", "0.1206", "0.615", "0.132", "0.576", "0.132" ], [ "0.1438", "0.1477", "0.2141", "0.1843", "0.1814", "0.2216", "0.1111", "0.1849" ], [ "0.5537", "0.1477", "0.589", "0.1477", "0.589", "0.1963", "0.5537", "0.1963" ], [ "0.581", "0.1509", "0.615", "0.1509", "0.615", "0.1622", "0.581", "0.1622" ], [ "0.581", "0.1799", "0.615", "0.1799", "0.615", "0.1913", "0.581", "0.1913" ], [ "0.1422", "0.1528", "0.1978", "0.1528", "0.1978", "0.1642", "0.1422", "0.1642" ], [ "0.1422", "0.1901", "0.1544", "0.1932", "0.1503", "0.2026", "0.1381", "0.1995" ], [ "0.5645", "0.1901", "0.574", "0.1963", "0.5664", "0.2026", "0.558", "0.1957" ], [ "0.1569", "0.2122", "0.1978", "0.2122", "0.1978", "0.2235", "0.1569", "0.2235" ], [ "0.584", "0.2102", "0.6177", "0.2102", "0.6177", "0.2216", "0.584", "0.2216" ], [ "0.1642", "0.2405", "0.1978", "0.2405", "0.1978", "0.2524", "0.1642", "0.2524" ], [ "0.5825", "0.2393", "0.6177", "0.2393", "0.6177", "0.252", "0.5825", "0.252" ], [ "0.609", "0.2715", "0.615", "0.2715", "0.615", "0.2778", "0.609", "0.2778" ], [ "0.286", "0.281", "0.3057", "0.2886", "0.2942", "0.3062", "0.2754", "0.2998" ], [ "0.4478", "0.2822", "0.4592", "0.2874", "0.451", "0.2974", "0.4395", "0.2922" ], [ "0.6406", "0.2798", "0.6567", "0.2827", "0.652", "0.2998", "0.6357", "0.2966" ], [ "0.701", "0.2842", "0.7197", "0.2842", "0.7197", "0.2966", "0.701", "0.2966" ], [ "0.8013", "0.286", "0.8145", "0.286", "0.8145", "0.2966", "0.8013", "0.2966" ], [ "0.834", "0.286", "0.846", "0.286", "0.846", "0.2954", "0.834", "0.2954" ], [ "0.8677", "0.2842", "0.8774", "0.2917", "0.8677", "0.2974", "0.859", "0.2903" ], [ "0.3506", "0.2886", "0.3594", "0.2886", "0.3594", "0.2922", "0.3506", "0.2922" ], [ "0.2443", "0.2874", "0.2688", "0.2822", "0.273", "0.2954", "0.25", "0.3005" ], [ "0.6665", "0.2886", "0.681", "0.2822", "0.689", "0.2917", "0.6733", "0.2974" ], [ "0.3179", "0.289", "0.33", "0.2874", "0.3308", "0.2917", "0.3186", "0.2935" ], [ "0.2427", "0.289", "0.25", "0.289", "0.25", "0.2942", "0.2427", "0.2942" ], [ "0.2754", "0.2903", "0.282", "0.2903", "0.282", "0.2942", "0.2754", "0.2942" ], [ "0.3245", "0.2922", "0.392", "0.289", "0.3945", "0.3208", "0.327", "0.324" ], [ "0.3782", "0.2917", "0.3865", "0.286", "0.3938", "0.2917", "0.3857", "0.2966" ], [ "0.4084", "0.2922", "0.4224", "0.2827", "0.4314", "0.2903", "0.4182", "0.2998" ], [ "0.7295", "0.2922", "0.7427", "0.2822", "0.7534", "0.2903", "0.74", "0.3005" ], [ "0.763", "0.2935", "0.777", "0.2822", "0.7876", "0.2903", "0.774", "0.3018" ], [ "0.312", "0.2954", "0.327", "0.2954", "0.327", "0.3037", "0.312", "0.3037" ], [ "0.2173", "0.2942", "0.2279", "0.2822", "0.2378", "0.2886", "0.2263", "0.2998" ], [ "0.6978", "0.2954", "0.716", "0.2986", "0.709", "0.3193", "0.6914", "0.3164" ], [ "0.857", "0.2954", "0.871", "0.2986", "0.8594", "0.3225", "0.846", "0.3193" ], [ "0.669", "0.2966", "0.681", "0.2966", "0.681", "0.303", "0.669", "0.303" ], [ "0.2173", "0.2998", "0.2296", "0.2998", "0.2296", "0.3093", "0.2173", "0.3093" ], [ "0.267", "0.2998", "0.3325", "0.2974", "0.3342", "0.3245", "0.2688", "0.327" ], [ "0.4395", "0.2966", "0.451", "0.3018", "0.4395", "0.3132", "0.4307", "0.3093" ], [ "0.7935", "0.2998", "0.808", "0.3037", "0.7983", "0.3257", "0.7837", "0.3213" ], [ "0.6274", "0.3005", "0.6436", "0.2935", "0.653", "0.3037", "0.636", "0.3125" ], [ "0.411", "0.3005", "0.4182", "0.3005", "0.4182", "0.305", "0.411", "0.305" ], [ "0.7656", "0.3005", "0.777", "0.3005", "0.777", "0.3037", "0.7656", "0.3037" ], [ "0.8267", "0.3018", "0.835", "0.2986", "0.8384", "0.3018", "0.83", "0.305" ], [ "0.6626", "0.3037", "0.681", "0.3037", "0.681", "0.3176", "0.6626", "0.3176" ], [ "0.2427", "0.303", "0.281", "0.303", "0.281", "0.318", "0.2427", "0.318" ], [ "0.33", "0.308", "0.3367", "0.308", "0.3367", "0.315", "0.33", "0.315" ], [ "0.3962", "0.308", "0.4182", "0.308", "0.4182", "0.3208", "0.3962", "0.3208" ], [ "0.7236", "0.3037", "0.7417", "0.3037", "0.7417", "0.3176", "0.7236", "0.3176" ], [ "0.2942", "0.3125", "0.304", "0.3125", "0.304", "0.3164", "0.2942", "0.3164" ], [ "0.433", "0.3145", "0.4395", "0.3145", "0.4395", "0.3208", "0.433", "0.3208" ], [ "0.819", "0.3113", "0.834", "0.3113", "0.834", "0.3213", "0.819", "0.3213" ], [ "0.7534", "0.3176", "0.7656", "0.307", "0.7754", "0.3145", "0.763", "0.3245" ], [ "0.2222", "0.3289", "0.4478", "0.3289", "0.4478", "0.339", "0.2222", "0.339" ], [ "0.6665", "0.3289", "0.8423", "0.33", "0.8423", "0.3416", "0.6665", "0.3403" ], [ "0.335", "0.3591", "0.6587", "0.361", "0.6587", "0.375", "0.335", "0.3726" ], [ "0.08905", "0.401", "0.4763", "0.401", "0.4763", "0.414", "0.08905", "0.414" ], [ "0.5186", "0.401", "0.911", "0.401", "0.911", "0.414", "0.5186", "0.414" ], [ "0.08826", "0.414", "0.478", "0.414", "0.478", "0.428", "0.08826", "0.428" ], [ "0.536", "0.414", "0.91", "0.414", "0.91", "0.4287", "0.536", "0.4287" ], [ "0.08905", "0.4287", "0.478", "0.4287", "0.478", "0.4426", "0.08905", "0.4426" ], [ "0.536", "0.4287", "0.9126", "0.428", "0.9126", "0.4414", "0.536", "0.4426" ], [ "0.08826", "0.4426", "0.4763", "0.4426", "0.4763", "0.4558", "0.08826", "0.4558" ], [ "0.536", "0.4426", "0.9126", "0.4426", "0.9126", "0.4558", "0.536", "0.4558" ], [ "0.08905", "0.4558", "0.4763", "0.4558", "0.4763", "0.4697", "0.08905", "0.4697" ], [ "0.5376", "0.457", "0.6313", "0.457", "0.6313", "0.4685", "0.5376", "0.4685" ], [ "0.08826", "0.4697", "0.4763", "0.4697", "0.4763", "0.483", "0.08826", "0.483" ], [ "0.08826", "0.4841", "0.478", "0.4841", "0.478", "0.4976", "0.08826", "0.4976" ], [ "0.664", "0.486", "0.7686", "0.486", "0.7686", "0.4988", "0.664", "0.4988" ], [ "0.08826", "0.4976", "0.4739", "0.4976", "0.4739", "0.511", "0.08826", "0.511" ], [ "0.5215", "0.505", "0.911", "0.505", "0.911", "0.518", "0.5215", "0.518" ], [ "0.08826", "0.511", "0.4763", "0.511", "0.4763", "0.5244", "0.08826", "0.5244" ], [ "0.08826", "0.526", "0.478", "0.526", "0.478", "0.539", "0.08826", "0.539" ], [ "0.523", "0.521", "0.91", "0.521", "0.91", "0.531", "0.523", "0.531" ], [ "0.5215", "0.5327", "0.911", "0.5327", "0.911", "0.547", "0.5215", "0.547" ], [ "0.08826", "0.539", "0.478", "0.539", "0.478", "0.553", "0.08826", "0.553" ], [ "0.523", "0.547", "0.911", "0.547", "0.911", "0.56", "0.523", "0.56" ], [ "0.08826", "0.5537", "0.4795", "0.5537", "0.4795", "0.5674", "0.08826", "0.5674" ], [ "0.08826", "0.5664", "0.3635", "0.5664", "0.3635", "0.58", "0.08826", "0.58" ], [ "0.523", "0.561", "0.911", "0.561", "0.911", "0.5747", "0.523", "0.5747" ], [ "0.1013", "0.58", "0.478", "0.579", "0.478", "0.5933", "0.1013", "0.5947" ], [ "0.523", "0.5747", "0.79", "0.5747", "0.79", "0.5884", "0.523", "0.5884" ], [ "0.08826", "0.5947", "0.4763", "0.5947", "0.4763", "0.608", "0.08826", "0.608" ], [ "0.523", "0.6016", "0.568", "0.6016", "0.568", "0.613", "0.523", "0.613" ], [ "0.08905", "0.608", "0.4763", "0.608", "0.4763", "0.622", "0.08905", "0.622" ], [ "0.562", "0.603", "0.79", "0.603", "0.79", "0.613", "0.562", "0.613" ], [ "0.08826", "0.6226", "0.4468", "0.6226", "0.4468", "0.636", "0.08826", "0.636" ], [ "0.523", "0.6206", "0.911", "0.6206", "0.911", "0.634", "0.523", "0.634" ], [ "0.523", "0.634", "0.911", "0.634", "0.911", "0.648", "0.523", "0.648" ], [ "0.5215", "0.648", "0.911", "0.6484", "0.911", "0.662", "0.5215", "0.661" ], [ "0.09314", "0.654", "0.1847", "0.654", "0.1847", "0.6655", "0.09314", "0.6655" ], [ "0.08496", "0.6694", "0.478", "0.6694", "0.478", "0.683", "0.08496", "0.683" ], [ "0.524", "0.664", "0.91", "0.664", "0.91", "0.675", "0.524", "0.675" ], [ "0.5195", "0.6753", "0.91", "0.675", "0.91", "0.6895", "0.5195", "0.69" ], [ "0.10297", "0.685", "0.4739", "0.685", "0.4739", "0.696", "0.10297", "0.696" ], [ "0.523", "0.6895", "0.911", "0.6895", "0.911", "0.704", "0.523", "0.704" ], [ "0.10297", "0.6978", "0.4756", "0.6978", "0.4756", "0.711", "0.10297", "0.711" ], [ "0.523", "0.7056", "0.6055", "0.7026", "0.6055", "0.714", "0.523", "0.7173" ], [ "0.603", "0.704", "0.9126", "0.704", "0.9126", "0.7173", "0.603", "0.7173" ], [ "0.1013", "0.712", "0.478", "0.7104", "0.478", "0.7236", "0.1013", "0.726" ], [ "0.5195", "0.7173", "0.863", "0.7163", "0.863", "0.731", "0.5195", "0.732" ], [ "0.10455", "0.7266", "0.4453", "0.7266", "0.4453", "0.7373", "0.10455", "0.7373" ], [ "0.5376", "0.733", "0.911", "0.733", "0.911", "0.747", "0.5376", "0.747" ], [ "0.08417", "0.742", "0.4795", "0.742", "0.4795", "0.756", "0.08417", "0.756" ], [ "0.5215", "0.7456", "0.9126", "0.7456", "0.9126", "0.759", "0.5215", "0.759" ], [ "0.10046", "0.756", "0.4763", "0.755", "0.4763", "0.7695", "0.10046", "0.771" ], [ "0.5195", "0.759", "0.732", "0.759", "0.732", "0.773", "0.5195", "0.773" ], [ "0.10046", "0.7715", "0.2361", "0.7695", "0.2361", "0.7827", "0.10046", "0.785" ], [ "0.539", "0.776", "0.5884", "0.776", "0.5884", "0.787", "0.539", "0.787" ], [ "0.08496", "0.7886", "0.4795", "0.7886", "0.4795", "0.803", "0.08496", "0.803" ], [ "0.6587", "0.7925", "0.6787", "0.797", "0.675", "0.81", "0.6543", "0.806" ], [ "0.7812", "0.7954", "0.7983", "0.7954", "0.7983", "0.8096", "0.7812", "0.8096" ], [ "0.10297", "0.8037", "0.4756", "0.8037", "0.4756", "0.8145", "0.10297", "0.8145" ], [ "0.6436", "0.802", "0.863", "0.802", "0.863", "0.8384", "0.6436", "0.8384" ], [ "0.58", "0.811", "0.6587", "0.811", "0.6587", "0.8257", "0.58", "0.8257" ], [ "0.8896", "0.811", "0.9126", "0.811", "0.9126", "0.8237", "0.8896", "0.8237" ], [ "0.773", "0.827", "0.8013", "0.827", "0.8013", "0.8384", "0.773", "0.8384" ], [ "0.10297", "0.8164", "0.312", "0.8164", "0.312", "0.83", "0.10297", "0.83" ], [ "0.08496", "0.834", "0.4795", "0.834", "0.4795", "0.8477", "0.08496", "0.8477" ], [ "0.652", "0.829", "0.6763", "0.8257", "0.6787", "0.837", "0.6543", "0.8403" ], [ "0.1013", "0.8467", "0.478", "0.8477", "0.478", "0.8623", "0.1013", "0.8613" ], [ "0.5186", "0.8477", "0.5537", "0.8447", "0.556", "0.8594", "0.5215", "0.8623" ], [ "0.10297", "0.8623", "0.4795", "0.8623", "0.4795", "0.8755", "0.10297", "0.8755" ], [ "0.5376", "0.8613", "0.9087", "0.8613", "0.9087", "0.875", "0.5376", "0.875" ], [ "0.10297", "0.8755", "0.2279", "0.8755", "0.2279", "0.8896", "0.10297", "0.8896" ], [ "0.5215", "0.8755", "0.911", "0.8755", "0.911", "0.8896", "0.5215", "0.8896" ] ]

[ "12000", "30000", "25000", "10000", "S", "sangea", "8000", "6000", "820000", "#", "&", "10000", "4000", "5000", "2000", "0", "0", "E", "0", "", "E", "E", "", "", "", "20", "0", "", "", "", "oK", "OK", "ok)", "bok)", "5", "50)", "0OT)", "wot]", "7", "H", "", "", "YOOT", "1.", "", "", "A", "0", "5", "", "", "M", "", "N", "", "LOK", "#FeatureOccurrenceinMovieLens-1oM", "#FeatureOccurrenceinAvazu", "Figurel:Distribution of Feature Occurrence", "只是忽略它们的相互作用。然后，我们使用一组共享参", ": We evaluate the performance of the proposed PQR", "数来进行高频和低频功能交互。实际上，我们可以将全", "model on both online rating prediction and online bi", "局偏差，特征权重向量和特征交互矩阵重写为增强的对", "nary classification tasks. Experimental results show", "称矩阵。我们上面提到的共享参数的直觉等价于将可行", "that PQR outperforms state-of-the-art online learn-", "集限制为对称矩阵的子集。我们证明了可行集实际上是", "ing methods.", "凸的，然后满足OCO的第一个要求。然后，我们将损", "失函数重写为关于增广矩阵的凸函数，从而满足OCO", "2Method", "的第二个要求。基于该方案，由此产生的PQR公式可", "在本节中，我们首先介绍在线学习中的预备知识，然后", "以无缝地满足OCO框架的上述要求。此外，由于其凸", "性，大多数优化算法如FTRL可以直接应用于PQR模", "转向针对在线凸问题的拟投影二次回归（PQR）模型的", "开发。PQR模型本质上是一个二次回归模型，在可行", "型，理论分析如后悔界和收敛速度仍然有效。此外，通", "集上有一个约束。这个模型背后的想法是共享参数。我", "过选择适当的超参数，这种矩阵可以用O(d）参数表示，", "使其具有与线性模型相同的空间复杂度", "们证明了PQR的凸性，使其适合在线学习环境，并且", "我们对现实世界的数据集进行了广泛的实验，以评估", "可以直接应用最先进的在线优化算法。", "PQR的经验性能。如在线评分预测和在线二进制分类", "2.1", "任务中的实验结果所示，PQR在准确性和效率方面均", "OnlineConvexOptimization", "优于最先进的在线学习算法，尤其是对于高维问题", "在线学习算法建立在以下假设之上：训练实例按顺序到", "达，而不是在学习任务之前可用。在线学习的最重要框", "架是在线凸优化（OCO）（Hazan2016：Shalev-Shwartz", "Contribution", ". PQR model captures first-order feature information", "2012）。它可以被看作是学习者和对手之间的结构化重", "复博奔。。在每轮tE{1,2,...,T}中，要求学习器从", "as well as second-order pairwise feature interaction", "凸集SCRn生成决策点t。然后，对手用凸损失函", "It is a convex formulation so that it fulfills the re", "数ft：s-", "→R回答学习者的决定，学习者遭受损失", "quirements of online convex optimization setting and", "ft（0t）。具体来说，在线学习设置可以描述如下", "the optimal solution is theoreticallyguaranteed.", "对于索引t=1,2,.·,T，学习者选择学习参数OtE", ". Optimization algorithms for convex models can be", "S，其中S是凸的。然后环境以凸函数f:S→R响", "applied directly to PQR and all the theoretical anal-", "应，并且该轮的输出为ft（0t）", "yses are still valid.", "遗憾由", ": PQR is a general framework and can be applied to", "T", "T", "many other tasks such as batch and stochastic set-", "(0) (0)", "regretT =", "(1)", "t=l", "tings besides online learning.", ":PQR is efficient in terms of space complexity and", "t=1", "computation cost. It has the same order of space", "定义", "and time complexity with linear model with a proper", "学习者的目标是生成一系列决策{0tt=1,2,..·,T", "hyper-parameter.", "，以便在T中相对于事后最佳固定决策的遗憾是次线" ]

{ "arxiv_id": "1911.10658", "language": "zh", "page_idx": 2 }

[ [ "0.1642", "0.11554", "0.8364", "0.11554", "0.8364", "0.1288", "0.1642", "0.1288" ], [ "0.1748", "0.1445", "0.814", "0.1445", "0.814", "0.159", "0.1748", "0.159" ], [ "0.2133", "0.1622", "0.7837", "0.1635", "0.7837", "0.1812", "0.2133", "0.1799" ], [ "0.1447", "0.1812", "0.8545", "0.1818", "0.8545", "0.1963", "0.1447", "0.1957" ], [ "0.1584", "0.1976", "0.839", "0.1989", "0.839", "0.2134", "0.1584", "0.2122" ], [ "0.3137", "0.2153", "0.6846", "0.2166", "0.6846", "0.2311", "0.3137", "0.2299" ], [ "0.1136", "0.2349", "0.886", "0.2349", "0.886", "0.25", "0.1136", "0.25" ], [ "0.2484", "0.2715", "0.326", "0.2715", "0.326", "0.289", "0.2484", "0.289" ], [ "0.5156", "0.2747", "0.91", "0.2747", "0.91", "0.2886", "0.5156", "0.2886" ], [ "0.5156", "0.2886", "0.911", "0.2886", "0.911", "0.3018", "0.5156", "0.3018" ], [ "0.11926", "0.2954", "0.4534", "0.2954", "0.4534", "0.31", "0.11926", "0.31" ], [ "0.5156", "0.3018", "0.91", "0.3018", "0.91", "0.315", "0.5156", "0.315" ], [ "0.1177", "0.3093", "0.451", "0.3093", "0.451", "0.3225", "0.1177", "0.3225" ], [ "0.5146", "0.3164", "0.91", "0.3164", "0.91", "0.33", "0.5146", "0.33" ], [ "0.1217", "0.3257", "0.4495", "0.3257", "0.4495", "0.336", "0.1217", "0.336" ], [ "0.5146", "0.33", "0.899", "0.33", "0.899", "0.3435", "0.5146", "0.3435" ], [ "0.11926", "0.3384", "0.4534", "0.3384", "0.4534", "0.3516", "0.11926", "0.3516" ], [ "0.5146", "0.3435", "0.91", "0.3435", "0.91", "0.3567", "0.5146", "0.3567" ], [ "0.1177", "0.351", "0.451", "0.351", "0.451", "0.3643", "0.1177", "0.3643" ], [ "0.5215", "0.356", "0.911", "0.3567", "0.911", "0.3718", "0.5215", "0.3706" ], [ "0.11926", "0.3643", "0.4495", "0.3643", "0.4495", "0.3774", "0.11926", "0.3774" ], [ "0.513", "0.3706", "0.6855", "0.3706", "0.6855", "0.3838", "0.513", "0.3838" ], [ "0.6953", "0.3718", "0.9126", "0.3718", "0.9126", "0.385", "0.6953", "0.385" ], [ "0.11926", "0.3801", "0.452", "0.3801", "0.452", "0.3933", "0.11926", "0.3933" ], [ "0.5146", "0.385", "0.7417", "0.385", "0.7417", "0.3984", "0.5146", "0.3984" ], [ "0.11926", "0.3933", "0.452", "0.3933", "0.452", "0.4072", "0.11926", "0.4072" ], [ "0.5195", "0.4048", "0.6543", "0.4048", "0.6543", "0.4153", "0.5195", "0.4153" ], [ "0.665", "0.404", "0.9126", "0.404", "0.9126", "0.4172", "0.665", "0.4172" ], [ "0.6562", "0.4092", "0.6685", "0.4092", "0.6685", "0.4124", "0.6562", "0.4124" ], [ "0.1177", "0.406", "0.452", "0.4072", "0.452", "0.4204", "0.1177", "0.4192" ], [ "0.11926", "0.4224", "0.451", "0.4224", "0.451", "0.433", "0.11926", "0.433" ], [ "0.5146", "0.4172", "0.9126", "0.4172", "0.9126", "0.4312", "0.5146", "0.4312" ], [ "0.1209", "0.435", "0.451", "0.435", "0.451", "0.449", "0.1209", "0.449" ], [ "0.513", "0.43", "0.91", "0.4312", "0.91", "0.4458", "0.513", "0.4443" ], [ "0.11926", "0.449", "0.452", "0.449", "0.452", "0.4622", "0.11926", "0.4622" ], [ "0.5156", "0.4458", "0.91", "0.4458", "0.91", "0.459", "0.5156", "0.459" ], [ "0.517", "0.4602", "0.5703", "0.4602", "0.5703", "0.4717", "0.517", "0.4717" ], [ "0.596", "0.4583", "0.91", "0.459", "0.91", "0.473", "0.596", "0.4717" ], [ "0.5146", "0.473", "0.91", "0.473", "0.91", "0.486", "0.5146", "0.486" ], [ "0.11926", "0.4634", "0.452", "0.4634", "0.452", "0.4766", "0.11926", "0.4766" ], [ "0.11926", "0.4766", "0.452", "0.4766", "0.452", "0.4905", "0.11926", "0.4905" ], [ "0.11926", "0.4905", "0.4495", "0.4905", "0.4495", "0.504", "0.11926", "0.504" ], [ "0.513", "0.486", "0.911", "0.4873", "0.911", "0.5005", "0.513", "0.5" ], [ "0.1209", "0.504", "0.452", "0.5024", "0.452", "0.5166", "0.1209", "0.5176" ], [ "0.5156", "0.5005", "0.9126", "0.5005", "0.9126", "0.5146", "0.5156", "0.5146" ], [ "0.1177", "0.518", "0.4534", "0.5166", "0.4534", "0.531", "0.1177", "0.5327" ], [ "0.5156", "0.5146", "0.911", "0.5146", "0.911", "0.528", "0.5156", "0.528" ], [ "0.11926", "0.532", "0.452", "0.532", "0.452", "0.5454", "0.11926", "0.5454" ], [ "0.5146", "0.5264", "0.9126", "0.529", "0.9126", "0.5425", "0.5146", "0.5405" ], [ "0.1177", "0.5444", "0.452", "0.5454", "0.452", "0.5596", "0.1177", "0.558" ], [ "0.5146", "0.5425", "0.7607", "0.5425", "0.7607", "0.556", "0.5146", "0.556" ], [ "0.11926", "0.5596", "0.451", "0.5596", "0.451", "0.5728", "0.11926", "0.5728" ], [ "0.5186", "0.56", "0.9126", "0.56", "0.9126", "0.5737", "0.5186", "0.5737" ], [ "0.1209", "0.5737", "0.4495", "0.5737", "0.4495", "0.584", "0.1209", "0.584" ], [ "0.5156", "0.5747", "0.911", "0.5747", "0.911", "0.5884", "0.5156", "0.5884" ], [ "0.11926", "0.587", "0.452", "0.587", "0.452", "0.601", "0.11926", "0.601" ], [ "0.5156", "0.5884", "0.911", "0.5884", "0.911", "0.6016", "0.5156", "0.6016" ], [ "0.1177", "0.5996", "0.452", "0.601", "0.452", "0.6157", "0.1177", "0.6143" ], [ "0.5146", "0.6016", "0.911", "0.6016", "0.911", "0.6157", "0.5146", "0.6157" ], [ "0.1177", "0.6143", "0.452", "0.613", "0.452", "0.6274", "0.1177", "0.629" ], [ "0.5156", "0.616", "0.911", "0.616", "0.911", "0.63", "0.5156", "0.63" ], [ "0.1177", "0.6274", "0.451", "0.6274", "0.451", "0.6416", "0.1177", "0.6416" ], [ "0.5146", "0.63", "0.9087", "0.63", "0.9087", "0.643", "0.5146", "0.643" ], [ "0.11926", "0.6426", "0.451", "0.6426", "0.451", "0.657", "0.11926", "0.657" ], [ "0.5146", "0.643", "0.911", "0.643", "0.911", "0.657", "0.5146", "0.657" ], [ "0.1209", "0.6577", "0.451", "0.657", "0.451", "0.6685", "0.1209", "0.6694" ], [ "0.5156", "0.6577", "0.911", "0.6577", "0.911", "0.672", "0.5156", "0.672" ], [ "0.1209", "0.672", "0.451", "0.6704", "0.451", "0.6816", "0.1209", "0.683" ], [ "0.513", "0.6704", "0.911", "0.6704", "0.911", "0.6846", "0.513", "0.6846" ], [ "0.11926", "0.6846", "0.4534", "0.6846", "0.4534", "0.6978", "0.11926", "0.6978" ], [ "0.5156", "0.6875", "0.8384", "0.6875", "0.8384", "0.6978", "0.5156", "0.6978" ], [ "0.1209", "0.6978", "0.4478", "0.6978", "0.4478", "0.711", "0.1209", "0.711" ], [ "0.5186", "0.704", "0.9126", "0.704", "0.9126", "0.7173", "0.5186", "0.7173" ], [ "0.1177", "0.7104", "0.452", "0.711", "0.452", "0.7246", "0.1177", "0.7236" ], [ "0.5156", "0.7173", "0.9087", "0.7173", "0.9087", "0.731", "0.5156", "0.731" ], [ "0.1177", "0.7246", "0.4495", "0.7246", "0.4495", "0.738", "0.1177", "0.738" ], [ "0.5146", "0.731", "0.91", "0.731", "0.91", "0.744", "0.5146", "0.744" ], [ "0.11926", "0.7393", "0.452", "0.7393", "0.452", "0.7524", "0.11926", "0.7524" ], [ "0.5146", "0.7456", "0.9126", "0.7456", "0.9126", "0.759", "0.5146", "0.759" ], [ "0.11926", "0.754", "0.452", "0.754", "0.452", "0.7676", "0.11926", "0.7676" ], [ "0.5156", "0.759", "0.91", "0.759", "0.91", "0.773", "0.5156", "0.773" ], [ "0.1177", "0.7676", "0.2942", "0.7676", "0.2942", "0.781", "0.1177", "0.781" ], [ "0.513", "0.773", "0.9087", "0.7734", "0.9087", "0.787", "0.513", "0.786" ], [ "0.5146", "0.787", "0.91", "0.787", "0.91", "0.8003", "0.5146", "0.8003" ], [ "0.0866", "0.8003", "0.1217", "0.8003", "0.1217", "0.8125", "0.0866", "0.8125" ], [ "0.1152", "0.8003", "0.2279", "0.8003", "0.2279", "0.8145", "0.1152", "0.8145" ], [ "0.511", "0.8", "0.911", "0.8003", "0.911", "0.815", "0.511", "0.8145" ], [ "0.08826", "0.821", "0.4844", "0.821", "0.4844", "0.834", "0.08826", "0.834" ], [ "0.5312", "0.8237", "0.8965", "0.8247", "0.8965", "0.839", "0.5312", "0.8384" ], [ "0.08905", "0.834", "0.4844", "0.834", "0.4844", "0.8477", "0.08905", "0.8477" ], [ "0.5146", "0.8384", "0.5986", "0.8384", "0.5986", "0.8496", "0.5146", "0.8496" ], [ "0.08826", "0.8477", "0.486", "0.8477", "0.486", "0.8613", "0.08826", "0.8613" ], [ "0.5312", "0.8486", "0.914", "0.8496", "0.914", "0.8643", "0.5312", "0.863" ], [ "0.08826", "0.8623", "0.4714", "0.8623", "0.4714", "0.8755", "0.08826", "0.8755" ], [ "0.5146", "0.863", "0.911", "0.863", "0.911", "0.877", "0.5146", "0.877" ], [ "0.08826", "0.8755", "0.486", "0.8755", "0.486", "0.8896", "0.08826", "0.8896" ], [ "0.513", "0.8755", "0.7197", "0.8755", "0.7197", "0.8896", "0.513", "0.8896" ] ]

[ "学习和推理印度农村地区贫困估计和生计能力的多方面和纵向数据", "Atharva Kulkarnil, Raya Das?, Ravi S. Srivastava3,Tanmoy Chakraborty", "ILanguage Technologies Institute, Carnegie Mellon University, USA", "2Indian Council for Research on International Economic Relations,New Delhi, India", "3Centrefor Employment Studies,Institute forHuman Development, Delhi, India", "4Indian Institute of Technology Delhi,India", "[email protected], [email protected], [email protected], [email protected]", "Abstract", "手段，例如食物，水，住所和金钱[A.M，2020]。虽然贫", "困一直是人类的长期社会经济问题，但在过去的二十年", "Poverty is a multifaceted phenomenon linked", "中，全球贫困状况持续下降[Bank，2020]。然而，由于", "to the lack of capabilities of households to earn", "COVID-19大流行，气候变化和社会经济冲突的复合影", "asustainablelivelihood,increasinglybeingas", "响，消除贫困的努力在一代人中首次遭受重大挫折。", "sessed using multidimensional indicators. Its", "流行导致几乎增加了???万人陷人极端贫困[Bank，2020]", "spatial pattern depends on social, economic.", "此外，随着联合国（UN）将消除贫困列为其主要可持", "political, and regional variables. Artificial in", "续发展目标（SDG）之", "，全球社区正在努力开发有效", "telligence has shown immense scope in analyz-", "的技术来跟踪，估计和消除贫困", "ing the complexities and nuances of poverty.", "Indiaunder crisis:", "根据2015-2016年1农业普查，约", "", "The proposed project aims to examine the", "povertysituation of rural Indiafor theperiod of", "有68个%人口居住在印度农村。最新的多维贫困指数", "1990-2022 based on the quality of life and liveli-", "（MPI）得分表明，农村人口的贫困得分为32.75%，与印", "hood indicators. The districts will be classified", "度城市的8.8%相反。BIMARU2州仍然是该国最贫困", "的地区，", "些统计数据与撒哈拉以南非洲国家相当，因", "此，印度政策的包容性受到质疑。全国大约有50%人口", "into ‘advanced', ‘catching up', ‘falling behind'", "and lagged' regions. The project proposes to", "integrate multiple data sources, including con", "从事农业和相关部门；因此，农业的区域发展对农村家", "ventional national-level large sample household", "庭的生活质量产生了巨大影响。印度社会有一个等级制", "surveys, census surveys, and proxy variables", "度，不同社会群体的流动性不同。根据NFHS-4数据，两", "like daytime, and nighttime data from satel-", "个少数群体，即在册种姓（SC）和在册部落（ST）家庭，", "lite images, and communication networks, to", "比一般社会群体的贫困发生率更高", "name a few, to provide a comprehensive view", "Povertyestimationiscrucial：对于任何国家，准确地", "ofpovertyatthedistrictlevel,Theproiectalso", "衡量贫困统计数据和人口的经济特征都会严重影响其研", "intends to examine causation and longitudinal", "究和国家政策[SkareandDruzeta，2016]。此外，经济", "analysis to examine the reasons for poverty.", "增长不能成为一个国家经济政策的唯一目标；确保经济", "Poverty and inequality could be widening in", "繁荣的好处惠及社会各阶层同样至关重要。这强调了评", "developing countries due to demographic and", "估所有表现形式的贫困的重要性。此外，贫困衡量对于", "growth-agglomerating policies. Therefore, tar-", "评估一个经济体在为所有公民提供一定的最低生活水平", "getingthelaggingregionsand thevulnerable", "方面的表现至关重要。总之，衡量贫困对政策起草和执", "populationis essential to eradicatepoverty and", "行有重大影响。因此，贫困估计和跟踪引起了经济学家、", "improve the quality of life to achieve the goal of", "社会科学家和统计学家的关注也就不足为奇了", "‘zero poverty. Thus, the study also focuses or", "Multifacetedmeasurementofpoverty：然而，贫困不", "the districts with a higher share of the marginal", "仅基于群众中财富的货币分配，而且是一个综合了社会", "section of the population compared to the na", "金融和政治成分的多方面思想。因此，为了解开贫困的结", "tional average to trace the performance of de-", "构并理解其细微差别，需要对其不同方面进行更深入和", "velopment indicators and their association with", "多层面的研究。这样的multidimensionalmeasurement", "poverty in these regions.", "ofpoverty包含两种方法-贫困作为能力剥夺，贫困作为", "剥夺的度量[Atkinson，2003]。由牛津贫困与人类发展倡", "1", "Introduction", "议（OPHI）和联合国开发计划署共同开发的多维贫困指", "贫困是一种复杂的情况，在这种情况下，缺乏能力会转", "*https://agcensus.nic.in/document/agcen1516/T1_ac", "化为家庭的低收人[NussbaumandSen，1993]。从货币", "201516.pdf", "的角度来看，贫困可以描述为低于阈值的收人分配的交", "2BIMARU is an acronym formed from the first letters of", "织，以及该边界[Balaji，2020]内存在的不平衡。这是", "the names of the Indian states of Bihar, Madhya Pradesh,", "种贫困状态，个人缺乏维持其日常生计所需的基本基本", "Rajasthan, and Uttar Pradesh." ]

{ "arxiv_id": "2304.13958", "language": "zh", "page_idx": 1 }

[ [ "0.08826", "0.0707", "0.4885", "0.0707", "0.4885", "0.084", "0.08826", "0.084" ], [ "0.513", "0.0707", "0.911", "0.0707", "0.911", "0.084", "0.513", "0.084" ], [ "0.08826", "0.0808", "0.125", "0.084", "0.1217", "0.0985", "0.0866", "0.0966" ], [ "0.5146", "0.0821", "0.91", "0.0833", "0.91", "0.09784", "0.5146", "0.0966" ], [ "0.5146", "0.09784", "0.911", "0.09784", "0.911", "0.1111", "0.5146", "0.1111" ], [ "0.08496", "0.1067", "0.4478", "0.108", "0.4478", "0.1225", "0.08496", "0.1218" ], [ "0.513", "0.11237", "0.839", "0.11237", "0.839", "0.1256", "0.513", "0.1256" ], [ "0.08826", "0.127", "0.4844", "0.127", "0.4844", "0.1401", "0.08826", "0.1401" ], [ "0.08826", "0.1401", "0.4844", "0.1401", "0.4844", "0.154", "0.08826", "0.154" ], [ "0.511", "0.1364", "0.7935", "0.1383", "0.7935", "0.1528", "0.511", "0.1509" ], [ "0.08826", "0.1547", "0.4844", "0.1547", "0.4844", "0.1686", "0.08826", "0.1686" ], [ "0.5146", "0.1572", "0.9087", "0.1578", "0.9087", "0.1692", "0.5146", "0.1686" ], [ "0.08826", "0.1686", "0.4844", "0.1686", "0.4844", "0.1818", "0.08826", "0.1818" ], [ "0.5146", "0.1704", "0.911", "0.1704", "0.911", "0.1843", "0.5146", "0.1843" ], [ "0.0866", "0.1818", "0.4844", "0.1812", "0.4844", "0.1945", "0.0866", "0.1957" ], [ "0.5156", "0.1843", "0.9087", "0.1843", "0.9087", "0.1976", "0.5156", "0.1976" ], [ "0.0989", "0.1963", "0.4844", "0.1963", "0.4844", "0.2102", "0.0989", "0.2102" ], [ "0.5146", "0.1989", "0.911", "0.1989", "0.911", "0.2122", "0.5146", "0.2122" ], [ "0.08826", "0.2102", "0.4844", "0.209", "0.4844", "0.2229", "0.08826", "0.2235" ], [ "0.517", "0.214", "0.9087", "0.214", "0.9087", "0.2247", "0.517", "0.2247" ], [ "0.08826", "0.2235", "0.4844", "0.2235", "0.4844", "0.2374", "0.08826", "0.2374" ], [ "0.5156", "0.2279", "0.9087", "0.2279", "0.9087", "0.238", "0.5156", "0.238" ], [ "0.08826", "0.2374", "0.4844", "0.2374", "0.4844", "0.2507", "0.08826", "0.2507" ], [ "0.513", "0.2405", "0.91", "0.2405", "0.91", "0.254", "0.513", "0.254" ], [ "0.08826", "0.2507", "0.486", "0.2507", "0.486", "0.2644", "0.08826", "0.2644" ], [ "0.5146", "0.2551", "0.91", "0.2551", "0.91", "0.2683", "0.5146", "0.2683" ], [ "0.08905", "0.2651", "0.486", "0.2651", "0.486", "0.279", "0.08905", "0.279" ], [ "0.5146", "0.2676", "0.91", "0.2676", "0.91", "0.281", "0.5146", "0.281" ], [ "0.08905", "0.279", "0.4844", "0.279", "0.4844", "0.2922", "0.08905", "0.2922" ], [ "0.513", "0.2822", "0.9087", "0.2822", "0.9087", "0.2954", "0.513", "0.2954" ], [ "0.08905", "0.2922", "0.4836", "0.2922", "0.4836", "0.3062", "0.08905", "0.3062" ], [ "0.5146", "0.2954", "0.911", "0.2954", "0.911", "0.3093", "0.5146", "0.3093" ], [ "0.0907", "0.307", "0.486", "0.307", "0.486", "0.3208", "0.0907", "0.3208" ], [ "0.5146", "0.3093", "0.6772", "0.3093", "0.6772", "0.3225", "0.5146", "0.3225" ], [ "0.08905", "0.3208", "0.4844", "0.3208", "0.4844", "0.334", "0.08905", "0.334" ], [ "0.08905", "0.334", "0.468", "0.334", "0.468", "0.348", "0.08905", "0.348" ], [ "0.5146", "0.3435", "0.527", "0.3435", "0.527", "0.3516", "0.5146", "0.3516" ], [ "0.524", "0.3435", "0.55", "0.3435", "0.55", "0.351", "0.524", "0.351" ], [ "0.544", "0.3384", "0.6567", "0.3416", "0.6567", "0.356", "0.544", "0.3528" ], [ "0.0907", "0.348", "0.4844", "0.348", "0.4844", "0.361", "0.0907", "0.361" ], [ "0.08905", "0.361", "0.4844", "0.361", "0.4844", "0.375", "0.08905", "0.375" ], [ "0.511", "0.3599", "0.911", "0.361", "0.911", "0.3757", "0.511", "0.375" ], [ "0.08905", "0.3757", "0.486", "0.3757", "0.486", "0.3894", "0.08905", "0.3894" ], [ "0.5146", "0.375", "0.911", "0.375", "0.911", "0.3882", "0.5146", "0.3882" ], [ "0.0866", "0.387", "0.486", "0.3882", "0.486", "0.4028", "0.0866", "0.4016" ], [ "0.513", "0.3882", "0.91", "0.3882", "0.91", "0.4016", "0.513", "0.4016" ], [ "0.08826", "0.4028", "0.4844", "0.4028", "0.4844", "0.4167", "0.08826", "0.4167" ], [ "0.513", "0.4016", "0.911", "0.404", "0.911", "0.4172", "0.513", "0.4153" ], [ "0.08826", "0.4172", "0.4844", "0.4172", "0.4844", "0.4312", "0.08826", "0.4312" ], [ "0.5146", "0.4167", "0.9126", "0.4167", "0.9126", "0.4312", "0.5146", "0.4312" ], [ "0.08826", "0.4312", "0.486", "0.4312", "0.486", "0.4443", "0.08826", "0.4443" ], [ "0.5146", "0.4312", "0.911", "0.4312", "0.911", "0.4443", "0.5146", "0.4443" ], [ "0.0866", "0.4458", "0.486", "0.4458", "0.486", "0.459", "0.0866", "0.459" ], [ "0.513", "0.443", "0.9126", "0.4443", "0.9126", "0.459", "0.513", "0.4583" ], [ "0.0866", "0.459", "0.4182", "0.459", "0.4182", "0.473", "0.0866", "0.473" ], [ "0.513", "0.459", "0.847", "0.459", "0.847", "0.473", "0.513", "0.473" ], [ "0.0866", "0.483", "0.4478", "0.483", "0.4478", "0.4968", "0.0866", "0.4968" ], [ "0.511", "0.483", "0.888", "0.4849", "0.888", "0.5", "0.511", "0.4976" ], [ "0.08826", "0.5005", "0.4805", "0.502", "0.4805", "0.515", "0.08826", "0.5146" ], [ "0.517", "0.506", "0.91", "0.506", "0.91", "0.5166", "0.517", "0.5166" ], [ "0.08826", "0.515", "0.4836", "0.515", "0.4836", "0.529", "0.08826", "0.529" ], [ "0.5156", "0.5176", "0.911", "0.5176", "0.911", "0.531", "0.5156", "0.531" ], [ "0.09235", "0.53", "0.4844", "0.53", "0.4844", "0.5435", "0.09235", "0.5435" ], [ "0.5146", "0.532", "0.9126", "0.532", "0.9126", "0.5454", "0.5146", "0.5454" ], [ "0.08826", "0.5425", "0.486", "0.5415", "0.486", "0.556", "0.08826", "0.5566" ], [ "0.513", "0.5444", "0.9126", "0.5454", "0.9126", "0.56", "0.513", "0.5596" ], [ "0.08905", "0.5566", "0.486", "0.5566", "0.486", "0.571", "0.08905", "0.571" ], [ "0.5146", "0.56", "0.694", "0.56", "0.694", "0.5737", "0.5146", "0.5737" ], [ "0.08826", "0.5713", "0.4844", "0.5713", "0.4844", "0.5854", "0.08826", "0.5854" ], [ "0.08905", "0.5854", "0.4844", "0.5854", "0.4844", "0.5986", "0.08905", "0.5986" ], [ "0.517", "0.582", "0.6313", "0.5854", "0.6313", "0.5967", "0.517", "0.5933" ], [ "0.08826", "0.5986", "0.486", "0.5986", "0.486", "0.6123", "0.08826", "0.6123" ], [ "0.5146", "0.5986", "0.91", "0.5986", "0.91", "0.6123", "0.5146", "0.6123" ], [ "0.08826", "0.613", "0.486", "0.613", "0.486", "0.627", "0.08826", "0.627" ], [ "0.513", "0.6123", "0.9087", "0.6113", "0.9087", "0.6255", "0.513", "0.627" ], [ "0.08826", "0.6255", "0.4844", "0.6255", "0.4844", "0.6396", "0.08826", "0.6396" ], [ "0.5146", "0.6255", "0.911", "0.6255", "0.911", "0.6396", "0.5146", "0.6396" ], [ "0.09235", "0.6416", "0.486", "0.6416", "0.486", "0.6514", "0.09235", "0.6514" ], [ "0.5146", "0.64", "0.911", "0.64", "0.911", "0.654", "0.5146", "0.654" ], [ "0.08826", "0.654", "0.486", "0.654", "0.486", "0.6675", "0.08826", "0.6675" ], [ "0.5146", "0.654", "0.911", "0.654", "0.911", "0.6675", "0.5146", "0.6675" ], [ "0.08826", "0.6675", "0.4844", "0.6675", "0.4844", "0.681", "0.08826", "0.681" ], [ "0.513", "0.6675", "0.9126", "0.6675", "0.9126", "0.681", "0.513", "0.681" ], [ "0.08826", "0.6816", "0.486", "0.6816", "0.486", "0.696", "0.08826", "0.696" ], [ "0.511", "0.681", "0.91", "0.681", "0.91", "0.6943", "0.511", "0.6943" ], [ "0.0866", "0.696", "0.4878", "0.696", "0.4878", "0.709", "0.0866", "0.709" ], [ "0.5146", "0.696", "0.911", "0.696", "0.911", "0.709", "0.5146", "0.709" ], [ "0.08826", "0.709", "0.4878", "0.709", "0.4878", "0.7227", "0.08826", "0.7227" ], [ "0.5146", "0.709", "0.911", "0.709", "0.911", "0.7227", "0.5146", "0.7227" ], [ "0.08905", "0.7236", "0.486", "0.7236", "0.486", "0.7373", "0.08905", "0.7373" ], [ "0.5107", "0.7217", "0.9087", "0.7227", "0.9087", "0.7373", "0.5107", "0.7363" ], [ "0.08826", "0.7363", "0.486", "0.735", "0.486", "0.75", "0.08826", "0.7505" ], [ "0.513", "0.7373", "0.911", "0.7373", "0.911", "0.7505", "0.513", "0.7505" ], [ "0.08826", "0.7505", "0.4844", "0.7505", "0.4844", "0.7646", "0.08826", "0.7646" ], [ "0.5146", "0.7505", "0.91", "0.7505", "0.91", "0.7646", "0.5146", "0.7646" ], [ "0.08826", "0.7646", "0.486", "0.763", "0.486", "0.778", "0.08826", "0.779" ], [ "0.5156", "0.7666", "0.9087", "0.7666", "0.9087", "0.7764", "0.5156", "0.7764" ], [ "0.08905", "0.779", "0.4844", "0.779", "0.4844", "0.7925", "0.08905", "0.7925" ], [ "0.5156", "0.778", "0.91", "0.778", "0.91", "0.7915", "0.5156", "0.7915" ], [ "0.08905", "0.7925", "0.4878", "0.7925", "0.4878", "0.806", "0.08905", "0.806" ], [ "0.511", "0.7905", "0.91", "0.7915", "0.91", "0.8047", "0.511", "0.8037" ], [ "0.08905", "0.806", "0.4902", "0.806", "0.4902", "0.8193", "0.08905", "0.8193" ], [ "0.5146", "0.806", "0.911", "0.806", "0.911", "0.8193", "0.5146", "0.8193" ], [ "0.08905", "0.8193", "0.486", "0.8193", "0.486", "0.8335", "0.08905", "0.8335" ], [ "0.5146", "0.8193", "0.911", "0.8193", "0.911", "0.8335", "0.5146", "0.8335" ], [ "0.08826", "0.834", "0.4836", "0.834", "0.4836", "0.8477", "0.08826", "0.8477" ], [ "0.513", "0.834", "0.911", "0.834", "0.911", "0.8477", "0.513", "0.8477" ], [ "0.08905", "0.8477", "0.4836", "0.8477", "0.4836", "0.8613", "0.08905", "0.8613" ], [ "0.513", "0.8477", "0.911", "0.8477", "0.911", "0.8613", "0.513", "0.8613" ], [ "0.08826", "0.8623", "0.486", "0.8623", "0.486", "0.8755", "0.08826", "0.8755" ], [ "0.513", "0.8623", "0.91", "0.8623", "0.91", "0.8755", "0.513", "0.8755" ], [ "0.08826", "0.8755", "0.4902", "0.8755", "0.4902", "0.8896", "0.08826", "0.8896" ], [ "0.511", "0.8735", "0.911", "0.875", "0.911", "0.8896", "0.511", "0.888" ] ]

[ "社会经济落后国家贫困的表现和扩大。本节扩展了这些", "此外，我们提供基于多输人深度学习的架构，用于聚合", "目标", "和处理来自多个来源的数据。最后，我们提供了调查各", "种输人变量与贫困估计之间的相关，关联和因果关系的", "2.1Limitations of Traditional Data Sources", "方法，对其进行识别可以帮助更好地制定政策", "来自家庭支出和收人调查的数据，例如国家家庭和健康", "调查（NFHS），不同轮的国家抽样调查（NSS）和印度人", "2.3Temporal Analysis ofPoverty", "类发展调查（IHDS），构成了识别和衡量印度家庭贫困", "贫困是一种随时间加剧或下降的时间现象。因此，为了", "状况的基础。虽然人口普查数据提供了对人口中个人物", "预测一个区域的贫困统计及其进展，还应考虑其表征时", "质生活水平的全面衡量，但它是在长周期（通常每十年", "间变化的历史数据。尽管这种方法很可能会产生更好", "一次）内进行的，只涵盖了目标人口的几个特征。另一方", "更准确和可靠的预测，但它也将有助于区分导致贫困的", "面，家庭调查涵盖了广泛的变量，但其对本地统计推断", "主要因素和较弱的因素。它还将帮助诊断在特定时间内", "的可靠性有限进行此类调查也很昂贵，耗时，不精确，有", "成功导致减贫或产生不利影响的计划，计划和政策，从", "时对于贫困评估也不可行。这些统计数据中的大多数是", "而导致贫困扩大。关于贫困估计的当代文学只关注静态", "在相当长的时间内生成的，并不能准确反映影响特定地", "和特定时间的数据，没有考虑到贫困的时间方面。在印", "区居民生计的属性。此外，所有数据集的目标人群也不", "度等快速发展的国家，最重要的是要考虑到贫困的时间", "相同。例如，阿鲁纳恰尔邦，比哈尔邦，贾坎德邦和奥里", "层面，以查明其核心原因并设计解决贫困的政策。因此", "萨邦的部落人口以及标准的农村和城市分布。这引起了", "本研究提出了一种时间数据收集、整合和预测方案，以", "人们对估计贫困数字可信度的担忧。除少数例外，例如", "实现更稳健的贫困预测", "有关教育和工作状况的信息，此类调查仅关注家庭货币", "信息[AlkireandFoster，2011]。然而，由于贫困不是", "3", "", "Methodology", "个单一的现象，而不仅仅是个人经济状况的反映，它还", "包括其他伴随的方面，如基础设施发展、农业增长、车辆", "在本节中，我们详细介绍了不同的代理数据源，数据集", "数量、网络塔、政治气候、种姓信息、用电量等数据。这", "成方法，使用ML和深度学习技术进行贫困估计以及对", "些因素在农村环境中更为相关，因为它们是繁荣的直接", "贫困的时间分析。此外，我们根据1993-2015年间的表", "标志。如上所述，虽然构成收人统计的家庭数据确实在", "现（MPI得分）对地区进行分类，考虑到几轮NFHS家", "贫困预测中占有一席之地，但它们有一定的局限性。因", "庭调查，分为“先进”，“追赶”，“落后”和“落后”地", "此, we propose a data integration mechanism that takes", "区。该研究进一步针对四个最落后的州-北方邦，贾坎德", "into account both the traditional datasets along with the", "邦，比哈尔邦和奥里萨邦，因为它们在2021年[Tripathi", "new proxy data sources for poverty prediction", "andYenneti,2020]的多维贫困指数中排名最高", "2.2Towards the Use of Proxy Data Sources", "3.1Proxy Data Sources and DataIntegration", "关于多维贫困估算的研究很多，尽管其中大多数除了", "要开发一种功能方法，供发展中国家政府用于准确的贫", "经济方面的[AlkireandFoster，2011]之外，还仅纳人", "困估计和跟踪，需要一个代表该国人口的数据集，可以", "了教育，健康和生活水平的条件。虽然这些战略无疑", "自动收集和及时更新，并且可以在精细的地理粒度。在", "优于单一来源的贫困评估方法，但它们确实有一些不", "本节中，我们将探讨用于贫困估计的不同代理数据源，", "利之处，如上一节所述。为了弥补这些缺陷，大数据", "并详细说明数据集成方法", "的出现，加上机器学习的技术突破，为贫困跟踪以及", "解释和预测社会经济状况提供了巨大的希望。。这说", "Remotesensing", "明了从社交媒体，遥感，农业增长，车辆交通，基础", "利用卫星图像进行的远程测量是一种低成本且可靠的方", "设施发展，手机元数据和住房详细信息收集的数据政", "法，可以以精细的时空分辨率[Jean etal.，2016]跟踪人", "治，文化和环境方面也是估算贫困的因素。有很多例", "类的发展。遥感涉及各种类型的卫星成像，在白天和黑", "子详细说明子如何成功使用这些代理数据源进行跨地", "夜传达不同类型的信息。白天卫星图像提供了有关该地", "域的贫困估计。例如，遥感图像，如陆地卫星数据和", "区地理，基础设施发展和人口增长的丰富信息。此外，它", "夜间光线图像，是最具代表性的数据源，因为它们提", "还可用于推断其他繁荣信号，例如道路网络的增长，建", "供了关于该地区景观的重要信息。关于中国[niu，2020;", "筑密度，森林覆盖率和基础设施的扩张。卫星图像提供", "Shietal.,2020]，泰国[Puttanapongetal.,2020]，菲律", "的夜间光度信息提供了一个区域夜间活动的镜头。它是", "宾[Hoferetal.,2020]，孟加拉国[Steeleetal.,2017]和", "用电量，电气化程度和人口增长的理想代理[Ghosh et", "非洲国家[Jeanetal.2016;Ayushetal.，2021]的各种", "al.，2013]。研究还表明，夜间光度与二氧化碳排放量", "案例研究表明，遥感可用于贫困跟踪和估计。其他一些", "GDP，人均GDP，不变价格GDP，非农业GDP和资", "研究表明，使用来自手机[Blumenstocketal.,2015]，通", "本存量[Addison and Stewart，2015]呈正相关。。因此", "信网络[Smith-Clarkeetal.，2014]和政治气候[Vander", "我们的目标是收集和利用白天和夜间卫星图像数据，因", "Bergetal.，2006]的数据也会产生有希望的结果。虽然", "为每种数据都有其自身的优势。我们建议使用2001年，", "这些研究工作确实验证了使用代理数据源进行贫困估计，", "2011年，2016年和2019年Landsat7任务的卫星图像", "但它们也揭示了没有这样的研究仔细研究印度次大陆。", "来跟踪白天的活动。夜间灯光数据可以从美国空军防御", "此外，这些工作使用独立的代理数据源进行贫困估计，", "气象卫星计划（DMSP）获得。卫星的操作线扫描系统", "并且没有考虑将不同的数据源组合和集成以进行更有效", "（OLS）传感器具有空间分辨率，使它们能够进行从整个", "可解释和可靠的预测的想法。因此，这项工作旨在通过", "大陆到不到一平方公里的观测。为了将村庄和地区映射", "提出一种数据汇总和集成方法来解决这一瓶颈，该方法", "到其卫星图像位置，我们还建议收集每个形状文件的信", "结合了传统和代理数据源，用于印度各州的贫困估算。", "息。一旦一个地区被链接到其卫星图像，我们就可以从" ]

{ "arxiv_id": "2304.13958", "language": "zh", "page_idx": 3 }

[ [ "0.11316", "0.11945", "0.1399", "0.11945", "0.1399", "0.134", "0.11316", "0.134" ], [ "0.1214", "0.1361", "0.2294", "0.1361", "0.2294", "0.148", "0.1214", "0.148" ], [ "0.0977", "0.1465", "0.815", "0.15", "0.815", "0.1708", "0.0977", "0.1674" ], [ "0.0936", "0.1715", "0.7817", "0.1736", "0.7817", "0.1924", "0.0936", "0.1903" ], [ "0.09467", "0.209", "0.856", "0.209", "0.856", "0.2236", "0.09467", "0.2236" ], [ "0.0936", "0.2355", "0.9043", "0.2355", "0.9043", "0.2487", "0.0936", "0.2487" ], [ "0.0936", "0.2507", "0.9023", "0.2507", "0.9023", "0.2646", "0.0936", "0.2646" ], [ "0.09155", "0.2654", "0.9053", "0.2654", "0.9053", "0.2798", "0.09155", "0.2798" ], [ "0.09155", "0.2798", "0.9053", "0.2798", "0.9053", "0.2937", "0.09155", "0.2937" ], [ "0.09155", "0.2966", "0.907", "0.2966", "0.907", "0.3103", "0.09155", "0.3103" ], [ "0.0936", "0.311", "0.907", "0.311", "0.907", "0.325", "0.0936", "0.325" ], [ "0.09155", "0.3271", "0.907", "0.3271", "0.907", "0.3403", "0.09155", "0.3403" ], [ "0.0905", "0.3403", "0.9053", "0.3416", "0.9053", "0.3562", "0.0905", "0.355" ], [ "0.09155", "0.357", "0.907", "0.357", "0.907", "0.3708", "0.09155", "0.3708" ], [ "0.09155", "0.3716", "0.9053", "0.3716", "0.9053", "0.3862", "0.09155", "0.3862" ], [ "0.0936", "0.3862", "0.9053", "0.3862", "0.9053", "0.4", "0.0936", "0.4" ], [ "0.09155", "0.4028", "0.907", "0.4028", "0.907", "0.4167", "0.09155", "0.4167" ], [ "0.09155", "0.4175", "0.9053", "0.4175", "0.9053", "0.432", "0.09155", "0.432" ], [ "0.09155", "0.4333", "0.9053", "0.4333", "0.9053", "0.4465", "0.09155", "0.4465" ], [ "0.09155", "0.4487", "0.9043", "0.4487", "0.9043", "0.4624", "0.09155", "0.4624" ], [ "0.0905", "0.4631", "0.7417", "0.4624", "0.7417", "0.477", "0.0905", "0.4778" ], [ "0.0936", "0.4841", "0.9053", "0.4841", "0.9053", "0.4985", "0.0936", "0.4985" ], [ "0.0936", "0.5005", "0.8496", "0.5005", "0.8496", "0.5146", "0.0936", "0.5146" ], [ "0.09155", "0.5215", "0.907", "0.5215", "0.907", "0.536", "0.09155", "0.536" ], [ "0.0936", "0.5396", "0.1543", "0.5396", "0.1543", "0.551", "0.0936", "0.551" ], [ "0.09155", "0.558", "0.2964", "0.558", "0.2964", "0.5713", "0.09155", "0.5713" ], [ "0.0936", "0.5737", "0.9053", "0.5737", "0.9053", "0.5874", "0.0936", "0.5874" ], [ "0.09155", "0.5894", "0.9053", "0.5894", "0.9053", "0.6025", "0.09155", "0.6025" ], [ "0.0905", "0.602", "0.6978", "0.6025", "0.6978", "0.6206", "0.0905", "0.6196" ], [ "0.1163", "0.6377", "0.1245", "0.6377", "0.1245", "0.6426", "0.1163", "0.6426" ], [ "0.1193", "0.6333", "0.2932", "0.6333", "0.2932", "0.6465", "0.1193", "0.6465" ], [ "0.0936", "0.6543", "0.9053", "0.6543", "0.9053", "0.669", "0.0936", "0.669" ], [ "0.0936", "0.6714", "0.9053", "0.6714", "0.9053", "0.686", "0.0936", "0.686" ], [ "0.09155", "0.684", "0.9053", "0.6855", "0.9053", "0.7026", "0.09155", "0.702" ], [ "0.09467", "0.704", "0.9043", "0.704", "0.9043", "0.7188", "0.09467", "0.7188" ], [ "0.09467", "0.7207", "0.36", "0.7207", "0.36", "0.734", "0.09467", "0.734" ], [ "0.11523", "0.7363", "0.9023", "0.7363", "0.9023", "0.7505", "0.11523", "0.7505" ], [ "0.09467", "0.7544", "0.9043", "0.7544", "0.9043", "0.7686", "0.09467", "0.7686" ], [ "0.0936", "0.7793", "0.748", "0.7793", "0.748", "0.7925", "0.0936", "0.7925" ], [ "0.09155", "0.8027", "0.9053", "0.8027", "0.9053", "0.817", "0.09155", "0.817" ], [ "0.09155", "0.817", "0.907", "0.817", "0.907", "0.83", "0.09155", "0.83" ], [ "0.0936", "0.831", "0.907", "0.831", "0.907", "0.844", "0.0936", "0.844" ], [ "0.0936", "0.8467", "0.9053", "0.8467", "0.9053", "0.857", "0.0936", "0.857" ], [ "0.09155", "0.859", "0.7427", "0.859", "0.7427", "0.873", "0.09155", "0.873" ], [ "0.0936", "0.8735", "0.6514", "0.8735", "0.6514", "0.8877", "0.0936", "0.8877" ], [ "0.09467", "0.89", "0.323", "0.89", "0.323", "0.898", "0.09467", "0.898" ], [ "0.09155", "0.902", "0.3057", "0.902", "0.3057", "0.915", "0.09155", "0.915" ], [ "0.2119", "0.934", "0.9053", "0.934", "0.9053", "0.9478", "0.2119", "0.9478" ] ]

[ "[1]", "ENEXT LINE", "Proactive Moderation of Online Discussions: Existing", "Practices and the Potential for Algorithmic Support", "CHARLOTTE SCHLUGER, CHARLOTTE SCHLUGER, and JONATHAN P.CHANG", "Toaddress the widespread problem of uncivil behavior,manyonline discussionplatforms employhuman", "moderators to take action against objectionable content, such as removing it or placing sanctions on its au", "thors. This reactive paradigm of taking action against already-posted antisocial content is currently the most", "common form of moderation, and has accordingly underpinned many recent efforts at introducing automa", "tion into the moderation process. Comparatively less work has been done to understand other moderation", "paradigms-such as proactively discouraging the emergence of antisocial behavior rather than reacting to", "it-and the role algorithmic support can play in these paradigms. In this work, we investigate such a proac-", "tive framework for moderation in a case study of a collaborative setting: Wikipedia Talk Pages. We employ", "a mixed methods approach, combining qualitative and design components for a holistic analysis.Through", "interviews with moderators, we find that despite a lack of technical and social support, moderators already", "engage in a number of proactive moderation behaviors, such as preemptively intervening in conversations", "to keep them on track. Further, we explore how automation could assist with this existing proactive modera-", "tion workflow by building a prototype tool, presenting it to moderators, and examining how the assistance it", "provides mightfit into their workflow.The resulting feedback uncovers both strengths and drawbacks of the", "prototype tool and suggests concrete steps towards further developing such assisting technology so it can", "most effectively support moderators in their existing proactive moderation workflow.", "CcS Concepts: · Human-centered computing - Interactive systems and tools; Collaborative and social", "computing systems and tools; .Computing methodologies→ Natural language processing.", "Additional Key Words and Phrases: antisocial behavior, content moderation, algorithmic assistance, hybrid", "systems", "ACM Reference Format:", "Charlotte Schluger, Charlotte Schluger,and Jonathan P.Chang.2022.Proactive Moderation of Online Dis", "cussions:ExistingPractices andthePotentialforAlgorithmicSupport.Proc.ACMHum.-Comput.Interact.6,", "CSCW2,Article 370 (November 2022),25pages.https://doi.org/10.1145/3555095", "", "INTRODUCTION", "Online discussion platforms enable new forms of interaction and knowledge sharing and have", "becomekey in supporting online collaboration.However, incivility challenges theirbenefits,损害", "遭受反社会行为或人身攻击的个人的精神和情感健康[1,53]，reducingengagement[14]，or", "distractingfromtheunderlyinggoalsofthediscussion[2]。因此，大多数在线平台都使用节制", "来将讨论保持在其社区准则内", "传统上，公共讨论平台的审核包括一个过程，在该过程中，人类主持人对他们发现的反社", "会行为的实例做出反应-通过手动搜索自已或通过社区成员的报告。这种方法是劳动密集型", "Authors\"address: Charlotte Schluger; Charlotte Schluger, [email protected]; Jonathan P. Chang.", "Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee", "provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and", "the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be", "honored.Abstracting with credit is permitted.To copy otherwise,orrepublish,topost on servers or toredistribute to lists", "requires prior specific permission and/or a fee.Request permissions from [email protected].", "2022 Copyright held by the owner/author(s).Publication rights licensed to ACM.", "2573-0142/2022/11-ART370S15.00", "https://doi.org/10.1145/3555095", "Proc.ACM Hum.-Comput. Interact.,Vol. 6, No.CSCW2, Article 370.Publication date:November 2022." ]

{ "arxiv_id": "2211.16525", "language": "zh", "page_idx": 0 }

[ [ "0.0905", "0.082", "0.1338", "0.082", "0.1338", "0.0979", "0.0905", "0.0979" ], [ "0.737", "0.082", "0.9014", "0.082", "0.9014", "0.0958", "0.737", "0.0958" ], [ "0.0936", "0.11945", "0.2603", "0.11945", "0.2603", "0.1333", "0.0936", "0.1333" ], [ "0.6387", "0.1215", "0.9014", "0.1215", "0.9014", "0.132", "0.6387", "0.132" ], [ "0.0967", "0.1382", "0.1965", "0.1382", "0.1965", "0.1486", "0.0967", "0.1486" ], [ "0.6885", "0.1354", "0.9043", "0.1354", "0.9043", "0.1486", "0.6885", "0.1486" ], [ "0.0967", "0.1542", "0.2026", "0.1542", "0.2026", "0.1653", "0.0967", "0.1653" ], [ "0.6997", "0.1528", "0.9043", "0.1528", "0.9043", "0.1666", "0.6997", "0.1666" ], [ "0.09467", "0.1694", "0.1821", "0.1694", "0.1821", "0.1813", "0.09467", "0.1813" ], [ "0.1183", "0.1875", "0.2068", "0.1875", "0.2068", "0.1979", "0.1183", "0.1979" ], [ "0.7593", "0.1862", "0.9053", "0.1862", "0.9053", "0.2", "0.7593", "0.2" ], [ "0.0967", "0.2042", "0.2089", "0.2042", "0.2089", "0.2153", "0.0967", "0.2153" ], [ "0.7446", "0.2028", "0.9023", "0.2028", "0.9023", "0.2167", "0.7446", "0.2167" ], [ "0.09467", "0.2195", "0.2418", "0.2195", "0.2418", "0.2333", "0.09467", "0.2333" ], [ "0.7964", "0.2216", "0.8994", "0.2216", "0.8994", "0.2319", "0.7964", "0.2319" ], [ "0.09467", "0.2361", "0.2603", "0.2361", "0.2603", "0.25", "0.09467", "0.25" ], [ "0.7603", "0.2375", "0.9014", "0.2375", "0.9014", "0.2479", "0.7603", "0.2479" ], [ "0.0967", "0.2527", "0.2479", "0.2527", "0.2479", "0.2666", "0.0967", "0.2666" ], [ "0.711", "0.2542", "0.9043", "0.2542", "0.9043", "0.2646", "0.711", "0.2646" ], [ "0.0936", "0.2695", "0.3044", "0.2695", "0.3044", "0.2834", "0.0936", "0.2834" ], [ "0.675", "0.2688", "0.9053", "0.2673", "0.9053", "0.282", "0.675", "0.2834" ], [ "0.1091", "0.284", "0.323", "0.2861", "0.323", "0.3", "0.1091", "0.2979" ], [ "0.672", "0.288", "0.75", "0.288", "0.75", "0.296", "0.672", "0.296" ], [ "0.11316", "0.302", "0.36", "0.302", "0.36", "0.3152", "0.11316", "0.3152" ], [ "0.6387", "0.3027", "0.9053", "0.3027", "0.9053", "0.3167", "0.6387", "0.3167" ], [ "0.09155", "0.3174", "0.3982", "0.3188", "0.3982", "0.3333", "0.09155", "0.332" ], [ "0.633", "0.3193", "0.9053", "0.3193", "0.9053", "0.3333", "0.633", "0.3333" ], [ "0.09467", "0.3362", "0.4229", "0.3362", "0.4229", "0.35", "0.09467", "0.35" ], [ "0.607", "0.3362", "0.907", "0.3362", "0.907", "0.3506", "0.607", "0.3506" ], [ "0.0936", "0.3528", "0.36", "0.3528", "0.36", "0.3667", "0.0936", "0.3667" ], [ "0.3684", "0.3584", "0.4053", "0.3584", "0.4053", "0.3625", "0.3684", "0.3625" ], [ "0.501", "0.3542", "0.9043", "0.3542", "0.9043", "0.3645", "0.501", "0.3645" ], [ "0.09467", "0.3694", "0.4731", "0.3694", "0.4731", "0.3833", "0.09467", "0.3833" ], [ "0.4773", "0.3708", "0.9014", "0.3708", "0.9014", "0.3813", "0.4773", "0.3813" ], [ "0.0936", "0.384", "0.9023", "0.3855", "0.9023", "0.4", "0.0936", "0.3987" ], [ "0.09467", "0.4006", "0.9043", "0.402", "0.9043", "0.4167", "0.09467", "0.4153" ], [ "0.09467", "0.4194", "0.9043", "0.4194", "0.9043", "0.434", "0.09467", "0.434" ], [ "0.0905", "0.4333", "0.9053", "0.434", "0.9053", "0.4521", "0.0905", "0.4507" ], [ "0.0936", "0.453", "0.1975", "0.453", "0.1975", "0.4646", "0.0936", "0.4646" ], [ "0.11523", "0.4688", "0.8354", "0.4688", "0.8354", "0.4834", "0.11523", "0.4834" ], [ "0.11316", "0.4883", "0.9053", "0.4883", "0.9053", "0.503", "0.11316", "0.503" ], [ "0.142", "0.505", "0.2325", "0.505", "0.2325", "0.5195", "0.142", "0.5195" ], [ "0.11316", "0.5215", "0.893", "0.5215", "0.893", "0.536", "0.11316", "0.536" ], [ "0.0936", "0.5415", "0.9053", "0.5415", "0.9053", "0.556", "0.0936", "0.556" ], [ "0.0936", "0.559", "0.907", "0.559", "0.907", "0.5737", "0.0936", "0.5737" ], [ "0.09155", "0.575", "0.2808", "0.575", "0.2808", "0.5884", "0.09155", "0.5884" ], [ "0.0936", "0.607", "0.1091", "0.607", "0.1091", "0.615", "0.0936", "0.615" ], [ "0.1163", "0.609", "0.1245", "0.609", "0.1245", "0.6133", "0.1163", "0.6133" ], [ "0.1245", "0.6025", "0.634", "0.604", "0.634", "0.6187", "0.1245", "0.617" ], [ "0.0936", "0.625", "0.9043", "0.625", "0.9043", "0.638", "0.0936", "0.638" ], [ "0.0936", "0.6426", "0.9053", "0.6426", "0.9053", "0.6567", "0.0936", "0.6567" ], [ "0.0936", "0.659", "0.907", "0.659", "0.907", "0.674", "0.0936", "0.674" ], [ "0.09155", "0.676", "0.9053", "0.676", "0.9053", "0.6895", "0.09155", "0.6895" ], [ "0.0936", "0.6924", "0.9053", "0.6924", "0.9053", "0.707", "0.0936", "0.707" ], [ "0.09155", "0.7104", "0.1399", "0.7104", "0.1399", "0.7217", "0.09155", "0.7217" ], [ "0.09155", "0.7373", "0.3662", "0.7373", "0.3662", "0.7505", "0.09155", "0.7505" ], [ "0.09155", "0.759", "0.9053", "0.759", "0.9053", "0.7734", "0.09155", "0.7734" ], [ "0.09155", "0.776", "0.907", "0.776", "0.907", "0.79", "0.09155", "0.79" ], [ "0.09155", "0.7925", "0.9053", "0.7925", "0.9053", "0.806", "0.09155", "0.806" ], [ "0.09155", "0.809", "0.9053", "0.809", "0.9053", "0.8228", "0.09155", "0.8228" ], [ "0.0936", "0.8257", "0.9053", "0.8257", "0.9053", "0.8403", "0.0936", "0.8403" ], [ "0.0936", "0.8423", "0.5225", "0.8423", "0.5225", "0.8564", "0.0936", "0.8564" ], [ "0.1121", "0.8584", "0.9053", "0.8584", "0.9053", "0.873", "0.1121", "0.873" ], [ "0.0936", "0.875", "0.9053", "0.875", "0.9053", "0.8896", "0.0936", "0.8896" ], [ "0.09155", "0.898", "0.32", "0.9", "0.32", "0.9175", "0.09155", "0.915" ], [ "0.09155", "0.934", "0.784", "0.934", "0.784", "0.9478", "0.09155", "0.9478" ] ]

[ "370:2", "Charlotte Schluger et al.", "的，很难扩展到现", "[23，25]来说可能是practically", "andemotion", "自动化内容审核过程的兴", "趣日益浓厚。", "习来自动检测反社会评", "论[51.65]", "量然这些", "关键的限制：整", "个过程基本：", "最好的反应性节制", "也只能在已经造", "力或阻止富", "有成效的讨论。尽", "技术关注，但合", "人惊的是，在", "面所做的工作很少。F", "[35]：出于反应性节制", "主持人需要算法支持的唯", "任务吗？或者，在审核", "开助", "在这项工作中，我们通过研", "替力来开始解决这些问题-主动", "审核-throughacasestudyofacolla", "Talk Pages.Wefocus onaspecific", "proactivemoderationstrategythatprio", "s needing algorithmic support: iden-", "tifying and monitoring conversa", "TS", "-riskofdevolvingintouncivilbehavior,withthe", "goal of intervening earlytoavoidderailment", "ortomitigateitseffectsinatimelymanner[39,59", "我们采访主持人，以了解他们的主动审核实践并确定他们对算法支持的需求。为了探索使用", "当今可用的技术解决这些需求的可行性，我们构建了一个原型工具来帮助主动审核，并观察", "主持人如何在其现有工作流程的背景下参与其中。Inouranalysiswepayparticularattention", "to the ethical issues that come to the fore, and discuss their implications for system design and", "future work.", "具体而言，ourcasestudyaimstoaddressthefollowingprimaryresearchquestions:", "(1) (How) do moderators act proactively to prevent the derailment of discussions into uncivil", "behavior?", "(2) (How) can an algorithmic tool assist moderators in their proactive moderation workflow?", "Building on the answers to these questions, we identify concrete next steps towards further de", "veloping such assisting technology so that it can support moderators in their existing proactive", "moderation workflow.", "2", "", "BACKGROUND:THELANDSCAPEOFMODERATION", "Our work contributes to an extensive line of research on online discussion moderation which has", "aimed to both understand existing practices and develop new ones. Thus, to motivate our focus", "on the proactive paradigm and put our contributions in context, we begin by briefly surveying", "how prior work has characterized moderation practices, with an emphasis on identifying where", "existing practices fall on the reactive-proactive spectrum and exploring the role of algorithmic", "tools.", "2.1Practices of Moderation", "In the popular consciousness, content moderation is often associated with the removal of illegal,", "hateful, or otherwise objectionable content [23], especially high-profile cases such as Reddit's 2015", "mass banning of hate communities [9]. However, removal of content is but one part of content", "moderation, which in reality involves a wide and diverse range of hidden labor [16, 48, 64] that,", "while less visible to the public eye than a high-profile ban, is no less important to maintaining the", "everyday functioning of an online community [29]", "The need to look beyond content removal is particularly salient in collaborative platforms like", "Wikipedia, which have eschewed top-down,platform-driven governance infavor of a community", "https://www.perspectiveapi.com", "Proc.ACM Hum.-Comput.Interact.,Vol. 6, No.CSCW2, Article 370.Publication date:November 2022." ]

{ "arxiv_id": "2211.16525", "language": "zh", "page_idx": 1 }

[ [ "0.09467", "0.08405", "0.39", "0.08405", "0.39", "0.0944", "0.09467", "0.0944" ], [ "0.869", "0.0833", "0.908", "0.0833", "0.908", "0.0958", "0.869", "0.0958" ], [ "0.1163", "0.12085", "0.3838", "0.12085", "0.3838", "0.1312", "0.1163", "0.1312" ], [ "0.4702", "0.12085", "0.534", "0.12085", "0.534", "0.1312", "0.4702", "0.1312" ], [ "0.568", "0.11945", "0.9043", "0.11945", "0.9043", "0.1333", "0.568", "0.1333" ], [ "0.5996", "0.1354", "0.9043", "0.1354", "0.9043", "0.1486", "0.5996", "0.1486" ], [ "0.3806", "0.1215", "0.4075", "0.1215", "0.4075", "0.1292", "0.3806", "0.1292" ], [ "0.0967", "0.1375", "0.3528", "0.1375", "0.3528", "0.148", "0.0967", "0.148" ], [ "0.09467", "0.1528", "0.3796", "0.1528", "0.3796", "0.1666", "0.09467", "0.1666" ], [ "0.534", "0.1528", "0.9023", "0.1528", "0.9023", "0.1666", "0.534", "0.1666" ], [ "0.0967", "0.1694", "0.4104", "0.1694", "0.4104", "0.1833", "0.0967", "0.1833" ], [ "0.4053", "0.175", "0.4187", "0.175", "0.4187", "0.1792", "0.4053", "0.1792" ], [ "0.4248", "0.175", "0.4434", "0.175", "0.4434", "0.1792", "0.4248", "0.1792" ], [ "0.4824", "0.175", "0.525", "0.175", "0.525", "0.1792", "0.4824", "0.1792" ], [ "0.54", "0.1694", "0.9053", "0.1694", "0.9053", "0.1841", "0.54", "0.1841" ], [ "0.09467", "0.1862", "0.512", "0.1862", "0.512", "0.2", "0.09467", "0.2" ], [ "0.4558", "0.1896", "0.4888", "0.1896", "0.4888", "0.1937", "0.4558", "0.1937" ], [ "0.11523", "0.2028", "0.9023", "0.2028", "0.9023", "0.2167", "0.11523", "0.2167" ], [ "0.0977", "0.2195", "0.9023", "0.2195", "0.9023", "0.234", "0.0977", "0.234" ], [ "0.0936", "0.2361", "0.604", "0.2361", "0.604", "0.2507", "0.0936", "0.2507" ], [ "0.09155", "0.2654", "0.5464", "0.2654", "0.5464", "0.279", "0.09155", "0.279" ], [ "0.09155", "0.2854", "0.9053", "0.2854", "0.9053", "0.3", "0.09155", "0.3" ], [ "0.0936", "0.3027", "0.9053", "0.3027", "0.9053", "0.3174", "0.0936", "0.3174" ], [ "0.0936", "0.3193", "0.9053", "0.3193", "0.9053", "0.3333", "0.0936", "0.3333" ], [ "0.09467", "0.3354", "0.9053", "0.3354", "0.9053", "0.35", "0.09467", "0.35" ], [ "0.09155", "0.3528", "0.5215", "0.3528", "0.5215", "0.3667", "0.09155", "0.3667" ], [ "0.1091", "0.3667", "0.907", "0.3674", "0.907", "0.3855", "0.1091", "0.384" ], [ "0.09155", "0.3862", "0.907", "0.3862", "0.907", "0.4", "0.09155", "0.4" ], [ "0.0936", "0.4028", "0.4229", "0.4028", "0.4229", "0.4167", "0.0936", "0.4167" ], [ "0.3118", "0.4062", "0.3323", "0.4062", "0.3323", "0.4104", "0.3118", "0.4104" ], [ "0.3374", "0.4062", "0.357", "0.4062", "0.357", "0.4104", "0.3374", "0.4104" ], [ "0.5186", "0.4006", "0.9053", "0.402", "0.9053", "0.4167", "0.5186", "0.4153" ], [ "0.285", "0.421", "0.4043", "0.421", "0.4043", "0.4312", "0.285", "0.4312" ], [ "0.538", "0.4194", "0.9053", "0.4194", "0.9053", "0.4333", "0.538", "0.4333" ], [ "0.3982", "0.4229", "0.4507", "0.4216", "0.4507", "0.4292", "0.3982", "0.43" ], [ "0.0967", "0.4229", "0.251", "0.4229", "0.251", "0.43", "0.0967", "0.43" ], [ "0.0936", "0.4375", "0.3015", "0.4375", "0.3015", "0.448", "0.0936", "0.448" ], [ "0.592", "0.4417", "0.616", "0.4417", "0.616", "0.4458", "0.592", "0.4458" ], [ "0.6387", "0.4375", "0.9053", "0.4353", "0.9053", "0.4487", "0.6387", "0.4507" ], [ "0.0936", "0.4521", "0.321", "0.4507", "0.321", "0.4646", "0.0936", "0.4653" ], [ "0.5464", "0.4548", "0.576", "0.4548", "0.576", "0.4624", "0.5464", "0.4624" ], [ "0.5845", "0.454", "0.9043", "0.454", "0.9043", "0.4646", "0.5845", "0.4646" ], [ "0.09155", "0.4688", "0.3796", "0.4688", "0.3796", "0.482", "0.09155", "0.482" ], [ "0.5215", "0.4688", "0.9053", "0.4688", "0.9053", "0.482", "0.5215", "0.482" ], [ "0.09467", "0.4854", "0.3633", "0.4854", "0.3633", "0.4985", "0.09467", "0.4985" ], [ "0.537", "0.4883", "0.57", "0.4883", "0.57", "0.4958", "0.537", "0.4958" ], [ "0.5874", "0.4854", "0.9053", "0.4854", "0.9053", "0.4985", "0.5874", "0.4985" ], [ "0.09467", "0.502", "0.461", "0.502", "0.461", "0.515", "0.09467", "0.515" ], [ "0.4824", "0.5005", "0.9043", "0.5005", "0.9043", "0.515", "0.4824", "0.515" ], [ "0.0936", "0.5186", "0.9053", "0.5186", "0.9053", "0.5317", "0.0936", "0.5317" ], [ "0.0936", "0.5356", "0.2212", "0.5356", "0.2212", "0.549", "0.0936", "0.549" ], [ "0.1121", "0.55", "0.9043", "0.55", "0.9043", "0.5645", "0.1121", "0.5645" ], [ "0.09155", "0.569", "0.9053", "0.569", "0.9053", "0.582", "0.09155", "0.582" ], [ "0.09155", "0.5854", "0.907", "0.5854", "0.907", "0.5986", "0.09155", "0.5986" ], [ "0.09467", "0.602", "0.9053", "0.602", "0.9053", "0.615", "0.09467", "0.615" ], [ "0.09467", "0.617", "0.9043", "0.617", "0.9043", "0.6313", "0.09467", "0.6313" ], [ "0.09467", "0.6353", "0.9053", "0.6353", "0.9053", "0.6484", "0.09467", "0.6484" ], [ "0.0905", "0.6484", "0.1965", "0.651", "0.1965", "0.665", "0.0905", "0.663" ], [ "0.11316", "0.6675", "0.9053", "0.6675", "0.9053", "0.682", "0.11316", "0.682" ], [ "0.0936", "0.684", "0.907", "0.684", "0.907", "0.6987", "0.0936", "0.6987" ], [ "0.0936", "0.7007", "0.501", "0.7007", "0.501", "0.7144", "0.0936", "0.7144" ], [ "0.64", "0.7026", "0.9043", "0.7026", "0.9043", "0.7134", "0.64", "0.7134" ], [ "0.09467", "0.7173", "0.3652", "0.717", "0.3652", "0.73", "0.09467", "0.7314" ], [ "0.6885", "0.7188", "0.9043", "0.7188", "0.9043", "0.732", "0.6885", "0.732" ], [ "0.0967", "0.7354", "0.3086", "0.7354", "0.3086", "0.7456", "0.0967", "0.7456" ], [ "0.7417", "0.7397", "0.775", "0.7397", "0.775", "0.7437", "0.7417", "0.7437" ], [ "0.7715", "0.7363", "0.9043", "0.7363", "0.9043", "0.7466", "0.7715", "0.7466" ], [ "0.09467", "0.7505", "0.329", "0.7505", "0.329", "0.7646", "0.09467", "0.7646" ], [ "0.7954", "0.753", "0.9043", "0.753", "0.9043", "0.763", "0.7954", "0.763" ], [ "0.09467", "0.7695", "0.1965", "0.7686", "0.1965", "0.78", "0.09467", "0.7812" ], [ "0.249", "0.773", "0.3179", "0.773", "0.3179", "0.777", "0.249", "0.777" ], [ "0.7573", "0.7695", "0.9043", "0.7695", "0.9043", "0.78", "0.7573", "0.78" ], [ "0.09155", "0.784", "0.2983", "0.7817", "0.2983", "0.796", "0.09155", "0.798" ], [ "0.711", "0.786", "0.9043", "0.786", "0.9043", "0.7964", "0.711", "0.7964" ], [ "0.09467", "0.8027", "0.2715", "0.8027", "0.2715", "0.813", "0.09467", "0.813" ], [ "0.778", "0.8027", "0.8994", "0.8027", "0.8994", "0.813", "0.778", "0.813" ], [ "0.1214", "0.8193", "0.2654", "0.8193", "0.2654", "0.83", "0.1214", "0.83" ], [ "0.3416", "0.821", "0.4011", "0.821", "0.4011", "0.8276", "0.3416", "0.8276" ], [ "0.746", "0.8174", "0.9023", "0.8174", "0.9023", "0.831", "0.746", "0.831" ], [ "0.09467", "0.834", "0.3108", "0.834", "0.3108", "0.848", "0.09467", "0.848" ], [ "0.3262", "0.8384", "0.3867", "0.8384", "0.3867", "0.8423", "0.3262", "0.8423" ], [ "0.7397", "0.834", "0.9023", "0.834", "0.9023", "0.848", "0.7397", "0.848" ], [ "0.09467", "0.8506", "0.4104", "0.8506", "0.4104", "0.8647", "0.09467", "0.8647" ], [ "0.6367", "0.855", "0.7183", "0.855", "0.7183", "0.859", "0.6367", "0.859" ], [ "0.715", "0.8506", "0.9043", "0.8525", "0.9043", "0.8633", "0.715", "0.8613" ], [ "0.09467", "0.867", "0.535", "0.867", "0.535", "0.8813", "0.09467", "0.8813" ], [ "0.568", "0.867", "0.9053", "0.867", "0.9053", "0.8813", "0.568", "0.8813" ], [ "0.0936", "0.884", "0.9053", "0.884", "0.9053", "0.898", "0.0936", "0.898" ], [ "0.0936", "0.901", "0.9053", "0.901", "0.9053", "0.9146", "0.0936", "0.9146" ], [ "0.214", "0.934", "0.9053", "0.934", "0.9053", "0.9478", "0.214", "0.9478" ] ]

[ "ProactiveModerationof OnlineDiscussions", "370:5", "但是，大多数平台都避免使用", "因为它", "发许多实际问题。预筛选是高度劳动", "使在算法预筛选的帮助下，《纽约", "", "密集型的、并随着平台的增", "时报》目前也只允许在工作日对", "时的评论。此外，预筛选通过在用户提交", "内容与版主审查平台上出现的内容", "", "E", "", "从而阻止了平台上用户之间的实时交互。", "最后，预筛选以压制言论自由[231为由受到批", "", "虽然预筛选可能会阻止反社会行为到达普通受众，但它仍然依赖于版主对用户在其平台", "上进行的所有反社会行为尝试做出反应和解决-这种反应刚刚从公共广场转移到版主的私人", "空间。这种转变通过再现上述问题限制了预筛选的有效性", "2.4ProactiveModerationofOnlineDiscussion", "In online discussion platforms,proactive approaches tomoderation can aim to discourage undesir", "able actions or to encourage prosocial behavior and productive conversations[58, 59].Proactive", "strategies can range from static design decisions to dynamic interventions in which moderators", "play a more active role. They can also be broadly applied to the entire community or targeted", "towards specific situations, users, or conversations.", "Static strategies primarily involve deliberate choices in platform design aimed at promoting pro-", "social behaviors.These have a long and established history in social computing; now-common de-", "sign choices such as activity indicators", "TF.", "c", "licitlylisted rules[40]wereinitiallydeveloped", "agethedevelop", "ion of social normswithin online communi", "pment", "asmeasurestoenc", "ties.Morerecent develop", "Ta", "ns on communitysizeorrateof", "participation[26]and syste", "afleo", "moredeeplyoncommentstheyhave", "read[33,43,44].While suchdesig", "historicallybeen broadly applied at thelevel", "of the entireplatform, there ha", "evel", "terventions that are more targeted to", "specificsituations[31,47].Inthespaceof o", "ne discussions,such work has includedpsychologi", "cal priming strategies that are deployed to users when they are about to comment in a discussion", "thread[56, 61].", "As platforms have grown and evolved, they have developed more dynamic strategies for proac", "tive community management, in which moderators take a more direct role.For instance, a natu", "ral development from staticlisting of rules involves moderatorspersonally sending reminders of", "communityrules inhigh-impact situations, such as whenwelcomingnewcomers[59].As a fur", "ther step from this, recent work has looked at howmoderators can model good behavior in their", "owninteractions,asawayofimplicitlysignalingtothe communitywhatproperbehaviorlooks", "like[33,58].", "One dynamic strategy in particular has picked up increasing interest: moderators actively mon", "itor ongoing conversations in order to proactively prevent them from derailing into uncivil be-", "havior or, at least, to be in a position that allow", "tetheeffectsofderailmentina", "timelymanner[39,47].Unliketh", "e, this dynamic and highly", "targetedmoderationstrat", "Te:", "moderatorsand", "thus scales poorly.As such", "oportforthis", "typeofstrat", "TeTorr", "isthatmightsoon", "devolve\"that could help", "eded\"[59].Thiscurrent", "workaimstodevelop", "stitsfeasibility", "与筛选前的节制", "E", "论自由的道德关", "注。我们参与这些问题，", "", "例如仅根据对未来", "事件的预测来关闭对话或主动指责", "", "Whiletheseconcerns", "arewarranted anddeservingofameaningfulinquir", "rearguethatthewell-established harms", "tousers, platforms, and moderators in the popular reactive moderation framework motivate us to", "explore alternate approaches tomoderation.Because the status quo is not a neutral ethical choice,", "Proc.ACM Hum.-Comput. Interact.,Vol.6,No.CSCW2,Article 370.Publication date: November 2022" ]

{ "arxiv_id": "2211.16525", "language": "zh", "page_idx": 4 }

[ [ "0.3064", "0.11554", "0.6685", "0.11554", "0.6685", "0.13", "0.3064", "0.13" ], [ "0.1772", "0.1768", "0.7915", "0.1768", "0.7915", "0.1913", "0.1772", "0.1913" ], [ "0.0866", "0.1932", "0.152", "0.1957", "0.152", "0.2102", "0.0866", "0.2084" ], [ "0.2427", "0.2052", "0.3203", "0.207", "0.3203", "0.2247", "0.2427", "0.2229" ], [ "0.5024", "0.209", "0.884", "0.209", "0.884", "0.2229", "0.5024", "0.2229" ], [ "0.5034", "0.226", "0.884", "0.226", "0.884", "0.2361", "0.5034", "0.2361" ], [ "0.1217", "0.2374", "0.4395", "0.2374", "0.4395", "0.2507", "0.1217", "0.2507" ], [ "0.5024", "0.2393", "0.8857", "0.2393", "0.8857", "0.2524", "0.5024", "0.2524" ], [ "0.12335", "0.254", "0.4395", "0.254", "0.4395", "0.2644", "0.12335", "0.2644" ], [ "0.5024", "0.2556", "0.884", "0.2556", "0.884", "0.2664", "0.5024", "0.2664" ], [ "0.1217", "0.2676", "0.4412", "0.2676", "0.4412", "0.281", "0.1217", "0.281" ], [ "0.5024", "0.2695", "0.884", "0.2695", "0.884", "0.2827", "0.5024", "0.2827" ], [ "0.1217", "0.2822", "0.4412", "0.2822", "0.4412", "0.2954", "0.1217", "0.2954" ], [ "0.5024", "0.2854", "0.8857", "0.2854", "0.8857", "0.2986", "0.5024", "0.2986" ], [ "0.1217", "0.2966", "0.4395", "0.2966", "0.4395", "0.31", "0.1217", "0.31" ], [ "0.5034", "0.3018", "0.8823", "0.3018", "0.8823", "0.3125", "0.5034", "0.3125" ], [ "0.1217", "0.3125", "0.4412", "0.3125", "0.4412", "0.3257", "0.1217", "0.3257" ], [ "0.5024", "0.315", "0.8857", "0.315", "0.8857", "0.3289", "0.5024", "0.3289" ], [ "0.1217", "0.3276", "0.4412", "0.3276", "0.4412", "0.3416", "0.1217", "0.3416" ], [ "0.5024", "0.3308", "0.8857", "0.3308", "0.8857", "0.3447", "0.5024", "0.3447" ], [ "0.1217", "0.3423", "0.4412", "0.3423", "0.4412", "0.356", "0.1217", "0.356" ], [ "0.5024", "0.3467", "0.6914", "0.3467", "0.6914", "0.3599", "0.5024", "0.3599" ], [ "0.1217", "0.358", "0.4412", "0.358", "0.4412", "0.3718", "0.1217", "0.3718" ], [ "0.1209", "0.3726", "0.4412", "0.3726", "0.4412", "0.3865", "0.1209", "0.3865" ], [ "0.5024", "0.3687", "0.8857", "0.3687", "0.8857", "0.3818", "0.5024", "0.3818" ], [ "0.11926", "0.387", "0.4412", "0.3882", "0.4412", "0.4028", "0.11926", "0.4016" ], [ "0.5024", "0.3833", "0.8823", "0.3833", "0.8823", "0.3965", "0.5024", "0.3965" ], [ "0.501", "0.3965", "0.8857", "0.3977", "0.8857", "0.411", "0.501", "0.4104" ], [ "0.1217", "0.404", "0.4395", "0.404", "0.4395", "0.4172", "0.1217", "0.4172" ], [ "0.5024", "0.4124", "0.8857", "0.4136", "0.8857", "0.4268", "0.5024", "0.4255" ], [ "0.1209", "0.4185", "0.4395", "0.4185", "0.4395", "0.432", "0.1209", "0.432" ], [ "0.5024", "0.4268", "0.8857", "0.4287", "0.8857", "0.4426", "0.5024", "0.44" ], [ "0.1217", "0.433", "0.4395", "0.433", "0.4395", "0.4463", "0.1217", "0.4463" ], [ "0.1217", "0.4475", "0.4395", "0.4475", "0.4395", "0.461", "0.1217", "0.461" ], [ "0.501", "0.443", "0.8384", "0.443", "0.8384", "0.457", "0.501", "0.457" ], [ "0.1217", "0.4634", "0.4412", "0.4634", "0.4412", "0.4766", "0.1217", "0.4766" ], [ "0.5034", "0.4666", "0.884", "0.4666", "0.884", "0.4797", "0.5034", "0.4797" ], [ "0.12335", "0.4797", "0.4387", "0.4797", "0.4387", "0.4905", "0.12335", "0.4905" ], [ "0.5024", "0.481", "0.8857", "0.481", "0.8857", "0.4944", "0.5024", "0.4944" ], [ "0.12335", "0.4956", "0.4395", "0.4956", "0.4395", "0.506", "0.12335", "0.506" ], [ "0.5024", "0.4968", "0.8857", "0.4968", "0.8857", "0.5103", "0.5024", "0.5103" ], [ "0.1209", "0.507", "0.4395", "0.5083", "0.4395", "0.523", "0.1209", "0.5215" ], [ "0.5034", "0.513", "0.884", "0.513", "0.884", "0.5234", "0.5034", "0.5234" ], [ "0.1217", "0.526", "0.437", "0.526", "0.437", "0.536", "0.1217", "0.536" ], [ "0.5024", "0.526", "0.8857", "0.526", "0.8857", "0.539", "0.5024", "0.539" ], [ "0.1209", "0.539", "0.4395", "0.539", "0.4395", "0.553", "0.1209", "0.553" ], [ "0.5024", "0.5415", "0.8857", "0.5415", "0.8857", "0.5547", "0.5024", "0.5547" ], [ "0.1209", "0.5537", "0.4395", "0.553", "0.4395", "0.5674", "0.1209", "0.5684" ], [ "0.501", "0.5547", "0.884", "0.556", "0.884", "0.571", "0.501", "0.5693" ], [ "0.1217", "0.5693", "0.4412", "0.5693", "0.4412", "0.5835", "0.1217", "0.5835" ], [ "0.501", "0.5713", "0.8823", "0.5713", "0.8823", "0.5854", "0.501", "0.5854" ], [ "0.1217", "0.5854", "0.2942", "0.5854", "0.2942", "0.5986", "0.1217", "0.5986" ], [ "0.5024", "0.5884", "0.884", "0.5884", "0.884", "0.5986", "0.5024", "0.5986" ], [ "0.5024", "0.603", "0.55", "0.603", "0.55", "0.6143", "0.5024", "0.6143" ], [ "0.501", "0.625", "0.8857", "0.625", "0.8857", "0.638", "0.501", "0.638" ], [ "0.08905", "0.634", "0.2214", "0.634", "0.2214", "0.648", "0.08905", "0.648" ], [ "0.5024", "0.64", "0.8857", "0.64", "0.8857", "0.654", "0.5024", "0.654" ], [ "0.0907", "0.66", "0.4722", "0.66", "0.4722", "0.674", "0.0907", "0.674" ], [ "0.5024", "0.655", "0.8823", "0.655", "0.8823", "0.6685", "0.5024", "0.6685" ], [ "0.0907", "0.6753", "0.4739", "0.6753", "0.4739", "0.6895", "0.0907", "0.6895" ], [ "0.5024", "0.672", "0.8857", "0.672", "0.8857", "0.6816", "0.5024", "0.6816" ], [ "0.5", "0.683", "0.884", "0.6846", "0.884", "0.6987", "0.5", "0.6978" ], [ "0.09235", "0.69", "0.4722", "0.69", "0.4722", "0.704", "0.09235", "0.704" ], [ "0.08905", "0.705", "0.4722", "0.705", "0.4722", "0.7183", "0.08905", "0.7183" ], [ "0.501", "0.6997", "0.8857", "0.6997", "0.8857", "0.7134", "0.501", "0.7134" ], [ "0.08905", "0.7197", "0.4722", "0.7197", "0.4722", "0.733", "0.08905", "0.733" ], [ "0.501", "0.714", "0.8823", "0.7153", "0.8823", "0.729", "0.501", "0.728" ], [ "0.08826", "0.7344", "0.4722", "0.735", "0.4722", "0.7485", "0.08826", "0.7476" ], [ "0.5024", "0.73", "0.8857", "0.73", "0.8857", "0.7437", "0.5024", "0.7437" ], [ "0.0907", "0.7505", "0.4722", "0.7505", "0.4722", "0.7646", "0.0907", "0.7646" ], [ "0.5034", "0.747", "0.581", "0.747", "0.581", "0.7583", "0.5034", "0.7583" ], [ "0.0907", "0.765", "0.4714", "0.765", "0.4714", "0.779", "0.0907", "0.779" ], [ "0.08905", "0.78", "0.1299", "0.782", "0.1283", "0.7935", "0.08826", "0.7915" ], [ "0.10864", "0.8037", "0.4722", "0.8037", "0.4722", "0.8184", "0.10864", "0.8184" ], [ "0.08905", "0.8164", "0.1528", "0.8193", "0.152", "0.8306", "0.08826", "0.8276" ], [ "0.1462", "0.8174", "0.4722", "0.8174", "0.4722", "0.8306", "0.1462", "0.8306" ], [ "0.08905", "0.83", "0.4722", "0.83", "0.4722", "0.8433", "0.08905", "0.8433" ], [ "0.08496", "0.8403", "0.3064", "0.8423", "0.3064", "0.857", "0.08496", "0.855" ], [ "0.08905", "0.866", "0.4739", "0.866", "0.4739", "0.881", "0.08905", "0.881" ], [ "0.08826", "0.879", "0.4739", "0.88", "0.4739", "0.8945", "0.08826", "0.893" ], [ "0.0866", "0.8916", "0.2524", "0.893", "0.2524", "0.907", "0.0866", "0.905" ] ]

[ "Secure Distributed Training at Scale", "Eduard Gorbunov *123 Alexander Borzunov * 43 Michael Diskin 43 Max Ryabinin 43", "???。3in", "Abstract", "最近，类似的技术被提出用于深度学习。他们的目标是", "解决许多机器学习任务的绝对计算复杂性所带来的挑", "Many areas of deep learning benefit from us", "战，例如为NLP（Devlinetal.,2019;Brownetal.，2020;", "ing increasingly larger neural networks trained", "Liuetal..2019）预训练变压器或在视觉（Sunetal.", "on public data, as is the case for pre-trained", "2017;Kolesnikov etal.,2020;Goyaletal.,2021)中学习", "models for NLP and computer vision. Training", "巨大的数据集。最近的工作（Kijsipongseetal.，2018；", "such models requires a lot of computational re.", "Ryabinin &Gusev,2020;Atre etal.,202l:Diskin etal.", "sources (e.g., HPC clusters) that are not avail-", "2021）提出了几个系统，在许多志愿者中共享计算，这", "able to small research groups and independent", "些志愿者捐赠了他们的计算机totrainlargemodelson", "researchers. One way to address it is for sev-", "publicdatasets.的空闲时间", "eral smaller groups to pool their computational", "resources together and train a model that benefits", "尽管志愿者计算系统有其优势，但到目前为止，它们", "all participants. Unfortunately, in this case, any", "的实际应用有限（Kijsipongseetal.，2018）。全球采用", "这些技术的主要障碍是信任每个参与者的可靠性。对", "participant can jeopardize the entire training run", "于分布式训练，如果单个对等方由于计算中的错误而", "by sending incorrect updates, deliberately or by", "发送不正确的输出，则可能会破坏协作所取得的所有", "mistake. Training in presence of such peers re", "quires specialized distributed training algorithms", "进展（Smith，2019）或恶意（Tolpeginetal.,2020)", "with Byzantine tolerance. These algorithms of-", "分散优化中的现有技术提出了几种可抵抗此类“拜占", "ten sacrificeefficiencybyintroducingredundant", "庭”故障的优化算法。但是，大多数容忍拜占庭的训", "communicationorpassingallupdatesthrougha", "练协议都需要通过受信任的中央服务器传递所有更新，", "trusted server, making it infeasible to apply them", "或者交换额外的消息：从而使网络负载增加数倍（Chen", "tolarge-scaledeeplearning,wheremodelscar", "etal.，2018;Rajputetal.，2019）。这是大规模分布式深", "have billions of parameters. In this work, we", "度学习的一个主要问题，数百个对等点必须定期交换", "propose a novel protocol for secure (Byzantine", "数百万个参数的更新（Lietal.，2020Sergeev&Balso", "tolerant) decentralized training that emphasizes", "2018;Shoeybietal.,2019）。因此，在许多实际场景中", "communicationefficiency", "拜占庭容忍算法的计算和通信开销超过了与他人协作", "的好处", "在这项工作中，我们着手通过提出一种新颖的拜占", "1.Introduction", "庭容忍分布式训练协议来解决此问题，该协议专为", "通过许多国家，团体和个人之间的合作解决了许多", "大规模深度学习工作负载而设计。我们的方法结合了", "棘手的科学问题。这在自然科学中尤为明显，研究", "现代分布式训练技术的可扩展性和通信效率，如All-", "ReduceSGD（Sergeev&Balso，2018），以及针对拜占庭", "人员形成了跨国合作来运行大规模实验并共享计算", "基础设施（Aad et al.,2012;Ruttley et al.,2017;Abbott", "和Sybil攻击者的弹性。为了实现这一目标，我们利", "etal.,2016)。像Folding@home(Beberg etal.,2009)和", "用加密技术来验证训练的完整性，并将开销降至最低", "BOINC（Anderson，2004)这样的项目通过招募志愿者", "而不依赖于模型大小。我们的协议不需要信任任何特", "来进一步推动这一趋势，这些志愿者捐赠了他们的", "定的对等方", "计算，以前所未有的规模集体运行计算实验（Merritt", "2020)", "Equal contribution MIPT 2 Mila - Quebec AI Institute", "3Yandex", "4HSE University :Correspondence to: Eduard Gor", "bunov<[email protected] >, Alexander Borzunov", "< [email protected] >.", "Proceedings of the 3gth International Conference on Machine", "Learning, Baltimore, Maryland, USA, PMLR 162, 2022. Copy", "right2022by the author(s)." ]

{ "arxiv_id": "2106.11257", "language": "zh", "page_idx": 0 }

[ [ "0.3743", "0.05936", "0.5986", "0.05936", "0.5986", "0.06946", "0.3743", "0.06946" ], [ "0.08905", "0.0871", "0.2737", "0.0871", "0.2737", "0.09784", "0.08905", "0.09784" ], [ "0.5522", "0.089", "0.5786", "0.089", "0.5786", "0.09784", "0.5522", "0.09784" ], [ "0.6772", "0.089", "0.709", "0.089", "0.709", "0.09784", "0.6772", "0.09784" ], [ "0.7935", "0.0884", "0.8325", "0.0884", "0.8325", "0.0985", "0.7935", "0.0985" ], [ "0.7363", "0.09155", "0.767", "0.09155", "0.767", "0.09534", "0.7363", "0.09534" ], [ "0.609", "0.0903", "0.636", "0.0903", "0.636", "0.09784", "0.609", "0.09784" ], [ "0.1013", "0.1048", "0.1217", "0.1048", "0.1217", "0.11743", "0.1013", "0.11743" ], [ "0.609", "0.11237", "0.6396", "0.11237", "0.6396", "0.1276", "0.609", "0.1276" ], [ "0.6445", "0.1111", "0.71", "0.1111", "0.71", "0.13", "0.6445", "0.13" ], [ "0.8486", "0.11237", "0.8774", "0.11237", "0.8774", "0.125", "0.8486", "0.125" ], [ "0.08905", "0.1364", "0.4697", "0.1364", "0.4697", "0.1497", "0.08905", "0.1497" ], [ "0.509", "0.113", "0.54", "0.113", "0.54", "0.1256", "0.509", "0.1256" ], [ "0.6167", "0.1433", "0.7114", "0.137", "0.716", "0.1674", "0.621", "0.1736" ], [ "0.10297", "0.1509", "0.4739", "0.1509", "0.4739", "0.1642", "0.10297", "0.1642" ], [ "0.5107", "0.1528", "0.5293", "0.1559", "0.527", "0.1635", "0.5073", "0.1604" ], [ "0.863", "0.1559", "0.869", "0.1559", "0.869", "0.159", "0.863", "0.159" ], [ "0.10455", "0.1666", "0.4697", "0.1666", "0.4697", "0.1799", "0.10455", "0.1799" ], [ "0.10297", "0.1818", "0.2737", "0.1818", "0.2737", "0.1957", "0.10297", "0.1957" ], [ "0.6772", "0.1913", "0.71", "0.1913", "0.71", "0.2026", "0.6772", "0.2026" ], [ "0.8486", "0.1913", "0.8774", "0.1913", "0.8774", "0.2026", "0.8486", "0.2026" ], [ "0.728", "0.1925", "0.761", "0.1925", "0.761", "0.202", "0.728", "0.202" ], [ "0.5073", "0.1925", "0.539", "0.1925", "0.539", "0.2039", "0.5073", "0.2039" ], [ "0.6045", "0.1932", "0.638", "0.1932", "0.638", "0.2052", "0.6045", "0.2052" ], [ "0.6587", "0.1945", "0.675", "0.1945", "0.675", "0.2026", "0.6587", "0.2026" ], [ "0.759", "0.1945", "0.7646", "0.1945", "0.7646", "0.2008", "0.759", "0.2008" ], [ "0.589", "0.1995", "0.6045", "0.1976", "0.6055", "0.2026", "0.591", "0.2052" ], [ "0.08905", "0.1945", "0.1095", "0.1945", "0.1095", "0.2052", "0.08905", "0.2052" ], [ "0.08905", "0.2184", "0.4714", "0.2184", "0.4714", "0.2317", "0.08905", "0.2317" ], [ "0.5024", "0.2247", "0.8813", "0.2247", "0.8813", "0.238", "0.5024", "0.238" ], [ "0.10297", "0.2329", "0.4722", "0.2329", "0.4722", "0.2468", "0.10297", "0.2468" ], [ "0.5", "0.2374", "0.886", "0.238", "0.886", "0.2524", "0.5", "0.252" ], [ "0.10297", "0.2488", "0.4739", "0.2488", "0.4739", "0.262", "0.10297", "0.262" ], [ "0.5024", "0.2524", "0.544", "0.2524", "0.544", "0.2644", "0.5024", "0.2644" ], [ "0.1013", "0.2632", "0.4722", "0.262", "0.4722", "0.2766", "0.1013", "0.2778" ], [ "0.1013", "0.2798", "0.3245", "0.279", "0.3245", "0.2922", "0.1013", "0.2935" ], [ "0.0907", "0.2903", "0.10864", "0.2903", "0.10864", "0.2986", "0.0907", "0.2986" ], [ "0.5024", "0.2854", "0.8857", "0.2854", "0.8857", "0.2998", "0.5024", "0.2998" ], [ "0.5034", "0.303", "0.7295", "0.303", "0.7295", "0.3132", "0.5034", "0.3132" ], [ "0.08826", "0.315", "0.4697", "0.315", "0.4697", "0.33", "0.08826", "0.33" ], [ "0.5024", "0.324", "0.8823", "0.324", "0.8823", "0.3372", "0.5024", "0.3372" ], [ "0.10297", "0.33", "0.4722", "0.33", "0.4722", "0.3447", "0.10297", "0.3447" ], [ "0.5024", "0.3384", "0.8823", "0.3384", "0.8823", "0.3516", "0.5024", "0.3516" ], [ "0.10297", "0.3467", "0.4722", "0.3467", "0.4722", "0.3599", "0.10297", "0.3599" ], [ "0.5034", "0.3547", "0.5596", "0.3547", "0.5596", "0.3655", "0.5034", "0.3655" ], [ "0.558", "0.3516", "0.8857", "0.3528", "0.8857", "0.3674", "0.558", "0.3662" ], [ "0.0768", "0.3694", "0.125", "0.3542", "0.1356", "0.3738", "0.08826", "0.3901" ], [ "0.5024", "0.3687", "0.8857", "0.3687", "0.8857", "0.3818", "0.5024", "0.3818" ], [ "0.5024", "0.3838", "0.884", "0.3838", "0.884", "0.3977", "0.5024", "0.3977" ], [ "0.08826", "0.3965", "0.4722", "0.3958", "0.4722", "0.4104", "0.08826", "0.411" ], [ "0.5034", "0.401", "0.884", "0.401", "0.884", "0.411", "0.5034", "0.411" ], [ "0.10297", "0.4124", "0.4722", "0.4124", "0.4722", "0.4255", "0.10297", "0.4255" ], [ "0.501", "0.414", "0.884", "0.414", "0.884", "0.428", "0.501", "0.428" ], [ "0.1013", "0.428", "0.4722", "0.4255", "0.4722", "0.4395", "0.1013", "0.4414" ], [ "0.501", "0.428", "0.8857", "0.4287", "0.8857", "0.4426", "0.501", "0.4414" ], [ "0.1013", "0.4414", "0.4722", "0.4426", "0.4722", "0.457", "0.1013", "0.4558" ], [ "0.5024", "0.443", "0.8857", "0.4443", "0.8857", "0.4583", "0.5024", "0.457" ], [ "0.10297", "0.4583", "0.4697", "0.4583", "0.4697", "0.4717", "0.10297", "0.4717" ], [ "0.501", "0.4602", "0.884", "0.4602", "0.884", "0.4734", "0.501", "0.4734" ], [ "0.501", "0.4734", "0.884", "0.4749", "0.884", "0.4893", "0.501", "0.488" ], [ "0.0866", "0.4905", "0.2279", "0.4924", "0.2279", "0.507", "0.0866", "0.505" ], [ "0.5024", "0.4912", "0.589", "0.4912", "0.589", "0.5024", "0.5024", "0.5024" ], [ "0.0866", "0.518", "0.2998", "0.518", "0.2998", "0.532", "0.0866", "0.532" ], [ "0.5", "0.5195", "0.7563", "0.521", "0.7563", "0.535", "0.5", "0.534" ], [ "0.09235", "0.5415", "0.4722", "0.5415", "0.4722", "0.5547", "0.09235", "0.5547" ], [ "0.501", "0.5444", "0.8857", "0.5444", "0.8857", "0.558", "0.501", "0.558" ], [ "0.08905", "0.556", "0.4722", "0.5566", "0.4722", "0.571", "0.08905", "0.5693" ], [ "0.501", "0.56", "0.884", "0.56", "0.884", "0.5737", "0.501", "0.5737" ], [ "0.08905", "0.571", "0.4739", "0.5713", "0.4739", "0.586", "0.08905", "0.5854" ], [ "0.5024", "0.5747", "0.8823", "0.5747", "0.8823", "0.5884", "0.5024", "0.5884" ], [ "0.0907", "0.587", "0.4739", "0.587", "0.4739", "0.601", "0.0907", "0.601" ], [ "0.5024", "0.5903", "0.884", "0.5903", "0.884", "0.604", "0.5024", "0.604" ], [ "0.0907", "0.6016", "0.4739", "0.6016", "0.4739", "0.6157", "0.0907", "0.6157" ], [ "0.501", "0.606", "0.6177", "0.606", "0.6177", "0.619", "0.501", "0.619" ], [ "0.0907", "0.6157", "0.4714", "0.616", "0.4714", "0.631", "0.0907", "0.63" ], [ "0.0907", "0.632", "0.4697", "0.632", "0.4697", "0.646", "0.0907", "0.646" ], [ "0.5024", "0.6274", "0.8823", "0.6274", "0.8823", "0.6416", "0.5024", "0.6416" ], [ "0.08905", "0.6465", "0.4739", "0.6465", "0.4739", "0.66", "0.08905", "0.66" ], [ "0.501", "0.643", "0.8857", "0.643", "0.8857", "0.657", "0.501", "0.657" ], [ "0.0907", "0.662", "0.4722", "0.662", "0.4722", "0.6753", "0.0907", "0.6753" ], [ "0.501", "0.6577", "0.884", "0.657", "0.884", "0.672", "0.501", "0.6724" ], [ "0.5024", "0.6724", "0.8813", "0.6724", "0.8813", "0.686", "0.5024", "0.686" ], [ "0.0907", "0.6787", "0.4714", "0.6787", "0.4714", "0.6895", "0.0907", "0.6895" ], [ "0.0907", "0.6924", "0.4722", "0.6924", "0.4722", "0.7056", "0.0907", "0.7056" ], [ "0.5024", "0.6895", "0.884", "0.6895", "0.884", "0.6997", "0.5024", "0.6997" ], [ "0.0907", "0.7085", "0.4722", "0.7085", "0.4722", "0.7217", "0.0907", "0.7217" ], [ "0.501", "0.7026", "0.8823", "0.7026", "0.8823", "0.7163", "0.501", "0.7163" ], [ "0.08905", "0.7227", "0.4722", "0.7217", "0.4722", "0.7363", "0.08905", "0.7373" ], [ "0.5024", "0.7173", "0.8813", "0.7173", "0.8813", "0.731", "0.5024", "0.731" ], [ "0.08905", "0.7373", "0.4739", "0.7363", "0.4739", "0.7505", "0.08905", "0.752" ], [ "0.501", "0.7344", "0.884", "0.7344", "0.884", "0.7476", "0.501", "0.7476" ], [ "0.08905", "0.752", "0.4656", "0.752", "0.4656", "0.765", "0.08905", "0.765" ], [ "0.5024", "0.7485", "0.8823", "0.7485", "0.8823", "0.762", "0.5024", "0.762" ], [ "0.09235", "0.7676", "0.4722", "0.7676", "0.4722", "0.781", "0.09235", "0.781" ], [ "0.501", "0.765", "0.584", "0.765", "0.584", "0.7764", "0.501", "0.7764" ], [ "0.09235", "0.7827", "0.4722", "0.7827", "0.4722", "0.797", "0.09235", "0.797" ], [ "0.5024", "0.786", "0.8823", "0.786", "0.8823", "0.8", "0.5024", "0.8" ], [ "0.0907", "0.7974", "0.4722", "0.7974", "0.4722", "0.811", "0.0907", "0.811" ], [ "0.501", "0.802", "0.8813", "0.802", "0.8813", "0.815", "0.501", "0.815" ], [ "0.0907", "0.813", "0.4739", "0.813", "0.4739", "0.827", "0.0907", "0.827" ], [ "0.501", "0.8174", "0.8823", "0.8174", "0.8823", "0.8306", "0.501", "0.8306" ], [ "0.09235", "0.83", "0.4714", "0.83", "0.4714", "0.8403", "0.09235", "0.8403" ], [ "0.5024", "0.8335", "0.8813", "0.8335", "0.8813", "0.8433", "0.5024", "0.8433" ], [ "0.08905", "0.8433", "0.4722", "0.8433", "0.4722", "0.857", "0.08905", "0.857" ], [ "0.5", "0.8447", "0.8857", "0.8467", "0.8857", "0.86", "0.5", "0.858" ], [ "0.1095", "0.8657", "0.4756", "0.866", "0.4756", "0.881", "0.1095", "0.88" ], [ "0.5034", "0.863", "0.8813", "0.863", "0.8813", "0.8735", "0.5034", "0.8735" ], [ "0.08905", "0.881", "0.4429", "0.881", "0.4429", "0.8945", "0.08905", "0.8945" ], [ "0.5024", "0.877", "0.8823", "0.877", "0.8823", "0.89", "0.5024", "0.89" ], [ "0.5", "0.8926", "0.8823", "0.8926", "0.8823", "0.906", "0.5", "0.906" ] ]

[ "Secure Distributed Training at Scale", "我们的贡献可以总结如下：", "SPLIT", "REDUG", "MERG", "ATT", "CATTE", "px", "", "", "=", ":We propose a novel protocol for decentralized Byzantine", "91=", "VD", "tolerant training on data available to all participants,", "132", "", "where the extra communication cost does not depend or", "the number of parameters.", "", "=9", "", "93=", "", "", "一", "L", "px", ": We rigorously analyze this protocol and prove conver", "Figure 1. A scheme of Butterfly All-Reduce (Li et al., 2017)", "gence bounds for convex and non-convex losses with", "Each peer transfers only O(d) data when averaging a vector of", "Byzantine attackers. Furthermore, we derive accelerated", "size d.", "convergence rates for the same task under realistic as", "sumptions about modelgradients.", "px", "和本地更新（Zinkevichetal.，2010）部分解决了这个问", "题，但它仍然是这些方法的瓶颈", ": We propose a heuristic for resisting Sybil attacks from", "实际上，大多数分布式培训系统都利用All-Reduce", "computationally constrained attackers, allowing to ac", "(AR)E(Goyal et al., 2017; Mikami et al.,2019; You et al.", "cept new untrusted peers joining midway through train", "2020)--", "一系列集体通信协议，允许服务器平均其数", "pxing.", "据并在每台机器上接收结果。由此产生的名为All-", "ReduceSGD（AR-SGD）的方法在每个对等方的局部梯", "We verify the effectiveness of our algorithm in con", "度上运行AR，以计算全局平均值。通常，AR-SGD使", "trolled experiments ' and actual large-scale training runs.", "用带宽最佳版本的All-ReduceE(Sergeev&Balso,2018：", "Specifically,westart withResNet-18forCIFAR-10clas", "Patarasuk&Yuan，2009），例如ButterflyAll-Reduce（请", "sification and follow up with pretraining ALBERT-large", "参见图1）。根据确切的算法，它们要求每个对等体", "in a setup where almost a half of all peers are malicious.", "在n对等体上平均大小为d的向量时仅传输O(d)或", "O(dlogn）数据（与基于PS的方法不同，其中PS传输", "2. Related work", "O（dn）数据）", "2.1. Distributed deep learning", "2.2.Byzantine-tolerant optimization", "训练现代神经网络通常需要在任何一台机器上都无法", "标准的分布式训练方法对拜占庭攻击并不稳健。在", "实现的计算量。人们必须使用分布式训练方法在多台", "vanilla并行SGD中，一个恶意工作者可以通过以任意", "机器上训练这样的模型。这些方法中的大多数分为两", "方式移动结果向量的均值来破坏整个方法的收敛性", "组：在数据并行训练中，每个工人通过从训练数据E", "因此，研究界发明了特殊的算法，即使在这种设置下", "(Sergeev&Balso,2018Goyaletal.，2017）中采样批次", "也可以训练模型", "来训练整个模型；相反，模型并行训练将模型的各个", "部分分配给不同的工人E（Huangetal.，2019；Narayanar", "基于参数服务器(PS)的方法。设计为具有拜占庭弹性", "etal.2019;Shoeybietal.，2019）。在这项研究中，我们只", "的大多数算法都依赖于可信参数服务器的存在。在这", "考虑第一组；值得注意的是，大多数模型并行系统仍然", "种方法中，标准均值估计器，e。g。，并行SGD中使", "用的那个，通常被更健壮的聚合规则（Blanchardetal.", "依赖于同一阶段节点之间的数据并行性（Rajbhandari", "etal.，2020；Narayananetal.，2021）。通常，数据并行", "2017:Yinetal.,2018:Damaskinosetal..2019:ElMhamdi", "训练包括两个阶段：首先，每个worker计算其数据上", "etal.2018;Pillutlaetal.,2019）替换。但是，最近的工作", "的梯度；然后，所有worker聚合梯度并运行SGD步", "表明，通过提出特殊类型的拜占庭攻击E(Baruchetal.", "骤。最简单的聚合策略称为ParameterServers（PS）E", "2019；Xieetal.，2020）并表明置换不变算法无法收敛到", "(Li,2014:Dean et al.,2012:Recht et al.,2011) :其中", "解决方案（Karimireddyetal.，2020）的任何预定义精度", "个服务器存储并更新模型参数，而所有其他服务器迭", "这是不够的", "代计算梯度，将其发送到PS并下载更新的参数。这种", "虽然有几种方法旨在规避这个问题，但它们中的大多", "策略对于少数工人来说可能是相当有效的；随着它的", "数都有很大的局限性，例如没有收敛分析（Chenetal.", "增加，参数服务器最终变得无法处理负载。虽然梯度压", "2018; Rajput et al., 2019; Rodriguez-Barroso et al., 2020", "缩 (Seideetal.,2014;Lin etal.,2018:Mishchenkoetal..", "Xu&Lyu.2020)，在分析（Alistarhetal.,2018:Allen-Zhu", "2019;Koloskova et al.,2020; Gorbunov et al.,2020; 2021)", "etal.,2021：Regattietal.,2020）中过于严格的假设，或", "ISource code for the experiments is available at", "者使用方差降低的估计器（Wuetal.,2020），这些估计", "https://github.com/yandex-research/btard", "器在深度学习应用程序中收敛缓慢（Defazio&Bottou", "2019）。唯一没有这种限制的论文是Karimireddyetal" ]

{ "arxiv_id": "2106.11257", "language": "zh", "page_idx": 1 }

[ [ "0.3733", "0.05936", "0.597", "0.05936", "0.597", "0.06946", "0.3733", "0.06946" ], [ "0.0907", "0.0884", "0.4722", "0.0884", "0.4722", "0.0985", "0.0907", "0.0985" ], [ "0.5", "0.0871", "0.8857", "0.0871", "0.8857", "0.101", "0.5", "0.101" ], [ "0.0907", "0.1016", "0.4739", "0.1016", "0.4739", "0.11554", "0.0907", "0.11554" ], [ "0.5", "0.1016", "0.6807", "0.1016", "0.6807", "0.11554", "0.5", "0.11554" ], [ "0.0907", "0.11743", "0.4722", "0.11743", "0.4722", "0.1307", "0.0907", "0.1307" ], [ "0.5", "0.125", "0.734", "0.125", "0.734", "0.1383", "0.5", "0.1383" ], [ "0.08905", "0.132", "0.4722", "0.132", "0.4722", "0.1459", "0.08905", "0.1459" ], [ "0.08905", "0.1477", "0.2378", "0.1477", "0.2378", "0.161", "0.08905", "0.161" ], [ "0.513", "0.1427", "0.5337", "0.1427", "0.5337", "0.1559", "0.513", "0.1559" ], [ "0.09235", "0.1718", "0.4673", "0.1718", "0.4673", "0.1818", "0.09235", "0.1818" ], [ "0.5", "0.1736", "0.8857", "0.1736", "0.8857", "0.1875", "0.5", "0.1875" ], [ "0.0907", "0.1849", "0.4722", "0.1849", "0.4722", "0.1989", "0.0907", "0.1989" ], [ "0.5107", "0.1895", "0.7197", "0.1895", "0.7197", "0.2026", "0.5107", "0.2026" ], [ "0.09235", "0.202", "0.4714", "0.202", "0.4714", "0.2122", "0.09235", "0.2122" ], [ "0.501", "0.202", "0.5186", "0.202", "0.5186", "0.2172", "0.501", "0.2172" ], [ "0.10455", "0.2166", "0.4714", "0.2166", "0.4714", "0.2267", "0.10455", "0.2267" ], [ "0.08905", "0.2299", "0.4739", "0.2299", "0.4739", "0.2437", "0.08905", "0.2437" ], [ "0.5", "0.2317", "0.884", "0.2317", "0.884", "0.2456", "0.5", "0.2456" ], [ "0.0907", "0.2468", "0.4722", "0.2468", "0.4722", "0.2568", "0.0907", "0.2568" ], [ "0.513", "0.2474", "0.83", "0.2474", "0.83", "0.262", "0.513", "0.262" ], [ "0.08905", "0.26", "0.4739", "0.26", "0.4739", "0.2734", "0.08905", "0.2734" ], [ "0.5024", "0.26", "0.5186", "0.26", "0.5186", "0.2695", "0.5024", "0.2695" ], [ "0.08905", "0.276", "0.4722", "0.276", "0.4722", "0.289", "0.08905", "0.289" ], [ "0.5", "0.2842", "0.884", "0.2842", "0.884", "0.2974", "0.5", "0.2974" ], [ "0.08826", "0.2903", "0.4739", "0.289", "0.4739", "0.303", "0.08826", "0.3037" ], [ "0.5156", "0.2998", "0.8857", "0.2998", "0.8857", "0.3132", "0.5156", "0.3132" ], [ "0.513", "0.3145", "0.7173", "0.3145", "0.7173", "0.3276", "0.513", "0.3276" ], [ "0.0907", "0.324", "0.335", "0.324", "0.335", "0.334", "0.0907", "0.334" ], [ "0.0907", "0.3452", "0.4739", "0.3452", "0.4739", "0.3591", "0.0907", "0.3591" ], [ "0.5", "0.348", "0.751", "0.348", "0.751", "0.361", "0.5", "0.361" ], [ "0.0907", "0.361", "0.4722", "0.361", "0.4722", "0.375", "0.0907", "0.375" ], [ "0.08905", "0.3757", "0.4739", "0.3757", "0.4739", "0.3894", "0.08905", "0.3894" ], [ "0.501", "0.3726", "0.884", "0.3726", "0.884", "0.3865", "0.501", "0.3865" ], [ "0.0907", "0.3914", "0.4739", "0.3914", "0.4739", "0.4048", "0.0907", "0.4048" ], [ "0.501", "0.387", "0.884", "0.387", "0.884", "0.401", "0.501", "0.401" ], [ "0.501", "0.4016", "0.8857", "0.4016", "0.8857", "0.4153", "0.501", "0.4153" ], [ "0.09235", "0.4077", "0.4714", "0.4077", "0.4714", "0.4185", "0.09235", "0.4185" ], [ "0.08826", "0.4216", "0.2197", "0.4216", "0.2197", "0.435", "0.08826", "0.435" ], [ "0.5024", "0.4172", "0.8857", "0.4172", "0.8857", "0.4312", "0.5024", "0.4312" ], [ "0.5024", "0.432", "0.8857", "0.432", "0.8857", "0.4458", "0.5024", "0.4458" ], [ "0.09314", "0.4458", "0.4722", "0.4458", "0.4722", "0.4558", "0.09314", "0.4558" ], [ "0.501", "0.4463", "0.884", "0.4463", "0.884", "0.4602", "0.501", "0.4602" ], [ "0.09235", "0.461", "0.4722", "0.461", "0.4722", "0.4717", "0.09235", "0.4717" ], [ "0.5", "0.4622", "0.884", "0.4622", "0.884", "0.476", "0.5", "0.476" ], [ "0.0907", "0.4749", "0.4722", "0.4749", "0.4722", "0.488", "0.0907", "0.488" ], [ "0.5", "0.476", "0.8857", "0.4766", "0.8857", "0.4905", "0.5", "0.4893" ], [ "0.09235", "0.4912", "0.4714", "0.4912", "0.4714", "0.502", "0.09235", "0.502" ], [ "0.5034", "0.4944", "0.8857", "0.4944", "0.8857", "0.505", "0.5034", "0.505" ], [ "0.08905", "0.504", "0.4739", "0.504", "0.4739", "0.5176", "0.08905", "0.5176" ], [ "0.5024", "0.5083", "0.8857", "0.5083", "0.8857", "0.5215", "0.5024", "0.5215" ], [ "0.09314", "0.521", "0.4722", "0.521", "0.4722", "0.531", "0.09314", "0.531" ], [ "0.5024", "0.5244", "0.8823", "0.5244", "0.8823", "0.535", "0.5024", "0.535" ], [ "0.0907", "0.536", "0.4722", "0.536", "0.4722", "0.547", "0.0907", "0.547" ], [ "0.5", "0.537", "0.626", "0.536", "0.626", "0.5503", "0.5", "0.552" ], [ "0.08905", "0.55", "0.10864", "0.55", "0.10864", "0.563", "0.08905", "0.563" ], [ "0.5", "0.56", "0.884", "0.56", "0.884", "0.5737", "0.5", "0.5737" ], [ "0.0907", "0.5728", "0.4739", "0.5728", "0.4739", "0.586", "0.0907", "0.586" ], [ "0.5", "0.5747", "0.8857", "0.5757", "0.8857", "0.5903", "0.5", "0.589" ], [ "0.09235", "0.589", "0.4722", "0.589", "0.4722", "0.5996", "0.09235", "0.5996" ], [ "0.501", "0.5903", "0.884", "0.5903", "0.884", "0.604", "0.501", "0.604" ], [ "0.09235", "0.604", "0.4722", "0.604", "0.4722", "0.6143", "0.09235", "0.6143" ], [ "0.501", "0.606", "0.8857", "0.606", "0.8857", "0.619", "0.501", "0.619" ], [ "0.09235", "0.619", "0.4714", "0.619", "0.4714", "0.63", "0.09235", "0.63" ], [ "0.501", "0.6206", "0.8857", "0.6206", "0.8857", "0.634", "0.501", "0.634" ], [ "0.08826", "0.634", "0.165", "0.634", "0.165", "0.6445", "0.08826", "0.6445" ], [ "0.5034", "0.637", "0.8823", "0.637", "0.8823", "0.648", "0.5034", "0.648" ], [ "0.09235", "0.656", "0.4739", "0.656", "0.4739", "0.6694", "0.09235", "0.6694" ], [ "0.5024", "0.6514", "0.8823", "0.6514", "0.8823", "0.662", "0.5024", "0.662" ], [ "0.09235", "0.672", "0.4722", "0.672", "0.4722", "0.6816", "0.09235", "0.6816" ], [ "0.5", "0.6665", "0.7935", "0.6665", "0.7935", "0.6797", "0.5", "0.6797" ], [ "0.09235", "0.6875", "0.4722", "0.6875", "0.4722", "0.6978", "0.09235", "0.6978" ], [ "0.5024", "0.6895", "0.8857", "0.6895", "0.8857", "0.7026", "0.5024", "0.7026" ], [ "0.0907", "0.7007", "0.4739", "0.7007", "0.4739", "0.714", "0.0907", "0.714" ], [ "0.5024", "0.704", "0.8857", "0.704", "0.8857", "0.7173", "0.5024", "0.7173" ], [ "0.09235", "0.7153", "0.4722", "0.7153", "0.4722", "0.729", "0.09235", "0.729" ], [ "0.5024", "0.72", "0.8823", "0.72", "0.8823", "0.731", "0.5024", "0.731" ], [ "0.0907", "0.731", "0.4739", "0.731", "0.4739", "0.744", "0.0907", "0.744" ], [ "0.501", "0.7344", "0.884", "0.7344", "0.884", "0.7476", "0.501", "0.7476" ], [ "0.0907", "0.7456", "0.4739", "0.7456", "0.4739", "0.759", "0.0907", "0.759" ], [ "0.5", "0.75", "0.8857", "0.7476", "0.8857", "0.761", "0.5", "0.763" ], [ "0.09235", "0.762", "0.183", "0.762", "0.183", "0.7734", "0.09235", "0.7734" ], [ "0.501", "0.7646", "0.8857", "0.7646", "0.8857", "0.778", "0.501", "0.778" ], [ "0.501", "0.78", "0.761", "0.78", "0.761", "0.7935", "0.501", "0.7935" ], [ "0.08905", "0.797", "0.1805", "0.797", "0.1805", "0.808", "0.08905", "0.808" ], [ "0.5215", "0.802", "0.8823", "0.8047", "0.8823", "0.8184", "0.5215", "0.815" ], [ "0.0907", "0.8213", "0.4739", "0.8213", "0.4739", "0.8354", "0.0907", "0.8354" ], [ "0.5034", "0.8184", "0.884", "0.8184", "0.884", "0.829", "0.5034", "0.829" ], [ "0.5024", "0.8306", "0.841", "0.8306", "0.841", "0.8413", "0.5024", "0.8413" ], [ "0.09314", "0.8384", "0.4722", "0.8384", "0.4722", "0.8486", "0.09314", "0.8486" ], [ "0.5215", "0.8423", "0.884", "0.845", "0.884", "0.856", "0.5215", "0.853" ], [ "0.08905", "0.853", "0.4739", "0.853", "0.4739", "0.866", "0.08905", "0.866" ], [ "0.501", "0.855", "0.8857", "0.855", "0.8857", "0.869", "0.501", "0.869" ], [ "0.0907", "0.867", "0.4739", "0.867", "0.4739", "0.881", "0.0907", "0.881" ], [ "0.5024", "0.869", "0.8813", "0.869", "0.8813", "0.88", "0.5024", "0.88" ], [ "0.09235", "0.8833", "0.4722", "0.8833", "0.4722", "0.893", "0.09235", "0.893" ], [ "0.501", "0.88", "0.8857", "0.88", "0.8857", "0.893", "0.501", "0.893" ], [ "0.501", "0.893", "0.88", "0.893", "0.88", "0.907", "0.501", "0.907" ] ]

[ "SecureDistributed Training at Scale", "（202O）：它提出了一种称为CENTEREDCLIP的新聚合", "据并行训练设置（如E2.1节中所述），其中对等方聚", "规则，将其应用于具有客户端动量的SGD，并在合理", "合其大小为d的梯度向量", "的假设下证明了所获得方法在非凸情况下的收敛结果。", "我们分几个阶段描述我们的策略", "我们在附录B.1.1中提供了有关基于拜占庭容忍PS的", "方法的更多详细信息", "px", "仅在儿篇论文中研究了拜占庭容忍优化的分散方法", ": Section 3.1 outlines our approach for B yzantine- T oler-", "不幸的是，已知的方法并不适合分布式深度学习，因", "ant A ll-Re d uce (BTARD).", "为它们要么依赖于全梯度计算（Yang&Bajwa，2019a;b）", "Bx", "or use redundantcommunications withmultiple servers", "(El-Mhamdietal.2020)，，要么在每个步骤（Guptaetal.", ": In Section 3.2 , we formulate the underlying optimiza", "2021：Gupta&Vaidya，2021）都需要全向量的对等通信", "tion problem and derive its convergence bounds.", "这是不可扩展的，或者提供了不如非并行SGDPeng", "px", "etal.，2021)的收敛保证，后者在现代深度学习任务上", ".In Section 3.3,we propose a heuristic for resisting Sybil", "的收敛速度非常慢。我们将更多细节推迟到附录B.1.2", "attacks, allowing our system to accept new untrusted", "peers midway through training", "2.3.Securityin distributed systems", "消息传播协议。在这项工作中，我们考虑分布式系统", "3.1. Byzantine-Tolerant All-Reduce", "完全依赖于对等连接(e。g。，通过互联网工作的人)。", "我们算法的几个关键阶段要求对等体向所有其他对等", "我们假设某些工作程序可能是恶意的，即。e。，他们可", "体广播小消息。为了保持一致性，如果至少有一个诚", "以arbitrarily偏离我们的算法：forinstance，发送任意向", "量而不是随机梯度或违反通信协议。这样的工人被称", "实的对等方收到消息，我们希望所有其他诚实的对等", "方最终也能收到它", "为拜占庭节点或只是拜占庭。我们假设它们是无所不", "知的（Karimireddyetal.，2020）（除了诚实节点的私钥", "一个天真的解决方案是让所有对等端将每个以前看不", "和MPRNG的内部）and ableto collude with eachother", "见的消息中继到所有其他对等端。在这种情况下，对", "。我们将所有“好”工人的集合表示为9，将拜占庭", "于n对等体和6位消息，一个所有广播将要求每个", "工人的集合表示为B。我们进一步假设B在整个优", "对等体传输O(n-6）数据。为了提高效率，我们使用", "化过程中是固定的，并且不到一半的节点是拜占庭的", "GossipSubE（Vyzovitisetal.，2020），通过将每条消息仅", "B≤8n，其中E[0,1/2)。最后，我们假设所有工作", "中继到D精心选择的邻居，从而将每个对等点的成本", "人员都可以访问定义自标函数的数据，从完整数据集", "降低到O(nb）数据，其中D是根据延迟要求选择的常", "中采样小批量。2", "数", "We design our algorithm in such a way that all types of", "数字签名。我们的方法依赖于这样一个事实，即攻击", "Byzantine faults have limited effect and chance of being", "者不能冒充诚实的对等方或更改诚实的对等方广播的", "discovered.为了限制单个SGD步骤的伤害，我们使用", "消息。为了实现这一点，我们要求所有对等方声明他们", "称为CENTEREDCLIP（Karimireddyetal.，2020）的强大", "的公钥，并使用数字签名对所有消息进行签名（Rivest", "聚合技术修改ButterflyAll-Reduce3E（见图E1）。我", "et al..1978", "们将CENTEREDCLIP应用于梯度向量的每个分区，而", "多方随机数生成器。为了确保对等体诚实地计算梯度，", "不是朴素平均。我们将此过程表示为BUTTERFLYCLIF", "我们的方法验证了所有计算梯度的随机子集。因此：我", "（有关其正式描述，请参见AlgorithmE2）", "们需要选择谁将被检查，以使攻击者既不能预测也不", "但是，拜占庭同行可以通过多次迭代攻击来规避此", "能影响随机抽取。这可以通过基于Blum（1983）的抛硬", "限制。为了防止这种情况，BTARD会定期选择随机", "币协议的多方随机数生成器（MPRNG）来完成。我们", "对等体作为验证器。验证器必须重新计算其他对等体", "在附录B.2中解释了完整的算法。该算法要求每个对", "的梯度并报告任何差异insteadofcomputingtheirown", "等体仅广播3个标量，因此其通信成本为每个对等体", "gradients.由于此类测试仅在攻击者无法预测它们是否", "的O(n）数据", "将被验证时才有效，因此我们使用多方随机数生成器", "（如E2.3节所述）选择验证的对等体", "3. Method", "2He et al. (2020) show that it is impossible to achieve any pre", "我们考虑在公共数据集上进行安全的分布式训练，其", "defined accuracy of the solution without this assumption, i.e., in", "the heterogeneous case (see discussion in Appendix F.2)", "中每个对等体都可以访问整个训练数据andcommuni", "3WechooseButterflyAll-Reducesothatpeersaggregatenon", "catewithanyotherpeer。在这种情况下，多方通过将其", "overlapping parts of the gradient vector. This helps to identify the", "计算资源组合在一起进行一次大规模训练运行来进行", "attacker if the gradients are aggregated incorrectly.Jiang et al", "合作。具体来说，我们考虑使用All-ReduceSGD的数", "(2020) report that Butterfly All-Reduce is near-optimal for dis-", "tributed training over high-latency networks such as the Internet" ]

{ "arxiv_id": "2106.11257", "language": "zh", "page_idx": 2 }

[ [ "0.3743", "0.05936", "0.5986", "0.05936", "0.5986", "0.06946", "0.3743", "0.06946" ], [ "0.1396", "0.0947", "0.2345", "0.0947", "0.2345", "0.11237", "0.1396", "0.11237" ], [ "0.4697", "0.09534", "0.5513", "0.09534", "0.5513", "0.1061", "0.4697", "0.1061" ], [ "0.7646", "0.0966", "0.814", "0.0966", "0.814", "0.1067", "0.7646", "0.1067" ], [ "0.3513", "0.1067", "0.361", "0.1067", "0.361", "0.1111", "0.3513", "0.1111" ], [ "0.7607", "0.1048", "0.8164", "0.1067", "0.8164", "0.11743", "0.7607", "0.11554" ], [ "0.7573", "0.11554", "0.8203", "0.11554", "0.8203", "0.1256", "0.7573", "0.1256" ], [ "0.3489", "0.1092", "0.3635", "0.1092", "0.3635", "0.1332", "0.3489", "0.1332" ], [ "0.2573", "0.125", "0.3162", "0.125", "0.3162", "0.1351", "0.2573", "0.1351" ], [ "0.7563", "0.125", "0.822", "0.125", "0.822", "0.1364", "0.7563", "0.1364" ], [ "0.6665", "0.132", "0.6978", "0.132", "0.6978", "0.1364", "0.6665", "0.1364" ], [ "0.2566", "0.1338", "0.3203", "0.1364", "0.3203", "0.1509", "0.2566", "0.1483" ], [ "0.1111", "0.13", "0.232", "0.13", "0.232", "0.1445", "0.1111", "0.1445" ], [ "0.6177", "0.1364", "0.6274", "0.1364", "0.6274", "0.1395", "0.6177", "0.1395" ], [ "0.6587", "0.1395", "0.726", "0.1395", "0.726", "0.1497", "0.6587", "0.1497" ], [ "0.7695", "0.1338", "0.806", "0.1338", "0.806", "0.1445", "0.7695", "0.1445" ], [ "0.3489", "0.1433", "0.357", "0.1401", "0.361", "0.1465", "0.353", "0.1497" ], [ "0.3506", "0.1497", "0.361", "0.1497", "0.361", "0.1635", "0.3506", "0.1635" ], [ "0.759", "0.1445", "0.819", "0.1445", "0.819", "0.1547", "0.759", "0.1547" ], [ "0.11273", "0.1635", "0.2336", "0.1604", "0.2336", "0.178", "0.11273", "0.1812" ], [ "0.4968", "0.1635", "0.546", "0.1654", "0.546", "0.1755", "0.4968", "0.1736" ], [ "0.7754", "0.1622", "0.8013", "0.1622", "0.8013", "0.1736", "0.7754", "0.1736" ], [ "0.1152", "0.1686", "0.11926", "0.1686", "0.11926", "0.1749", "0.1152", "0.1749" ], [ "0.5073", "0.1749", "0.5376", "0.1749", "0.5376", "0.1831", "0.5073", "0.1831" ], [ "0.1111", "0.1976", "0.1323", "0.1976", "0.1323", "0.2108", "0.1111", "0.2108" ], [ "0.2646", "0.1989", "0.705", "0.1963", "0.705", "0.2108", "0.2646", "0.2134" ], [ "0.0907", "0.2279", "0.8857", "0.2279", "0.8857", "0.2412", "0.0907", "0.2412" ], [ "0.08905", "0.2412", "0.853", "0.2412", "0.853", "0.2551", "0.08905", "0.2551" ], [ "0.08905", "0.2683", "0.4756", "0.2683", "0.4756", "0.2822", "0.08905", "0.2822" ], [ "0.501", "0.2683", "0.886", "0.2683", "0.886", "0.2822", "0.501", "0.2822" ], [ "0.08826", "0.2827", "0.4756", "0.2822", "0.4756", "0.2966", "0.08826", "0.2974" ], [ "0.5034", "0.2854", "0.8823", "0.2854", "0.8823", "0.2954", "0.5034", "0.2954" ], [ "0.08905", "0.2986", "0.4739", "0.2986", "0.4739", "0.3125", "0.08905", "0.3125" ], [ "0.5", "0.2974", "0.884", "0.2974", "0.884", "0.3113", "0.5", "0.3113" ], [ "0.0907", "0.3145", "0.4722", "0.3145", "0.4722", "0.3276", "0.0907", "0.3276" ], [ "0.501", "0.3132", "0.884", "0.3132", "0.884", "0.327", "0.501", "0.327" ], [ "0.0907", "0.3289", "0.286", "0.3289", "0.286", "0.3423", "0.0907", "0.3423" ], [ "0.501", "0.327", "0.884", "0.3276", "0.884", "0.3423", "0.501", "0.3416" ], [ "0.5024", "0.3435", "0.884", "0.3435", "0.884", "0.3567", "0.5024", "0.3567" ], [ "0.09235", "0.3516", "0.4722", "0.3516", "0.4722", "0.3655", "0.09235", "0.3655" ], [ "0.501", "0.3591", "0.884", "0.3591", "0.884", "0.3726", "0.501", "0.3726" ], [ "0.0907", "0.3662", "0.4722", "0.3662", "0.4722", "0.3801", "0.0907", "0.3801" ], [ "0.0907", "0.3806", "0.4697", "0.3806", "0.4697", "0.3945", "0.0907", "0.3945" ], [ "0.501", "0.375", "0.8857", "0.375", "0.8857", "0.3882", "0.501", "0.3882" ], [ "0.501", "0.3882", "0.884", "0.3882", "0.884", "0.4016", "0.501", "0.4016" ], [ "0.0907", "0.3965", "0.4722", "0.3965", "0.4722", "0.4104", "0.0907", "0.4104" ], [ "0.501", "0.404", "0.884", "0.404", "0.884", "0.4172", "0.501", "0.4172" ], [ "0.0907", "0.411", "0.4722", "0.411", "0.4722", "0.4248", "0.0907", "0.4248" ], [ "0.5024", "0.4185", "0.884", "0.4185", "0.884", "0.432", "0.5024", "0.432" ], [ "0.08905", "0.4268", "0.4722", "0.4268", "0.4722", "0.44", "0.08905", "0.44" ], [ "0.0907", "0.44", "0.4722", "0.4414", "0.4722", "0.4558", "0.0907", "0.455" ], [ "0.5024", "0.4343", "0.8823", "0.4343", "0.8823", "0.4475", "0.5024", "0.4475" ], [ "0.08905", "0.455", "0.3718", "0.4558", "0.3718", "0.4702", "0.08905", "0.4697" ], [ "0.5024", "0.4495", "0.8823", "0.4495", "0.8823", "0.4634", "0.5024", "0.4634" ], [ "0.08905", "0.4778", "0.4697", "0.4792", "0.4697", "0.4937", "0.08905", "0.4924" ], [ "0.5024", "0.473", "0.8667", "0.473", "0.8667", "0.486", "0.5024", "0.486" ], [ "0.0907", "0.4944", "0.4722", "0.4944", "0.4722", "0.5083", "0.0907", "0.5083" ], [ "0.5", "0.502", "0.6846", "0.5024", "0.6846", "0.5176", "0.5", "0.5166" ], [ "0.0907", "0.5103", "0.4722", "0.5103", "0.4722", "0.5234", "0.0907", "0.5234" ], [ "0.09314", "0.5264", "0.4714", "0.5264", "0.4714", "0.537", "0.09314", "0.537" ], [ "0.5024", "0.528", "0.8857", "0.528", "0.8857", "0.5415", "0.5024", "0.5415" ], [ "0.09235", "0.5405", "0.4722", "0.5405", "0.4722", "0.5537", "0.09235", "0.5537" ], [ "0.5024", "0.5435", "0.5254", "0.5435", "0.5254", "0.5547", "0.5024", "0.5547" ], [ "0.0907", "0.5537", "0.4714", "0.5537", "0.4714", "0.5684", "0.0907", "0.5684" ], [ "0.5884", "0.5596", "0.806", "0.5596", "0.806", "0.5728", "0.5884", "0.5728" ], [ "0.864", "0.558", "0.886", "0.558", "0.886", "0.5713", "0.864", "0.5713" ], [ "0.0907", "0.5693", "0.3594", "0.5693", "0.3594", "0.584", "0.0907", "0.584" ], [ "0.58", "0.5693", "0.6274", "0.5693", "0.6274", "0.58", "0.58", "0.58" ], [ "0.5024", "0.5903", "0.884", "0.5903", "0.884", "0.604", "0.5024", "0.604" ], [ "0.1544", "0.6016", "0.2778", "0.6016", "0.2778", "0.627", "0.1544", "0.627" ], [ "0.165", "0.5986", "0.1772", "0.5986", "0.1772", "0.605", "0.165", "0.605" ], [ "0.452", "0.6094", "0.4739", "0.6094", "0.4739", "0.622", "0.452", "0.622" ], [ "0.5024", "0.605", "0.8857", "0.605", "0.8857", "0.619", "0.5024", "0.619" ], [ "0.2974", "0.6187", "0.3708", "0.6187", "0.3708", "0.63", "0.2974", "0.63" ], [ "0.3823", "0.61", "0.4004", "0.61", "0.4004", "0.6187", "0.3823", "0.6187" ], [ "0.1609", "0.6255", "0.1814", "0.6255", "0.1814", "0.636", "0.1609", "0.636" ], [ "0.5024", "0.6206", "0.8823", "0.6206", "0.8823", "0.634", "0.5024", "0.634" ], [ "0.5024", "0.6353", "0.884", "0.6353", "0.884", "0.6484", "0.5024", "0.6484" ], [ "0.0907", "0.6484", "0.4722", "0.6484", "0.4722", "0.662", "0.0907", "0.662" ], [ "0.5024", "0.651", "0.8857", "0.651", "0.8857", "0.664", "0.5024", "0.664" ], [ "0.0907", "0.6636", "0.4739", "0.6636", "0.4739", "0.677", "0.0907", "0.677" ], [ "0.501", "0.6655", "0.732", "0.664", "0.732", "0.678", "0.501", "0.6787" ], [ "0.09235", "0.6797", "0.1396", "0.6797", "0.1396", "0.6914", "0.09235", "0.6914" ], [ "0.5024", "0.688", "0.8857", "0.688", "0.8857", "0.702", "0.5024", "0.702" ], [ "0.09235", "0.7026", "0.4714", "0.7026", "0.4714", "0.7134", "0.09235", "0.7134" ], [ "0.5024", "0.7026", "0.884", "0.7026", "0.884", "0.7163", "0.5024", "0.7163" ], [ "0.0907", "0.7163", "0.4714", "0.7163", "0.4714", "0.73", "0.0907", "0.73" ], [ "0.5024", "0.7183", "0.884", "0.7183", "0.884", "0.732", "0.5024", "0.732" ], [ "0.0907", "0.733", "0.1929", "0.733", "0.1929", "0.744", "0.0907", "0.744" ], [ "0.5034", "0.7344", "0.8857", "0.7344", "0.8857", "0.7476", "0.5034", "0.7476" ], [ "0.5034", "0.7485", "0.8857", "0.7485", "0.8857", "0.762", "0.5034", "0.762" ], [ "0.1854", "0.7646", "0.3096", "0.7646", "0.3096", "0.782", "0.1854", "0.782" ], [ "0.5", "0.761", "0.884", "0.763", "0.884", "0.78", "0.5", "0.778" ], [ "0.1544", "0.7676", "0.1814", "0.7676", "0.1814", "0.78", "0.1544", "0.78" ], [ "0.452", "0.7676", "0.4739", "0.7676", "0.4739", "0.78", "0.452", "0.78" ], [ "0.3228", "0.7764", "0.3816", "0.7764", "0.3816", "0.7886", "0.3228", "0.7886" ], [ "0.5034", "0.779", "0.8857", "0.779", "0.8857", "0.7925", "0.5034", "0.7925" ], [ "0.5034", "0.7944", "0.6177", "0.7944", "0.6177", "0.808", "0.5034", "0.808" ], [ "0.0907", "0.8047", "0.339", "0.8047", "0.339", "0.815", "0.0907", "0.815" ], [ "0.3406", "0.803", "0.4722", "0.802", "0.4722", "0.8164", "0.3406", "0.8174" ], [ "0.08905", "0.8164", "0.4739", "0.8174", "0.4739", "0.832", "0.08905", "0.8306" ], [ "0.501", "0.815", "0.8857", "0.8164", "0.8857", "0.8306", "0.501", "0.83" ], [ "0.5024", "0.8306", "0.8857", "0.8306", "0.8857", "0.845", "0.5024", "0.845" ], [ "0.10864", "0.839", "0.4739", "0.8413", "0.4739", "0.856", "0.10864", "0.854" ], [ "0.5024", "0.8467", "0.8823", "0.8467", "0.8823", "0.86", "0.5024", "0.86" ], [ "0.08905", "0.853", "0.4722", "0.854", "0.4722", "0.8687", "0.08905", "0.867" ], [ "0.08905", "0.866", "0.4739", "0.866", "0.4739", "0.881", "0.08905", "0.881" ], [ "0.501", "0.8613", "0.8857", "0.8613", "0.8857", "0.875", "0.501", "0.875" ], [ "0.08905", "0.88", "0.4722", "0.88", "0.4722", "0.893", "0.08905", "0.893" ], [ "0.5024", "0.877", "0.884", "0.877", "0.884", "0.89", "0.5024", "0.89" ], [ "0.08826", "0.8926", "0.4617", "0.8916", "0.4617", "0.905", "0.08826", "0.906" ], [ "0.5024", "0.8926", "0.8857", "0.8926", "0.8857", "0.906", "0.5024", "0.906" ] ]

[ "Secure Distributed Training at Scale", "91=(as)", "CENTEREDCLIP", "Verify that", "R", "CenteredClip", "was performed", "DAR", "BROADCAST", "correctly using", "Sonor", "hash(g[3])", "O 92 = V(a,$)", "", "withMPRNG", "random", "1", "E", "projections to", "O-g=$）", "Outliers are", "z(rt)", "", "clipped", "O", "Peer i validates g, from the previous step, i and j are randomly chosen using rt-l", "Figure 2. A scheme illustrating one step of Byzantine-Tolerant All-Reduce- a part of Algorithm 1 executed between the consecutive", "SGD steps. Here, t is the step number, at is the model weights, and st is a publicly known random seed for sampling a minibatch.", "在每个训练步骤之后，对等方使用MPRNG选择m", "聚合结果的哈希（Alg的第6行。2）在他们学习", "validators and m peers to validate (each validator checks", "之前。这确保了恶意聚合器不能以与之正交的方式修", "onepeer).拜占庭人在遭受攻击之前无法预测“安全”", "改结果（此攻击向量和更复杂的攻击向量在附录D", "迭代。因此，更频繁的攻击（具有更大的总伤害）更有", "和E.5中进行了分析)。我们将所有这些过程组合在算", "可能被诚实的验证者检测到", "法E1中（请参见图E2中的方案及其在Alg中的详细", "版本。6-7）。至关重要的是，算法E1的一个步骤要", "由于验证程序也可能是恶意的，因此BTARD使用特", "求每个对等方（a)E接收和发送梯度向量的n分区（与", "殊的ACCUSE过程来检测虚假报告。在每个平均回合", "之前，对等方使用GossipSub4（算法E2的第2行）广", "ButterflyAll-Reduce中完全相同），（b）E广播O(n）标", "量值（所有哈希，内积s和必要的指控），以及（c)F", "播其梯度的哈希值。然后，如果验证器i指责对等方", "运行MPRNG一次。根据E2.1和2.3节，这些过程的", "i修改梯度，则所有其他对等方将能够重新计算i的", "总通信成本为每个对等方的O(d+n2）数据。这接近带", "梯度并将其哈希与广播的。如果对等体发现i的渐变", "是正确的，对等体i将被禁止（Hammurabi&Harper，", "宽最优版本All-Reduce的O(d)成本：对于受益于分布", "1904）。此过程在Algorithm3中有描述", "式训练的模型，O(n2)的额外成本通常比O(d)小得多", "生成的算法对通过不正确的梯度进行的攻击具有弹性", "该算法的同步点和计算开销在附录C中进行了回顾", "但是，恶意对等体也可能通过违反CENTEREDCLIP过", "3.2. Convergence analysis", "程来聚合它们正在聚合的梯度部分而损害训练。幸运", "的是，我们可以设计一个测试，通过该测试，对等方", "从优化理论的角度来看，我们的任务是期望最小化问", "可以验证他们收到的向量确实是CENTEREDCLIP的输", "题：", "出。我们需要将CENTEREDCLIP视为方程式的定点迭", "min(f(α) :=E~D [f(a,)])", "(3)", "代（请参阅附录EE.2中的详细信息）：", "TEQCR", "在这里，目标函数f是光滑且一致的下界，QRd是", "(g: - a) min ", "n", "(1)", "可容许参数的封闭凸集，ε是随机性的来源，例如小批", "g-", "0", "2=1", "量索引。我们假设问题（3）可以以分布式的方式解决", "即。e。，可以使用nworker并行计算（迷你批处理）随", "工人无法测试（1）是否直接成立，因为收集9将需", "机梯度，并根据某种协议进行通信。我们将工人集表", "要每个对等端发送O(dn）数据，从而破坏了我们算法", "示为[n]:={1,2,...,n]=gUB", "的目的", "拜占庭人有很多方法可以影响培训。我们可以将它们", "相反，工作人员应使用MPRNG输出在模型梯度空间", "全部分为四类：（a）E梯度攻击，其中拜占庭人修改", "中对随机方向之进行采样。然后，每个对等体计算并", "其9，但行为正常；（b)E聚合攻击，其中恶意聚合", "广播内积（2）：", "器返回错误的9i，并依靠其他人通过误报si来掩盖", "它；（c)E声誉攻击，例如通过虚假AcCUSE(ij）进", "z,(gi -)min", "行的诽谤；以及（4)E违反协议，。g。，拒绝在预定义", "Si=", "(2)", "1g-", "的超时时间内发送数据)。我们在附录D中详细说明", "了每种攻击类型", "最后，所有对等体都可以验证", "18i=0。与我们", "以前使用的MPRNG类似，所有聚合器都必须广播其", "出于此分析的目的，可以使用额外的策略来击退后两", "种攻击，该策略允许活动的工作者ELIMINATE任何其", "4we assume that peers declare their public key when joining", "他工作者，但代价是也被禁止。如果对等体i遇到来自", "and sign all broadcasted messages with the respective private key.", "Any peer broadcasting contradicting messages (e.g., different gra-", "对等体i的协议冲突，它将广播一条消息，要求从训", "dient hashes) should be banned, since it could break the eventual", "练中删除两个对等体i和j。此策略的设计确保每个", "consistency (other peers may receive them in a different order).", "此类消息，无论是由诚实的还是拜占庭的对等方发送" ]

{ "arxiv_id": "2106.11257", "language": "zh", "page_idx": 3 }

[ [ "0.1422", "0.11237", "0.831", "0.113", "0.831", "0.1307", "0.1422", "0.13" ], [ "0.3203", "0.1768", "0.649", "0.1755", "0.649", "0.1932", "0.3203", "0.1945" ], [ "0.08826", "0.1945", "0.1544", "0.1945", "0.1544", "0.209", "0.08826", "0.209" ], [ "0.2427", "0.2052", "0.3203", "0.207", "0.3203", "0.2247", "0.2427", "0.2229" ], [ "0.509", "0.2134", "0.638", "0.214", "0.638", "0.226", "0.509", "0.2247" ], [ "0.697", "0.2235", "0.7466", "0.226", "0.7466", "0.2374", "0.697", "0.2349" ], [ "0.6465", "0.2299", "0.6807", "0.2299", "0.6807", "0.2412", "0.6465", "0.2412" ], [ "0.11926", "0.2361", "0.4412", "0.2349", "0.4412", "0.25", "0.11926", "0.2507" ], [ "0.7983", "0.2343", "0.8716", "0.2343", "0.8716", "0.2412", "0.7983", "0.2412" ], [ "0.1209", "0.252", "0.4412", "0.252", "0.4412", "0.2651", "0.1209", "0.2651" ], [ "0.83", "0.2468", "0.8364", "0.2468", "0.8364", "0.25", "0.83", "0.25" ], [ "0.759", "0.26", "0.786", "0.26", "0.786", "0.2683", "0.759", "0.2683" ], [ "0.1209", "0.2664", "0.4412", "0.2664", "0.4412", "0.281", "0.1209", "0.281" ], [ "0.1209", "0.2798", "0.4395", "0.2798", "0.4395", "0.2935", "0.1209", "0.2935" ], [ "0.63", "0.2942", "0.652", "0.2942", "0.652", "0.2974", "0.63", "0.2974" ], [ "0.661", "0.2942", "0.6763", "0.2942", "0.6763", "0.2974", "0.661", "0.2974" ], [ "0.6895", "0.2986", "0.761", "0.2986", "0.761", "0.31", "0.6895", "0.31" ], [ "0.626", "0.3018", "0.6665", "0.3018", "0.6665", "0.3125", "0.626", "0.3125" ], [ "0.1209", "0.2966", "0.4412", "0.2966", "0.4412", "0.31", "0.1209", "0.31" ], [ "0.1209", "0.31", "0.4412", "0.3113", "0.4412", "0.3257", "0.1209", "0.3245" ], [ "0.5146", "0.31", "0.652", "0.308", "0.652", "0.3257", "0.5146", "0.3276" ], [ "0.734", "0.3113", "0.8364", "0.3113", "0.8364", "0.3416", "0.734", "0.3416" ], [ "0.1217", "0.327", "0.4429", "0.327", "0.4429", "0.3403", "0.1217", "0.3403" ], [ "0.5073", "0.334", "0.5864", "0.334", "0.5864", "0.3447", "0.5073", "0.3447" ], [ "0.613", "0.3352", "0.665", "0.3352", "0.665", "0.3423", "0.613", "0.3423" ], [ "0.12335", "0.3435", "0.4395", "0.3435", "0.4395", "0.3542", "0.12335", "0.3542" ], [ "0.1209", "0.3567", "0.4412", "0.3567", "0.4412", "0.3706", "0.1209", "0.3706" ], [ "0.636", "0.3516", "0.697", "0.3516", "0.697", "0.3623", "0.636", "0.3623" ], [ "0.5186", "0.361", "0.589", "0.3599", "0.589", "0.3687", "0.5186", "0.3694" ], [ "0.5986", "0.3599", "0.6323", "0.3623", "0.6313", "0.3726", "0.597", "0.3706" ], [ "0.1209", "0.3718", "0.4429", "0.3718", "0.4429", "0.3865", "0.1209", "0.3865" ], [ "0.7715", "0.3757", "0.806", "0.3655", "0.822", "0.3984", "0.786", "0.4077" ], [ "0.1209", "0.387", "0.4429", "0.387", "0.4429", "0.401", "0.1209", "0.401" ], [ "0.709", "0.3865", "0.739", "0.3726", "0.7524", "0.3894", "0.721", "0.4028" ], [ "0.11926", "0.4016", "0.4412", "0.401", "0.4412", "0.4153", "0.11926", "0.4167" ], [ "0.591", "0.411", "0.636", "0.411", "0.636", "0.4216", "0.591", "0.4216" ], [ "0.7295", "0.4092", "0.8384", "0.4092", "0.8384", "0.4192", "0.7295", "0.4192" ], [ "0.1209", "0.4172", "0.4412", "0.4172", "0.4412", "0.4312", "0.1209", "0.4312" ], [ "0.835", "0.4124", "0.859", "0.4124", "0.859", "0.4153", "0.835", "0.4153" ], [ "0.759", "0.4192", "0.8423", "0.4192", "0.8423", "0.43", "0.759", "0.43" ], [ "0.1209", "0.433", "0.4412", "0.433", "0.4412", "0.4463", "0.1209", "0.4463" ], [ "0.648", "0.4312", "0.687", "0.4312", "0.687", "0.435", "0.648", "0.435" ], [ "0.1209", "0.4475", "0.4429", "0.4475", "0.4429", "0.461", "0.1209", "0.461" ], [ "0.11926", "0.461", "0.4412", "0.4622", "0.4412", "0.4766", "0.11926", "0.476" ], [ "0.1217", "0.4778", "0.4412", "0.4778", "0.4412", "0.4912", "0.1217", "0.4912" ], [ "0.501", "0.473", "0.8857", "0.473", "0.8857", "0.486", "0.501", "0.486" ], [ "0.11926", "0.4924", "0.4412", "0.4912", "0.4412", "0.506", "0.11926", "0.507" ], [ "0.501", "0.488", "0.8857", "0.488", "0.8857", "0.4988", "0.501", "0.4988" ], [ "0.5", "0.502", "0.8857", "0.502", "0.8857", "0.515", "0.5", "0.515" ], [ "0.1209", "0.5083", "0.4412", "0.5083", "0.4412", "0.5215", "0.1209", "0.5215" ], [ "0.5", "0.515", "0.884", "0.515", "0.884", "0.529", "0.5", "0.529" ], [ "0.11926", "0.523", "0.4412", "0.5215", "0.4412", "0.536", "0.11926", "0.537" ], [ "0.4983", "0.528", "0.8857", "0.529", "0.8857", "0.5425", "0.4983", "0.5415" ], [ "0.1217", "0.5386", "0.4412", "0.5386", "0.4412", "0.552", "0.1217", "0.552" ], [ "0.501", "0.5444", "0.5474", "0.5444", "0.5474", "0.556", "0.501", "0.556" ], [ "0.1217", "0.553", "0.4429", "0.553", "0.4429", "0.5664", "0.1217", "0.5664" ], [ "0.1217", "0.5684", "0.4412", "0.5684", "0.4412", "0.582", "0.1217", "0.582" ], [ "0.11926", "0.5835", "0.4412", "0.582", "0.4412", "0.5967", "0.11926", "0.5977" ], [ "0.5024", "0.582", "0.8857", "0.582", "0.8857", "0.595", "0.5024", "0.595" ], [ "0.1209", "0.5986", "0.4412", "0.5986", "0.4412", "0.6123", "0.1209", "0.6123" ], [ "0.5024", "0.5977", "0.7793", "0.5977", "0.7793", "0.6113", "0.5024", "0.6113" ], [ "0.11926", "0.6123", "0.4412", "0.613", "0.4412", "0.6274", "0.11926", "0.627" ], [ "0.501", "0.6187", "0.884", "0.619", "0.884", "0.634", "0.501", "0.6333" ], [ "0.1217", "0.629", "0.4412", "0.629", "0.4412", "0.6426", "0.1217", "0.6426" ], [ "0.5034", "0.6353", "0.8857", "0.6353", "0.8857", "0.6484", "0.5034", "0.6484" ], [ "0.1209", "0.643", "0.4429", "0.643", "0.4429", "0.657", "0.1209", "0.657" ], [ "0.5034", "0.651", "0.884", "0.651", "0.884", "0.664", "0.5034", "0.664" ], [ "0.1217", "0.66", "0.2263", "0.66", "0.2263", "0.6724", "0.1217", "0.6724" ], [ "0.501", "0.664", "0.88", "0.664", "0.88", "0.678", "0.501", "0.678" ], [ "0.501", "0.6787", "0.884", "0.6797", "0.884", "0.6943", "0.501", "0.6934" ], [ "0.5", "0.6943", "0.884", "0.6943", "0.884", "0.7085", "0.5", "0.7085" ], [ "0.08826", "0.7085", "0.2239", "0.709", "0.2239", "0.7236", "0.08826", "0.7227" ], [ "0.501", "0.709", "0.709", "0.7104", "0.709", "0.7246", "0.501", "0.7236" ], [ "0.0907", "0.735", "0.4739", "0.735", "0.4739", "0.7485", "0.0907", "0.7485" ], [ "0.501", "0.732", "0.884", "0.733", "0.884", "0.7476", "0.501", "0.747" ], [ "0.08905", "0.75", "0.4739", "0.75", "0.4739", "0.763", "0.08905", "0.763" ], [ "0.5024", "0.7485", "0.884", "0.7485", "0.884", "0.762", "0.5024", "0.762" ], [ "0.0907", "0.765", "0.4739", "0.765", "0.4739", "0.779", "0.0907", "0.779" ], [ "0.5", "0.762", "0.8857", "0.763", "0.8857", "0.778", "0.5", "0.7764" ], [ "0.0907", "0.78", "0.4739", "0.78", "0.4739", "0.7935", "0.0907", "0.7935" ], [ "0.5024", "0.779", "0.8857", "0.779", "0.8857", "0.7925", "0.5024", "0.7925" ], [ "0.5024", "0.7925", "0.875", "0.7935", "0.875", "0.808", "0.5024", "0.8066" ], [ "0.1323", "0.8037", "0.4739", "0.8037", "0.4739", "0.8174", "0.1323", "0.8174" ], [ "0.08826", "0.8164", "0.1528", "0.8184", "0.1528", "0.83", "0.08826", "0.8276" ], [ "0.1814", "0.8174", "0.4739", "0.8174", "0.4739", "0.8306", "0.1814", "0.8306" ], [ "0.5024", "0.8164", "0.8857", "0.8164", "0.8857", "0.83", "0.5024", "0.83" ], [ "0.08826", "0.829", "0.4739", "0.829", "0.4739", "0.8423", "0.08826", "0.8423" ], [ "0.501", "0.8306", "0.8857", "0.8306", "0.8857", "0.8447", "0.501", "0.8447" ], [ "0.0866", "0.8413", "0.4265", "0.8413", "0.4265", "0.855", "0.0866", "0.855" ], [ "0.5024", "0.8467", "0.8857", "0.8467", "0.8857", "0.86", "0.5024", "0.86" ], [ "0.08826", "0.8657", "0.4756", "0.8657", "0.4756", "0.88", "0.08826", "0.88" ], [ "0.5024", "0.8613", "0.8857", "0.8613", "0.8857", "0.875", "0.5024", "0.875" ], [ "0.08905", "0.88", "0.4739", "0.88", "0.4739", "0.893", "0.08905", "0.893" ], [ "0.5", "0.875", "0.884", "0.8755", "0.884", "0.89", "0.5", "0.8896" ], [ "0.08826", "0.8926", "0.3096", "0.8926", "0.3096", "0.906", "0.08826", "0.906" ], [ "0.501", "0.89", "0.8857", "0.8916", "0.8857", "0.906", "0.501", "0.905" ], [ "0.482", "0.925", "0.4927", "0.925", "0.4927", "0.9346", "0.482", "0.9346" ] ]

[ "Protecting Language Generation Models via Invisible Watermarking", "Xuandong Zhao1 Yu-Xiang Wang 1 Lei Lil", "???。3in", "Abstract", "(a) Process of watermarking", "Rawinput", "Query", "Language generation models have been an in-", "ExtractedMode", "creasingly powerful enabler for many appli-", "", "Train", "cations. Many such models offer free or af-", "fordable API access, which makes them po-", "te", "r", "Generated text", "response", "tentially vulnerable to model extraction at-", "tacks through distillation. To protect intel-", " Victim Model API", "Adversary", "lectual property (IP) and ensure fair use of", "(b) Process of wa", "rkdetectior", "these models, various techniques such as lex", "ical watermarking and synonym replacement", "Probing input", "Suspect Model", "Query", "have been proposed. However, these methods", "", "can be nullified by obvious countermeasures", "", "such as “synonym randomization\" To ad", "Response", "Does the suspect.model", "dress this issue, we propose Ginsew, a novel", "Levtr", "the victim model?", "method to protect text generation models", "oeOut", "from being stolen through distillation. The", "key idea of our method is to inject secret sig", "nals into the probability vector of the decod-", "Figure 1. Overview of the process of watermarking and", "ing steps for each target token. We can then", "the process of watermark detection. The victim model API", "embeds watermarks in the response to input queries from", "detect the secret message by probing a sus-", "the adversaries. The APIowner can then use a key to verify", "pect model to tell if it is distilled from the pro-", "if the suspect model has been distilled from the victim", "tected one. Experimental results show that", "model.", "Ginsew can effectively identify instances of", "IP infringement with minimal impact on the", "generation quality of protected APIs. Our", "和维护成本。虽然许多应用程序受益于这些API，但", "method demonstrates an absolute improve-", "有些应用程序正在寻找更便宜的替代品", "ment of 19 to 29 points on mean average pre-", "虽然健康的竞争有利于防止LLM服务市场的垒断，但", "cision (mAP) in detecting suspects compared", "价格合理的API服务可能会暴露出一条“捷径”，使", "to previous methods against watermark re-", "公司能够以低得多的成本提炼可比较的模型-迫使原始", "moval attacks.", "模型创建者退出市场并扼杀未来的创新。预计这一点", "理性的LLM所有者可能永远不会以公平的价格提供", "API访问。相反，他们必须大幅增加费用-远远超出服", "1.Introduction", "务成本-使模型蒸馏无利可图。", "大型语言模型（LLM）已经变得越来越强大（Brow", "对于具有良好商业价值的下游应用程序，增加的费用", "etal.，2020Ouyangetal.，2022），但是由于高昂的培", "可能是可以容忍的，但它使研究和LLM应用程序无", "训成本，它们的所有者不愿意开源它们。大多数公司", "法负担公共利益。为了解决这个问题，政府可以资助", "仅免费或付费提供对其模型的API访问，以支付创新", "公共LLM进行研究和公共服务。但是，不道德的实体", "也可以将公共LLM蒸馏以获取商业利益或军事用途", "Department of Computer Science, UC Santa", "Barbara", "Correspondence to:Xuandong Zhao <", "最近，斯坦福大学的一个小组（Taorietal.，2023）声", "[email protected] >, Yu-Xiang Wang < yuxi-", "称在API调用中只为$600蒸馏了ChatGPT，这表", "[email protected] >, Lei Li < [email protected] >.", "明了问题的紧迫性。从理论上讲，可以证明与重新训", "Proceedings of the 4oth International Conference on Ma-", "练模型相比，提取模型的成本更低（Shalev-Shwartz&", "chine Learning, Honolulu, Hawaii, USA. PMLR 202, 2023.", "Ben-David，2014）：将模型从原始数据训练到e的样", "Copyright 2023 by the author(s).", "本复杂性-超额风险约为O(1/e2）；而使用模型生成的", "1" ]

{ "arxiv_id": "2302.03162", "language": "zh", "page_idx": 0 }

[ [ "0.3594", "0.05936", "0.614", "0.05936", "0.614", "0.06946", "0.3594", "0.06946" ], [ "0.0907", "0.0884", "0.4722", "0.0884", "0.4722", "0.0985", "0.0907", "0.0985" ], [ "0.5024", "0.0884", "0.884", "0.0884", "0.884", "0.0985", "0.5024", "0.0985" ], [ "0.08905", "0.1016", "0.1993", "0.1016", "0.1993", "0.11554", "0.08905", "0.11554" ], [ "0.5024", "0.1016", "0.8823", "0.1016", "0.8823", "0.11554", "0.5024", "0.11554" ], [ "0.5024", "0.1187", "0.884", "0.1187", "0.884", "0.1288", "0.5024", "0.1288" ], [ "0.08905", "0.125", "0.4756", "0.125", "0.4756", "0.1383", "0.08905", "0.1383" ], [ "0.5034", "0.1338", "0.8823", "0.1338", "0.8823", "0.1445", "0.5034", "0.1445" ], [ "0.09235", "0.1414", "0.2279", "0.1414", "0.2279", "0.1515", "0.09235", "0.1515" ], [ "0.5024", "0.1465", "0.8857", "0.1465", "0.8857", "0.161", "0.5024", "0.161" ], [ "0.0907", "0.1622", "0.4739", "0.1622", "0.4739", "0.1755", "0.0907", "0.1755" ], [ "0.5024", "0.1622", "0.884", "0.1622", "0.884", "0.1755", "0.5024", "0.1755" ], [ "0.0907", "0.178", "0.4722", "0.178", "0.4722", "0.1913", "0.0907", "0.1913" ], [ "0.5", "0.1768", "0.824", "0.1768", "0.824", "0.1901", "0.5", "0.1901" ], [ "0.0907", "0.1925", "0.4722", "0.1925", "0.4722", "0.2058", "0.0907", "0.2058" ], [ "0.0907", "0.207", "0.4739", "0.207", "0.4739", "0.2203", "0.0907", "0.2203" ], [ "0.5", "0.2108", "0.8857", "0.2122", "0.8857", "0.2267", "0.5", "0.226" ], [ "0.0907", "0.2229", "0.4722", "0.2229", "0.4722", "0.2361", "0.0907", "0.2361" ], [ "0.5034", "0.2299", "0.8823", "0.2299", "0.8823", "0.2405", "0.5034", "0.2405" ], [ "0.0907", "0.2374", "0.4739", "0.2374", "0.4739", "0.2507", "0.0907", "0.2507" ], [ "0.501", "0.2412", "0.8857", "0.2424", "0.8857", "0.2568", "0.501", "0.2556" ], [ "0.09235", "0.254", "0.4714", "0.254", "0.4714", "0.2644", "0.09235", "0.2644" ], [ "0.5", "0.2583", "0.8813", "0.2583", "0.8813", "0.2715", "0.5", "0.2715" ], [ "0.09235", "0.2695", "0.4714", "0.2695", "0.4714", "0.2798", "0.09235", "0.2798" ], [ "0.501", "0.2727", "0.8857", "0.2727", "0.8857", "0.286", "0.501", "0.286" ], [ "0.09235", "0.2854", "0.4722", "0.2854", "0.4722", "0.2954", "0.09235", "0.2954" ], [ "0.5024", "0.2886", "0.8857", "0.2886", "0.8857", "0.3018", "0.5024", "0.3018" ], [ "0.0907", "0.2974", "0.4739", "0.2974", "0.4739", "0.3113", "0.0907", "0.3113" ], [ "0.5034", "0.305", "0.884", "0.305", "0.884", "0.315", "0.5034", "0.315" ], [ "0.0907", "0.3132", "0.4722", "0.3132", "0.4722", "0.327", "0.0907", "0.327" ], [ "0.5024", "0.318", "0.884", "0.318", "0.884", "0.332", "0.5024", "0.332" ], [ "0.0907", "0.3289", "0.415", "0.3289", "0.415", "0.3423", "0.0907", "0.3423" ], [ "0.5024", "0.334", "0.7466", "0.334", "0.7466", "0.348", "0.5024", "0.348" ], [ "0.0907", "0.351", "0.4722", "0.351", "0.4722", "0.3643", "0.0907", "0.3643" ], [ "0.0907", "0.3674", "0.4722", "0.3674", "0.4722", "0.3806", "0.0907", "0.3806" ], [ "0.501", "0.3687", "0.8857", "0.3687", "0.8857", "0.3833", "0.501", "0.3833" ], [ "0.08905", "0.3818", "0.4739", "0.3818", "0.4739", "0.3958", "0.08905", "0.3958" ], [ "0.5034", "0.3865", "0.8823", "0.3865", "0.8823", "0.3965", "0.5034", "0.3965" ], [ "0.0907", "0.3977", "0.4739", "0.3977", "0.4739", "0.411", "0.0907", "0.411" ], [ "0.5024", "0.401", "0.8813", "0.401", "0.8813", "0.411", "0.5024", "0.411" ], [ "0.08905", "0.4124", "0.4739", "0.411", "0.4739", "0.4248", "0.08905", "0.4255" ], [ "0.5", "0.414", "0.8823", "0.414", "0.8823", "0.428", "0.5", "0.428" ], [ "0.0907", "0.428", "0.4739", "0.428", "0.4739", "0.4414", "0.0907", "0.4414" ], [ "0.5024", "0.43", "0.8857", "0.43", "0.8857", "0.443", "0.5024", "0.443" ], [ "0.09314", "0.4443", "0.4722", "0.4443", "0.4722", "0.455", "0.09314", "0.455" ], [ "0.5034", "0.4463", "0.8823", "0.4463", "0.8823", "0.457", "0.5034", "0.457" ], [ "0.0907", "0.4583", "0.4722", "0.4583", "0.4722", "0.4717", "0.0907", "0.4717" ], [ "0.5024", "0.4602", "0.884", "0.4602", "0.884", "0.4734", "0.5024", "0.4734" ], [ "0.09235", "0.4749", "0.2917", "0.4749", "0.2917", "0.4849", "0.09235", "0.4849" ], [ "0.5024", "0.4749", "0.884", "0.4749", "0.884", "0.488", "0.5024", "0.488" ], [ "0.501", "0.4893", "0.8823", "0.4893", "0.8823", "0.5024", "0.501", "0.5024" ], [ "0.0866", "0.4968", "0.4756", "0.4968", "0.4756", "0.511", "0.0866", "0.511" ], [ "0.5", "0.505", "0.884", "0.505", "0.884", "0.518", "0.5", "0.518" ], [ "0.10455", "0.512", "0.4739", "0.512", "0.4739", "0.5264", "0.10455", "0.5264" ], [ "0.5024", "0.5195", "0.8857", "0.5195", "0.8857", "0.5327", "0.5024", "0.5327" ], [ "0.1054", "0.528", "0.4714", "0.528", "0.4714", "0.5415", "0.1054", "0.5415" ], [ "0.5024", "0.535", "0.884", "0.535", "0.884", "0.549", "0.5024", "0.549" ], [ "0.10455", "0.5415", "0.255", "0.5415", "0.255", "0.5547", "0.10455", "0.5547" ], [ "0.501", "0.55", "0.8823", "0.55", "0.8823", "0.563", "0.501", "0.563" ], [ "0.08826", "0.5693", "0.4722", "0.5693", "0.4722", "0.5835", "0.08826", "0.5835" ], [ "0.501", "0.5645", "0.8823", "0.565", "0.8823", "0.579", "0.501", "0.5776" ], [ "0.10455", "0.5835", "0.4722", "0.5835", "0.4722", "0.5967", "0.10455", "0.5967" ], [ "0.5024", "0.582", "0.8813", "0.582", "0.8813", "0.5923", "0.5024", "0.5923" ], [ "0.10706", "0.601", "0.4714", "0.601", "0.4714", "0.6113", "0.10706", "0.6113" ], [ "0.5024", "0.595", "0.884", "0.595", "0.884", "0.6094", "0.5024", "0.6094" ], [ "0.10706", "0.6143", "0.4722", "0.6143", "0.4722", "0.6274", "0.10706", "0.6274" ], [ "0.501", "0.6094", "0.8823", "0.61", "0.8823", "0.625", "0.501", "0.6235" ], [ "0.1054", "0.63", "0.4722", "0.63", "0.4722", "0.643", "0.1054", "0.643" ], [ "0.501", "0.625", "0.8823", "0.6255", "0.8823", "0.64", "0.501", "0.6396" ], [ "0.1054", "0.643", "0.4714", "0.643", "0.4714", "0.657", "0.1054", "0.657" ], [ "0.5034", "0.6416", "0.5986", "0.6416", "0.5986", "0.653", "0.5034", "0.653" ], [ "0.10706", "0.662", "0.4714", "0.662", "0.4714", "0.6724", "0.10706", "0.6724" ], [ "0.1054", "0.6753", "0.4722", "0.6753", "0.4722", "0.6895", "0.1054", "0.6895" ], [ "0.4983", "0.6753", "0.749", "0.6753", "0.749", "0.69", "0.4983", "0.69" ], [ "0.10297", "0.69", "0.2793", "0.6895", "0.2793", "0.704", "0.10297", "0.705" ], [ "0.5", "0.709", "0.642", "0.7104", "0.642", "0.7246", "0.5", "0.7236" ], [ "0.08826", "0.7236", "0.2296", "0.7246", "0.2296", "0.7393", "0.08826", "0.738" ], [ "0.5024", "0.733", "0.8813", "0.733", "0.8813", "0.747", "0.5024", "0.747" ], [ "0.08826", "0.7485", "0.4739", "0.75", "0.4739", "0.7646", "0.08826", "0.763" ], [ "0.5034", "0.7505", "0.884", "0.7505", "0.884", "0.761", "0.5034", "0.761" ], [ "0.0907", "0.765", "0.4739", "0.765", "0.4739", "0.779", "0.0907", "0.779" ], [ "0.5", "0.762", "0.8857", "0.763", "0.8857", "0.778", "0.5", "0.7764" ], [ "0.08905", "0.78", "0.4739", "0.78", "0.4739", "0.7935", "0.08905", "0.7935" ], [ "0.501", "0.778", "0.884", "0.779", "0.884", "0.7935", "0.501", "0.7925" ], [ "0.0907", "0.7954", "0.4739", "0.7954", "0.4739", "0.8096", "0.0907", "0.8096" ], [ "0.5", "0.7925", "0.568", "0.7944", "0.568", "0.8096", "0.5", "0.8066" ], [ "0.09235", "0.8125", "0.4714", "0.8125", "0.4714", "0.8228", "0.09235", "0.8228" ], [ "0.5024", "0.8174", "0.884", "0.8174", "0.884", "0.8306", "0.5024", "0.8306" ], [ "0.0907", "0.8257", "0.4739", "0.8257", "0.4739", "0.839", "0.0907", "0.839" ], [ "0.501", "0.832", "0.884", "0.832", "0.884", "0.845", "0.501", "0.845" ], [ "0.0907", "0.8413", "0.4739", "0.8413", "0.4739", "0.855", "0.0907", "0.855" ], [ "0.501", "0.8467", "0.8857", "0.8467", "0.8857", "0.86", "0.501", "0.86" ], [ "0.0907", "0.856", "0.4739", "0.856", "0.4739", "0.869", "0.0907", "0.869" ], [ "0.501", "0.8613", "0.884", "0.8623", "0.884", "0.8755", "0.501", "0.875" ], [ "0.10864", "0.8774", "0.4739", "0.88", "0.4739", "0.8945", "0.10864", "0.8926" ], [ "0.5034", "0.879", "0.8823", "0.879", "0.8823", "0.8896", "0.5034", "0.8896" ], [ "0.08826", "0.8916", "0.2524", "0.893", "0.2524", "0.907", "0.08826", "0.905" ], [ "0.501", "0.8926", "0.8857", "0.8926", "0.8857", "0.906", "0.501", "0.906" ], [ "0.482", "0.924", "0.4927", "0.924", "0.4927", "0.936", "0.482", "0.936" ] ]

[ "Title Suppressed Due to ExcessiveSize", "标签的模型蒸馏以“可实现”的方式运行，因此仅需", "较低的价格将其作为云服务安装。这种类型的攻击与", "要0(1/e)样本", "知识提炼（KD）（Hintonetal.，2015）密切相关，其中", "攻击者充当近似受害者模型行为的学生模型。在本文", "这给我们带来了一个中心问题：我们如何通过蒸馏来", "中，我们特别关注文本生成中的模型提取攻击。先前", "防止模型窃取攻击？", "的作品（Wallaceetal.，2020;Xuetal.,2021）已经表", "这似乎是一个不可能的任务。通过提供最先进的LLM", "明，对手可以使用序列级知识提炼来模仿商业文本生", "服务，复制它所需的信息已被编码在提供的API输", "成API的功能，这对云平台构成了严重威胁。", "出本身中。研究人员决定放弃阻止可学习性，而是尝", "试对API输出进行水印处理，以便使用水印输出训练", "Watermarking水印是一种技术，用于将看不见的标", "的模型可以追溯到原始模型。最近的作品（Adietal.", "签嵌入到音频，视频或图像等信号中，以识别信号", "2018；Zhangetal.，2018）使用触发器集在分发之前在", "版权的所有者。在机器学习模型的背景下，一些研", "商业模型的神经元上嵌入不可见的水印。当怀疑模型", "究 (Merrer et al., 2017; Adi et al., 2018; Zhang et al.", "被盗时，模型所有者可以借助触发器集进行官方所有", "2018）使用水印来防止这些模型的精确重复，方法是", "权索赔。现有作品（Heetal.2021；2022）使用同义词", "将其插人受保护模型的参数中或构建激活特定预测的", "替换策略将表面级水印添加到文本生成模型中。但是，", "后门图像。但是，保护模型免受模型提取攻击是困难", "攻击者可以绕过这些方法，随机替换输出中的同义词", "的，因为可疑模型的参数可能与受害者模型的参数不", "并从提取的模型中删除水印，从而使保护无效", "同，并且后门行为可能不会被转移", "在本文中，我们提出了genergativeinvisiblesequence", "watermarking（Ginsew），一种保护文本生成模型并", "Watermarkingagainstmodel extraction最近，一些", "检测被盗模型的方法。图1说明了整个过程。核心思", "研究试图解决识别从受害者模型中提取的提取模型的", "想是将一个秘密的正弦信号注人到模型的生成概率中。", "挑战，在图像分类（Charetteetal.2022：Jiaetal.", "此信号不会损害模型的生成质量。为了检测候选模型", "2021）和文本分类（Zha0etal.，2022）方面取得了有希", "是否从目标模型中窃取，Ginsew识别正弦频率并将其", "望的结果。CosWM（Charetteetal.，2022）将余弦信", "与密钥进行比较。Ginsew提供了一种更强大的机制来", "号形式的水印嵌入到受保护模型的输出中。这种水印", "保护模型的知识产权，即使在随机同义词替换攻击下", "很难消除，使其成为识别从受保护模型中提取的模型", "也是如此。本文的贡献如下：", "的有效方法。尽管如此，CosWM仅适用于图像分类任", "务。对于文本生成，Heetal.（2021）提出了一种词汇", ": We propose g enerative in visible se quence w ater-", "水印方法来识别由提取攻击引起的IP侵权。该方法涉", "marking ( Ginsew), a method to protect text gener-", "及从受害者模型的训练数据中选择一组单词，为它们", "ation models against model extraction attacks with", "找到语义等效的替换，并将它们替换为替换。另一种", "invisible watermarks", "方法CATER（Heetal.,2022）通过基于语言特征替换", ". We carry out experiments on machine translation", "某些单词的同义词来提出条件水印。然而，这两种方", "and story generation on a variety of models. Exper", "法都是表面级水印。攻击者可以通过随机替换输出中", "imental results showthat our method Ginsew out", "的同义词来轻松绕过这些方法，从而难以通过探测可", "performs the previous methods in both generation", "疑模型进行验证。Ginsew直接修改输出令牌的概率分", "quality and robustness of IP infringement detection", "布，这使得水印不可见，并为识别提取的模型提供了", "ability. Even with adversarial watermark removal", "更强大的机制", "attacks,Ginsewstillgainsasignificantimprovement", "of 19 to 29 points in mean average precision (mAP)", "3. Proposed method: Ginsew", "of detecting suspects 1.", "3.1. Problem setup", "2. Related work", "我们的目标是保护文本生成模型免受模型提取攻击", "Modelextractionattacks模型提取攻击，也称为模", "它使受害模型所有者或第三方仲裁员能够将从受害", "型反演或模型窃取，对机器学习模型（Trameretal.，", "API中提取的可疑模型的所有权归于。我们利用提取", "2016; Orekondy et al., 2019; Wallace et al., 2020; He", "的模型从受害者模型中学到的秘密知识作为归属所有", "etal.，2020）的机密性构成重大威胁。这些攻击旨在通", "权的签名", "过创建或收集替代数据集来模仿黑盒受害者模型的功", "在模型提取攻击中，表示为S的对手只能黑盒访问受", "能。然后，攻击者使用受害者模型的API来预测替代", "害者模型的API，表示为V。攻击者可以使用辅助的", "数据集的标签。有了这个伪标签数据集，攻击者可以", "未标记数据集查询√，但只能观察文本输出，而不能", "训练一个模仿受害者模型的高性能模型，甚至可以以", "观察API产生的底层概率。因此，攻击者接收V的输", "lOur source code is available at https://github.com/", "出作为生成输出或伪标签。攻击者的自目标是采用序列", "XuandongZhao/Ginsew", "级知识提炼来复制V的功能。模型提取攻击如图1(a)", "2" ]

{ "arxiv_id": "2302.03162", "language": "zh", "page_idx": 1 }

[ [ "0.3586", "0.05682", "0.6167", "0.05682", "0.6167", "0.0707", "0.3586", "0.0707" ], [ "0.08905", "0.0859", "0.3342", "0.0871", "0.3342", "0.101", "0.08905", "0.09973" ], [ "0.5024", "0.0859", "0.8594", "0.0859", "0.8594", "0.09973", "0.5024", "0.09973" ], [ "0.09314", "0.1029", "0.4739", "0.1016", "0.4739", "0.11615", "0.09314", "0.11743" ], [ "0.505", "0.1029", "0.8857", "0.1016", "0.8857", "0.11615", "0.505", "0.11743" ], [ "0.1177", "0.1187", "0.4722", "0.1187", "0.4722", "0.132", "0.1177", "0.132" ], [ "0.528", "0.1187", "0.884", "0.1187", "0.884", "0.132", "0.528", "0.132" ], [ "0.11926", "0.1338", "0.4739", "0.1338", "0.4739", "0.1477", "0.11926", "0.1477" ], [ "0.5312", "0.1332", "0.8203", "0.1332", "0.8203", "0.1477", "0.5312", "0.1477" ], [ "0.1168", "0.1465", "0.2615", "0.1497", "0.2615", "0.1635", "0.1168", "0.1604" ], [ "0.509", "0.1483", "0.7036", "0.1483", "0.7036", "0.1622", "0.509", "0.1622" ], [ "0.0948", "0.1635", "0.2917", "0.1635", "0.2917", "0.178", "0.0948", "0.178" ], [ "0.5073", "0.1635", "0.6445", "0.1635", "0.6445", "0.1768", "0.5073", "0.1768" ], [ "0.0948", "0.178", "0.232", "0.178", "0.232", "0.1913", "0.0948", "0.1913" ], [ "0.5063", "0.178", "0.7134", "0.1768", "0.7134", "0.1913", "0.5063", "0.1925" ], [ "0.09644", "0.1932", "0.2998", "0.1932", "0.2998", "0.207", "0.09644", "0.207" ], [ "0.509", "0.1957", "0.524", "0.1957", "0.524", "0.2058", "0.509", "0.2058" ], [ "0.5415", "0.1932", "0.6504", "0.1932", "0.6504", "0.207", "0.5415", "0.207" ], [ "0.1299", "0.2084", "0.2385", "0.207", "0.2385", "0.2216", "0.1299", "0.2229" ], [ "0.5435", "0.2084", "0.6177", "0.2084", "0.6177", "0.2229", "0.5435", "0.2229" ], [ "0.509", "0.2102", "0.523", "0.2102", "0.523", "0.2216", "0.509", "0.2216" ], [ "0.09644", "0.2102", "0.1111", "0.2102", "0.1111", "0.2203", "0.09644", "0.2203" ], [ "0.1316", "0.2235", "0.3816", "0.2235", "0.3816", "0.238", "0.1316", "0.238" ], [ "0.509", "0.2247", "0.523", "0.2247", "0.523", "0.2361", "0.509", "0.2361" ], [ "0.5435", "0.2235", "0.71", "0.2235", "0.71", "0.238", "0.5435", "0.238" ], [ "0.1478", "0.238", "0.4756", "0.2374", "0.4756", "0.252", "0.1478", "0.2524" ], [ "0.0972", "0.226", "0.11273", "0.226", "0.11273", "0.2361", "0.0972", "0.2361" ], [ "0.0948", "0.2393", "0.1136", "0.2393", "0.1136", "0.2524", "0.0948", "0.2524" ], [ "0.509", "0.2405", "0.523", "0.2405", "0.523", "0.252", "0.509", "0.252" ], [ "0.5596", "0.2393", "0.734", "0.2393", "0.734", "0.254", "0.5596", "0.254" ], [ "0.1487", "0.254", "0.2852", "0.254", "0.2852", "0.2676", "0.1487", "0.2676" ], [ "0.509", "0.2556", "0.524", "0.2556", "0.524", "0.2664", "0.509", "0.2664" ], [ "0.5435", "0.2551", "0.597", "0.2551", "0.597", "0.2664", "0.5435", "0.2664" ], [ "0.0948", "0.2695", "0.1136", "0.2695", "0.1136", "0.2827", "0.0948", "0.2827" ], [ "0.1462", "0.2664", "0.2646", "0.2695", "0.2646", "0.2886", "0.1462", "0.2854" ], [ "0.5034", "0.2695", "0.58", "0.2695", "0.58", "0.281", "0.5034", "0.281" ], [ "0.1478", "0.2854", "0.2754", "0.2886", "0.2754", "0.3037", "0.1478", "0.3005" ], [ "0.5", "0.2827", "0.8857", "0.2842", "0.8857", "0.2986", "0.5", "0.2974" ], [ "0.528", "0.2986", "0.763", "0.2986", "0.763", "0.3125", "0.528", "0.3125" ], [ "0.0948", "0.2874", "0.11273", "0.2874", "0.11273", "0.3005", "0.0948", "0.3005" ], [ "0.0907", "0.303", "0.11273", "0.303", "0.11273", "0.315", "0.0907", "0.315" ], [ "0.1299", "0.3037", "0.1854", "0.3037", "0.1854", "0.315", "0.1299", "0.315" ], [ "0.08826", "0.315", "0.1691", "0.3176", "0.1691", "0.332", "0.08826", "0.33" ], [ "0.501", "0.315", "0.6094", "0.315", "0.6094", "0.3289", "0.501", "0.3289" ], [ "0.0866", "0.3289", "0.4739", "0.3308", "0.4739", "0.3484", "0.0866", "0.3467" ], [ "0.1152", "0.3499", "0.3308", "0.348", "0.3308", "0.361", "0.1152", "0.363" ], [ "0.08905", "0.3643", "0.4739", "0.3643", "0.4739", "0.379", "0.08905", "0.379" ], [ "0.505", "0.3623", "0.6006", "0.3623", "0.6006", "0.3738", "0.505", "0.3738" ], [ "0.11273", "0.3757", "0.273", "0.3757", "0.273", "0.401", "0.11273", "0.401" ], [ "0.5024", "0.3838", "0.884", "0.3838", "0.884", "0.3977", "0.5024", "0.3977" ], [ "0.08826", "0.3965", "0.3228", "0.3965", "0.3228", "0.411", "0.08826", "0.411" ], [ "0.5024", "0.3984", "0.8857", "0.3984", "0.8857", "0.4124", "0.5024", "0.4124" ], [ "0.08826", "0.411", "0.1993", "0.4124", "0.1993", "0.4268", "0.08826", "0.4255" ], [ "0.5024", "0.4136", "0.884", "0.4136", "0.884", "0.4268", "0.5024", "0.4268" ], [ "0.5034", "0.43", "0.8823", "0.43", "0.8823", "0.44", "0.5034", "0.44" ], [ "0.5024", "0.443", "0.884", "0.443", "0.884", "0.457", "0.5024", "0.457" ], [ "0.08826", "0.457", "0.4272", "0.457", "0.4272", "0.4702", "0.08826", "0.4702" ], [ "0.5024", "0.459", "0.8857", "0.459", "0.8857", "0.473", "0.5024", "0.473" ], [ "0.501", "0.4734", "0.6167", "0.4734", "0.6167", "0.4873", "0.501", "0.4873" ], [ "0.0907", "0.4797", "0.4722", "0.4797", "0.4722", "0.4937", "0.0907", "0.4937" ], [ "0.0907", "0.4956", "0.4722", "0.4956", "0.4722", "0.509", "0.0907", "0.509" ], [ "0.0907", "0.5103", "0.4714", "0.5103", "0.4714", "0.5234", "0.0907", "0.5234" ], [ "0.5", "0.509", "0.6357", "0.509", "0.6357", "0.5234", "0.5", "0.5234" ], [ "0.0907", "0.526", "0.4739", "0.526", "0.4739", "0.539", "0.0907", "0.539" ], [ "0.501", "0.534", "0.8857", "0.5327", "0.8857", "0.5474", "0.501", "0.549" ], [ "0.0907", "0.5405", "0.335", "0.5405", "0.335", "0.5537", "0.0907", "0.5537" ], [ "0.5024", "0.55", "0.884", "0.55", "0.884", "0.563", "0.5024", "0.563" ], [ "0.0907", "0.5625", "0.4739", "0.5625", "0.4739", "0.5757", "0.0907", "0.5757" ], [ "0.4983", "0.5645", "0.544", "0.561", "0.5474", "0.5747", "0.501", "0.5776" ], [ "0.09235", "0.577", "0.4617", "0.577", "0.4617", "0.5903", "0.09235", "0.5903" ], [ "0.08905", "0.5933", "0.4739", "0.5933", "0.4739", "0.607", "0.08905", "0.607" ], [ "0.5", "0.5947", "0.884", "0.595", "0.884", "0.61", "0.5", "0.6094" ], [ "0.08826", "0.608", "0.4739", "0.607", "0.4739", "0.622", "0.08826", "0.6226" ], [ "0.5", "0.61", "0.884", "0.61", "0.884", "0.6235", "0.5", "0.6235" ], [ "0.0907", "0.6235", "0.4739", "0.6235", "0.4739", "0.638", "0.0907", "0.638" ], [ "0.501", "0.627", "0.8857", "0.627", "0.8857", "0.64", "0.501", "0.64" ], [ "0.0866", "0.637", "0.4756", "0.636", "0.4756", "0.654", "0.0866", "0.655" ], [ "0.5024", "0.6416", "0.8857", "0.6416", "0.8857", "0.655", "0.5024", "0.655" ], [ "0.0907", "0.653", "0.4739", "0.653", "0.4739", "0.6675", "0.0907", "0.6675" ], [ "0.501", "0.657", "0.8823", "0.657", "0.8823", "0.6704", "0.501", "0.6704" ], [ "0.0907", "0.6685", "0.4722", "0.6685", "0.4722", "0.6816", "0.0907", "0.6816" ], [ "0.501", "0.672", "0.8857", "0.672", "0.8857", "0.686", "0.501", "0.686" ], [ "0.0907", "0.6846", "0.4722", "0.6846", "0.4722", "0.6978", "0.0907", "0.6978" ], [ "0.5024", "0.6875", "0.8857", "0.6875", "0.8857", "0.7007", "0.5024", "0.7007" ], [ "0.0907", "0.6997", "0.4739", "0.6997", "0.4739", "0.7134", "0.0907", "0.7134" ], [ "0.501", "0.7007", "0.884", "0.702", "0.884", "0.7163", "0.501", "0.7153" ], [ "0.08826", "0.712", "0.2532", "0.714", "0.2532", "0.729", "0.08826", "0.7266" ], [ "0.5024", "0.7163", "0.8857", "0.7163", "0.8857", "0.731", "0.5024", "0.731" ], [ "0.501", "0.732", "0.777", "0.732", "0.777", "0.7456", "0.501", "0.7456" ], [ "0.2893", "0.7407", "0.3203", "0.7407", "0.3203", "0.752", "0.2893", "0.752" ], [ "0.1854", "0.7505", "0.2427", "0.756", "0.2402", "0.7695", "0.183", "0.7646" ], [ "0.326", "0.752", "0.3757", "0.752", "0.3757", "0.7666", "0.326", "0.7666" ], [ "0.501", "0.763", "0.8857", "0.763", "0.8857", "0.7764", "0.501", "0.7764" ], [ "0.501", "0.779", "0.884", "0.779", "0.884", "0.7925", "0.501", "0.7925" ], [ "0.1136", "0.7915", "0.1584", "0.7935", "0.1569", "0.8047", "0.11273", "0.803" ], [ "0.2778", "0.7925", "0.3308", "0.7925", "0.3308", "0.8037", "0.2778", "0.8037" ], [ "0.3865", "0.7905", "0.4355", "0.7905", "0.4355", "0.8047", "0.3865", "0.8047" ], [ "0.1748", "0.799", "0.2091", "0.799", "0.2091", "0.81", "0.1748", "0.81" ], [ "0.5034", "0.7944", "0.884", "0.7944", "0.884", "0.808", "0.5034", "0.808" ], [ "0.3513", "0.8037", "0.3823", "0.8066", "0.3816", "0.815", "0.3506", "0.8125" ], [ "0.505", "0.8096", "0.8857", "0.8096", "0.8857", "0.8228", "0.505", "0.8228" ], [ "0.2035", "0.8193", "0.367", "0.8193", "0.367", "0.834", "0.2035", "0.834" ], [ "0.452", "0.821", "0.4722", "0.821", "0.4722", "0.8335", "0.452", "0.8335" ], [ "0.5024", "0.8237", "0.8823", "0.8237", "0.8823", "0.837", "0.5024", "0.837" ], [ "0.501", "0.839", "0.8823", "0.839", "0.8823", "0.853", "0.501", "0.853" ], [ "0.09235", "0.845", "0.4722", "0.845", "0.4722", "0.86", "0.09235", "0.86" ], [ "0.501", "0.854", "0.8857", "0.853", "0.8857", "0.867", "0.501", "0.8687" ], [ "0.0907", "0.8613", "0.4722", "0.8613", "0.4722", "0.875", "0.0907", "0.875" ], [ "0.0907", "0.877", "0.4722", "0.877", "0.4722", "0.89", "0.0907", "0.89" ], [ "0.5195", "0.877", "0.722", "0.879", "0.722", "0.893", "0.5195", "0.8916" ], [ "0.08905", "0.8926", "0.4739", "0.8926", "0.4739", "0.906", "0.08905", "0.906" ], [ "0.5195", "0.89", "0.7534", "0.8926", "0.7534", "0.907", "0.5195", "0.905" ], [ "0.482", "0.924", "0.4944", "0.924", "0.4944", "0.935", "0.482", "0.935" ] ]

[ "Title Suppressed Due to Excessive Size", "Algorithm 2 Watermark detection", "Algorithm 3 Watermark detection with text alone", "1:Inputs: Suspect model S, sample probing data D", "1:Inputs: Suspect model S, sample probing data D", "from the training data of S , vocab V, group 1 Gi", "from the training data of S , vocab V, group 1 Gi", ", group 2 G2, hash function g(a, u, M) , filtering", ", group 2 G2 , hash function g(α, u, M) .", "threshold value qmin", "2: Output: Signal strength", "2: Output: Signal strength", "3: Initialize H =0", "3: Initialize H = 0", "4: for each input in D do", "4: for each input in D do", "5:", "t= g(v,a, M)", "t= g(v,a, M)", "y←s(α)", "6:", "5:", "for each decoding step of S(α) do", "7:", "for each token of y do", "Get probabilityvectorpfrom the decoder of", "6:", "7:", "8:", "H <H U (t, 1(y; E Gi))", "the suspect model.", "9:", "end for", "8:", "Qgi =Eiegr Pi", "10:end for", "H←HU(t,Qg1)", "1l: Compute the Lomb-Scargle periodogram from H,", "and compute Psnr in Equation 5", "9:", "10:", "end for", "11: end for", "12: return Pnr", "12: Filter out elements in H where Qgi ≤ qmin , re-", "maining pairs form the set H", "13: Compute the Lomb-Scargle periodogram from the", "了受害者模型", "pairs (t(k),Qd) eH", "值得注意的是，可以调整Lomb-Scargle周期图方法中", "14: Compute Psnr in Equation 5", "使用的阈值和频率参数，以优化所提出方法的性能。我", "15: return Psnr", "们实验中使用的具体设置将在本文的实验部分进行讨", "论。：此外，我们提出了3算法，证明我们可以通过仅", "分析生成的文本本身来检测水印，而不依赖于模型的", "3.3. Detecting watermark from suspect models", "预测概率。关于这种纯文本方法的更详细的讨论可以", "在5.2节中找到", "为了检测IP侵权的实例，我们首先创建一个探测数据", "集D，以在解码步骤的概率向量中提取带有水印的信", "号。D可以从提取模型的训练数据中获得，因为所有", "4.Experiments", "者能够存储特定最终用户发送的任何查询。此外，D", "我们评估Ginsew在两个常见文本生成任务上的性能：", "也可以根据需要从其他分布中提取", "机器翻译和故事生成。有多个公共API可用于这些模", "从可疑模型中检测水印的过程在算法2中概述。对", "型23", "于每个探测文本输入，该方法首先获取哈希值t", "g(a,u,M）。接下来，对于可疑模型的每个解码步骤，", "Machinetranslation在机器翻译任务中，我们利用", "我们将对（t,Qg）添加到集合H中。为了消除低置信", "IWSLT14和WMT14数据集（Cettolo etal..2014:", "度的输出，我们使用Qg1≤qmin过滤掉对，其中阈值", "Bojaretal.,2014），特别关注德语（De）到英语（En）", "qmin是提取过程的恒定参数。然后使用Lomb-Scargle", "的翻译。我们基于BLEU（Papinenietal.，2002）和", "周期图（Scargle，1982）方法来估计探测集H中特定频", "BERTScore（Zhangetal.，2019）指标评估翻译质量，", "率fu下的傅立叶功率谱P(f）。通过应用近似傅里叶", "我们采用火车/有效/测试集的官方拆分。BLEU通过", "变换，我们放大了概率向量中的细微扰动。这样我们", "比较n-gram来关注词汇相似性，而BERTScore则通", "就可以在频率fu的功率谱中检测到一个峰值。这允", "过上下文嵌入来关注语义等价。对于IWSLT14，使用", "许通过计算信噪比Psnr", "由7,000BPE（Sennrichetal.，2016）单位组成的词汇", "表，而WMT14使用32.000BPE单位", "fw+号", "Psignal", "P(f)df", "Storygeneration对于故事生成任务，我们使用ROC", "stories（Mostafazadehetal.，2016）语料库。此数据集", "Pnoise", "P(f)df", "P(f)df", "F-", "中的每个故事包括5个句子，前4个句子作为故事的", "fw+", "上下文和模型的输人，第5个句子是要预测的故事的", "Psnr = Psignal /Pnoises,", "(5)", "结尾。训练集中有90.000个样本，验证集和测试集中有", "4081个样本。基于ROUGE（Lin，2004）和BERTScore", "来评估信号强度，其中控制「fu一号，fw+号l的窗", "指标评估发电质量。此任务中使用了25,000个BPE", "口宽度；F是最大频率，fu是嵌入受害者模型的角频", "率。较高的Psnr表示频域中的峰值较高，因此在可疑", "2https://translate.google.com/", "模型中存在秘密信号的可能性较高，这证实了它提取", "3https://beta.openai.com/overview", "4" ]

{ "arxiv_id": "2302.03162", "language": "zh", "page_idx": 3 }

[ [ "0.3525", "0.0861", "0.6543", "0.0861", "0.6543", "0.10095", "0.3525", "0.10095" ], [ "0.3955", "0.1051", "0.608", "0.1051", "0.608", "0.1194", "0.3955", "0.1194" ], [ "0.1646", "0.1443", "0.8403", "0.1443", "0.8403", "0.158", "0.1646", "0.158" ], [ "0.3315", "0.1609", "0.6733", "0.1609", "0.6733", "0.1758", "0.3315", "0.1758" ], [ "0.3535", "0.1758", "0.6533", "0.1746", "0.6533", "0.1924", "0.3535", "0.1936" ], [ "0.3525", "0.1948", "0.656", "0.1948", "0.656", "0.209", "0.3525", "0.209" ], [ "0.1855", "0.212", "0.8213", "0.212", "0.8213", "0.2257", "0.1855", "0.2257" ], [ "0.266", "0.2654", "0.3435", "0.2654", "0.3435", "0.2798", "0.266", "0.2798" ], [ "0.5225", "0.2666", "0.5728", "0.2666", "0.5728", "0.2808", "0.5225", "0.2808" ], [ "0.6465", "0.2654", "0.705", "0.2673", "0.7036", "0.282", "0.6455", "0.2798" ], [ "0.5845", "0.2715", "0.634", "0.2715", "0.634", "0.2861", "0.5845", "0.2861" ], [ "0.8613", "0.2666", "0.9194", "0.2642", "0.921", "0.279", "0.8623", "0.2808" ], [ "0.7197", "0.2725", "0.777", "0.2725", "0.777", "0.2874", "0.7197", "0.2874" ], [ "0.7866", "0.2725", "0.846", "0.2708", "0.848", "0.2852", "0.7886", "0.2874" ], [ "0.6465", "0.279", "0.6963", "0.279", "0.6963", "0.2935", "0.6465", "0.2935" ], [ "0.8613", "0.279", "0.906", "0.279", "0.906", "0.2935", "0.8613", "0.2935" ], [ "0.1453", "0.2922", "0.4575", "0.2935", "0.4575", "0.308", "0.1453", "0.3071" ], [ "0.5205", "0.2766", "0.5615", "0.279", "0.561", "0.2944", "0.519", "0.2922" ], [ "0.586", "0.2944", "0.621", "0.2976", "0.6187", "0.3135", "0.5825", "0.31" ], [ "0.86", "0.2944", "0.9146", "0.2964", "0.9136", "0.3123", "0.858", "0.31" ], [ "0.6455", "0.2964", "0.6943", "0.2964", "0.6943", "0.3113", "0.6455", "0.3113" ], [ "0.517", "0.2964", "0.538", "0.2964", "0.538", "0.3113", "0.517", "0.3113" ], [ "0.7173", "0.2964", "0.756", "0.2964", "0.756", "0.3123", "0.7173", "0.3123" ], [ "0.7866", "0.2964", "0.837", "0.2964", "0.837", "0.3113", "0.7866", "0.3113" ], [ "0.1453", "0.308", "0.4592", "0.3088", "0.4592", "0.3225", "0.1453", "0.3218" ], [ "0.517", "0.3154", "0.538", "0.3154", "0.538", "0.3308", "0.517", "0.3308" ], [ "0.6465", "0.3154", "0.6987", "0.3154", "0.6987", "0.33", "0.6465", "0.33" ], [ "0.7886", "0.3154", "0.837", "0.3154", "0.837", "0.33", "0.7886", "0.33" ], [ "0.8613", "0.3154", "0.9146", "0.3154", "0.9146", "0.33", "0.8613", "0.33" ], [ "0.7197", "0.3164", "0.754", "0.3164", "0.754", "0.3289", "0.7197", "0.3289" ], [ "0.1453", "0.3218", "0.4592", "0.3225", "0.4592", "0.3374", "0.1453", "0.3362" ], [ "0.588", "0.3164", "0.6187", "0.3164", "0.6187", "0.3308", "0.588", "0.3308" ], [ "0.5845", "0.332", "0.6206", "0.332", "0.6206", "0.3486", "0.5845", "0.3486" ], [ "0.7173", "0.332", "0.748", "0.332", "0.748", "0.3486", "0.7173", "0.3486" ], [ "0.6465", "0.3333", "0.687", "0.3333", "0.687", "0.3486", "0.6465", "0.3486" ], [ "0.5205", "0.3333", "0.544", "0.3333", "0.544", "0.3474", "0.5205", "0.3474" ], [ "0.7886", "0.3333", "0.837", "0.3333", "0.837", "0.3474", "0.7886", "0.3474" ], [ "0.86", "0.3333", "0.9146", "0.3333", "0.9146", "0.3474", "0.86", "0.3474" ], [ "0.86", "0.3499", "0.902", "0.3516", "0.9", "0.3682", "0.858", "0.3665" ], [ "0.147", "0.3374", "0.4592", "0.3374", "0.4592", "0.3508", "0.147", "0.3508" ], [ "0.147", "0.3508", "0.4575", "0.3508", "0.4575", "0.3652", "0.147", "0.3652" ], [ "0.519", "0.3516", "0.544", "0.3516", "0.544", "0.3665", "0.519", "0.3665" ], [ "0.5845", "0.3508", "0.6187", "0.3508", "0.6187", "0.3677", "0.5845", "0.3677" ], [ "0.6465", "0.3516", "0.69", "0.3516", "0.69", "0.3677", "0.6465", "0.3677" ], [ "0.7188", "0.3528", "0.7573", "0.3528", "0.7573", "0.3652", "0.7188", "0.3652" ], [ "0.7886", "0.3516", "0.8364", "0.3516", "0.8364", "0.3665", "0.7886", "0.3665" ], [ "0.1486", "0.3677", "0.4575", "0.3677", "0.4575", "0.3777", "0.1486", "0.3777" ], [ "0.517", "0.3694", "0.553", "0.3694", "0.553", "0.3855", "0.517", "0.3855" ], [ "0.6465", "0.3694", "0.692", "0.3694", "0.692", "0.3855", "0.6465", "0.3855" ], [ "0.7886", "0.3694", "0.8364", "0.3694", "0.8364", "0.3843", "0.7886", "0.3843" ], [ "0.7188", "0.3706", "0.7466", "0.3706", "0.7466", "0.3843", "0.7188", "0.3843" ], [ "0.586", "0.3706", "0.6113", "0.3706", "0.6113", "0.3855", "0.586", "0.3855" ], [ "0.8623", "0.3718", "0.9", "0.3718", "0.9", "0.3843", "0.8623", "0.3843" ], [ "0.1486", "0.3813", "0.4575", "0.3813", "0.4575", "0.3914", "0.1486", "0.3914" ], [ "0.147", "0.3938", "0.461", "0.3938", "0.461", "0.4067", "0.147", "0.4067" ], [ "0.513", "0.398", "0.881", "0.3984", "0.881", "0.412", "0.513", "0.411" ], [ "0.1486", "0.4092", "0.4575", "0.4092", "0.4575", "0.4192", "0.1486", "0.4192" ], [ "0.5146", "0.4133", "0.882", "0.4133", "0.882", "0.427", "0.5146", "0.427" ], [ "0.147", "0.4216", "0.461", "0.4216", "0.461", "0.4353", "0.147", "0.4353" ], [ "0.511", "0.427", "0.882", "0.4275", "0.882", "0.4412", "0.511", "0.4407" ], [ "0.147", "0.4358", "0.4575", "0.4358", "0.4575", "0.4495", "0.147", "0.4495" ], [ "0.5156", "0.4436", "0.881", "0.4436", "0.881", "0.4536", "0.5156", "0.4536" ], [ "0.147", "0.4507", "0.4592", "0.4507", "0.4592", "0.4644", "0.147", "0.4644" ], [ "0.5146", "0.456", "0.885", "0.456", "0.885", "0.4697", "0.5146", "0.4697" ], [ "0.147", "0.466", "0.199", "0.466", "0.199", "0.4768", "0.147", "0.4768" ], [ "0.2091", "0.4644", "0.4592", "0.4644", "0.4592", "0.478", "0.2091", "0.478" ], [ "0.513", "0.4697", "0.692", "0.4697", "0.692", "0.4827", "0.513", "0.4827" ], [ "0.147", "0.4785", "0.4592", "0.4785", "0.4592", "0.4922", "0.147", "0.4922" ], [ "0.1453", "0.4934", "0.4592", "0.4922", "0.4592", "0.506", "0.1453", "0.5073" ], [ "0.1486", "0.509", "0.4568", "0.509", "0.4568", "0.5195", "0.1486", "0.5195" ], [ "0.147", "0.5215", "0.4592", "0.5215", "0.4592", "0.535", "0.147", "0.535" ], [ "0.5146", "0.5195", "0.882", "0.5195", "0.882", "0.533", "0.5146", "0.533" ], [ "0.2778", "0.5337", "0.4575", "0.536", "0.4575", "0.55", "0.2778", "0.5474" ], [ "0.5156", "0.536", "0.882", "0.536", "0.882", "0.55", "0.5156", "0.55" ], [ "0.1486", "0.55", "0.4575", "0.55", "0.4575", "0.5635", "0.1486", "0.5635" ], [ "0.1486", "0.5376", "0.2854", "0.5376", "0.2854", "0.5474", "0.1486", "0.5474" ], [ "0.5146", "0.5527", "0.882", "0.5527", "0.882", "0.5664", "0.5146", "0.5664" ], [ "0.1486", "0.5635", "0.4575", "0.5635", "0.4575", "0.5767", "0.1486", "0.5767" ], [ "0.5156", "0.5684", "0.881", "0.5684", "0.881", "0.582", "0.5156", "0.582" ], [ "0.147", "0.5776", "0.4575", "0.5776", "0.4575", "0.5913", "0.147", "0.5913" ], [ "0.513", "0.5845", "0.882", "0.583", "0.882", "0.597", "0.513", "0.5986" ], [ "0.147", "0.5913", "0.4575", "0.5913", "0.4575", "0.605", "0.147", "0.605" ], [ "0.5156", "0.601", "0.882", "0.601", "0.882", "0.615", "0.5156", "0.615" ], [ "0.1486", "0.608", "0.4568", "0.608", "0.4568", "0.618", "0.1486", "0.618" ], [ "0.1486", "0.6206", "0.4592", "0.6206", "0.4592", "0.6343", "0.1486", "0.6343" ], [ "0.5146", "0.616", "0.882", "0.615", "0.882", "0.63", "0.5146", "0.6313" ], [ "0.147", "0.6353", "0.4575", "0.6353", "0.4575", "0.6484", "0.147", "0.6484" ], [ "0.5146", "0.633", "0.5903", "0.633", "0.5903", "0.648", "0.5146", "0.648" ], [ "0.147", "0.6494", "0.4592", "0.6494", "0.4592", "0.663", "0.147", "0.663" ], [ "0.531", "0.652", "0.8833", "0.653", "0.8833", "0.6675", "0.531", "0.6665" ], [ "0.1486", "0.663", "0.4592", "0.663", "0.4592", "0.677", "0.1486", "0.677" ], [ "0.5146", "0.6685", "0.882", "0.6694", "0.882", "0.684", "0.5146", "0.683" ], [ "0.147", "0.6777", "0.4592", "0.6777", "0.4592", "0.6914", "0.147", "0.6914" ], [ "0.5146", "0.685", "0.881", "0.685", "0.881", "0.6987", "0.5146", "0.6987" ], [ "0.1486", "0.6934", "0.4575", "0.6934", "0.4575", "0.7036", "0.1486", "0.7036" ], [ "0.1486", "0.706", "0.4575", "0.706", "0.4575", "0.7197", "0.1486", "0.7197" ], [ "0.5146", "0.703", "0.882", "0.703", "0.882", "0.7163", "0.5146", "0.7163" ], [ "0.1486", "0.7217", "0.4575", "0.7217", "0.4575", "0.7324", "0.1486", "0.7324" ], [ "0.5156", "0.7188", "0.881", "0.7188", "0.881", "0.7324", "0.5156", "0.7324" ], [ "0.1486", "0.734", "0.4592", "0.734", "0.4592", "0.7476", "0.1486", "0.7476" ], [ "0.517", "0.7354", "0.882", "0.7354", "0.882", "0.749", "0.517", "0.749" ], [ "0.1453", "0.7476", "0.4114", "0.75", "0.4114", "0.763", "0.1453", "0.761" ], [ "0.517", "0.7505", "0.882", "0.7505", "0.882", "0.764", "0.517", "0.764" ], [ "0.5146", "0.7656", "0.881", "0.7656", "0.881", "0.7793", "0.5146", "0.7793" ], [ "0.5156", "0.783", "0.882", "0.783", "0.882", "0.7964", "0.5156", "0.7964" ], [ "0.12006", "0.7905", "0.1335", "0.7905", "0.1335", "0.8003", "0.12006", "0.8003" ], [ "0.1561", "0.7905", "0.2644", "0.7905", "0.2644", "0.802", "0.1561", "0.802" ], [ "0.5156", "0.799", "0.881", "0.799", "0.881", "0.8125", "0.5156", "0.8125" ], [ "0.12177", "0.814", "0.486", "0.814", "0.486", "0.8276", "0.12177", "0.8276" ], [ "0.511", "0.814", "0.881", "0.814", "0.881", "0.829", "0.511", "0.829" ], [ "0.12006", "0.8306", "0.4878", "0.8306", "0.4878", "0.844", "0.12006", "0.844" ], [ "0.513", "0.8306", "0.881", "0.8306", "0.881", "0.844", "0.513", "0.844" ], [ "0.12006", "0.847", "0.4878", "0.847", "0.4878", "0.861", "0.12006", "0.861" ], [ "0.517", "0.847", "0.882", "0.847", "0.882", "0.861", "0.517", "0.861" ], [ "0.12006", "0.8633", "0.4895", "0.8633", "0.4895", "0.877", "0.12006", "0.877" ], [ "0.5146", "0.8633", "0.881", "0.8633", "0.881", "0.877", "0.5146", "0.877" ], [ "0.12177", "0.879", "0.4895", "0.879", "0.4895", "0.8926", "0.12177", "0.8926" ], [ "0.513", "0.8774", "0.882", "0.879", "0.882", "0.8936", "0.513", "0.8926" ], [ "0.12006", "0.8955", "0.4895", "0.8955", "0.4895", "0.909", "0.12006", "0.909" ], [ "0.5146", "0.8955", "0.679", "0.8955", "0.679", "0.909", "0.5146", "0.909" ] ]

[ "文本分类需要多天的词汇量？", "词汇选择的变分方法", "Wenhu Chent, Yu Su , Yilin Shent, Zhiyu Chent, Xifeng Yant, William Wang", "University of California, Santa Barbara", "Ohio State University, Columbusi", "Samsung Research, Moutain Viewt", "[ wenhuchen,zhiyuchen,xyan,william ↓ @cs.ucsb.edu [email protected] [email protected]", "Abstract", "Cutoff", "Remain", "Vocab", "#Emb", "# Emb", "# CNN", "Vocab", "Ratio", "With the rapid development in deep learn", "Freq.", "60K", "97.6 %", "100 %", "1", "15M", "0.36M", "ing, deep neural networks have been widely", "5", "21.7%", "0.36M", "95.6 %", "10M", "adopted in many real-life natural language", "40K", "24K", "6M", "13 %", "10", "0.36M", "94.3 %", "90%", "applications. Under deep neural networks,", "a pre-defined vocabulary is required to vec-", "20", "14K", "9.4%", "3.5M", "0.36M", "torizetextinputs.Thecanonicalapproach", "100", "2.7 %", "0.36M", "1M", "4K", "73%", "to select pre-defined vocabulary is based", "on the word frequency, where a threshold", "Table 1:Illustration of the frequency-based vocab-", "is selected to cut off the longtail distri-", "ulary selection heuristic on a typical CNN-based", "bution. However, we observed that such a", "document classification model (Section 4.1). #", "simple approach could easily lead to under-", "Embisthenumberofparametersinthewordem-", "sized vocabulary or over-sized vocabulary", "bedding matrix (256 dimensions), and # CNN is", "issues.", "Therefore, we are interested in", "that in the CNN model.", "understanding how the end-task classif-", "cation accuracy is related to the vocab-", "ulary size and what is the minimum re-", "quired vocabulary size to achieve a spe-", "化，而单词嵌入矩阵又假定预定义的词汇集。", "In this paper, we pro-", "对于许多NLP任务，词汇量很容易达到数万", "vide a more sophisticated variational vo-", "cific performance.", "的数量级，这可能会使单词嵌入可训练参数的", "cabulary dropout (VVD) based on varia-", "最大部分。例如，像AG-news （Zhang et al.", "tional dropout to perform vocabulary se-", "2015）这样的文档分类任务可以包含多达60K", "lection, which can intelligently select the", "个唯一单词，嵌人矩阵占97。6可训练参数", "subset of the vocabulary to achieve the re-", "quired performance. To evaluate different", "的%（Table1），这导致神经网络自身参数的", "algorithms on the newly proposed vocab-", "表示不足", "ulary selection problem, we propose two", "从直觉上讲，使用完整或非常大的词汇表既", "new metrics: Area Under Accuracy-Vocab", "不经济，因为它限制了模型在计算或内存受", "Curve and Vocab Size under X % Accuracy", "限情况下的适用性E（Yogatamaetal.，2015", "Drop.Through extensive experiments on", "various NLP classification tasks, our varia-", "Faruquietal.，2015），也不是必需的，因为", "tional framework is shown to significantly", "许多单词可能对最终任务贡献不大，并且可", "outperform the frequency-based and other", "以安全地从词汇表中删除。因此，如何选择最", "selection baselines on these metrics.", "佳词汇是一个理论和实践都感兴趣的问题。有", "点令人惊的是，这个词汇选择问题在文献中", "基本上没有得到解决：事实上的标准做法是进", "1", "Introduction", "行基于频率的截止E(Luongetal.,2015;Kim，", "在过去的十年中，深度神经网络已成为大量自", "2014），并且仅保留比某个值（Table1）更", "然语言处理（NLP）任务中最受欢迎的模型选", "频繁的单词。虽然这个简单的启发式已经证明", "择，并且一直在提供最新的结果。由于神经网", "了强大的经验性能，其任务不可知的性质意味", "络模型假定数据是连续的，因此要在任何文本", "着它可能不是许多任务（或任何任务）的最佳", "数据上应用神经网络，第一步是通过查找操作", "策略。。任务感知的词汇选择策略和不同策略", "使用单词嵌人矩阵对离散文本输入进行向量", "的系统比较仍然缺乏" ]

{ "arxiv_id": "1902.10339", "language": "zh", "page_idx": 0 }

[ [ "0.1368", "0.076", "0.4878", "0.0766", "0.4878", "0.09143", "0.1368", "0.0903" ], [ "0.55", "0.076", "0.6733", "0.0766", "0.6733", "0.09143", "0.55", "0.0903" ], [ "0.12177", "0.0932", "0.4895", "0.0932", "0.4895", "0.1069", "0.12177", "0.1069" ], [ "0.525", "0.0932", "0.882", "0.0932", "0.882", "0.1081", "0.525", "0.1081" ], [ "0.12006", "0.10925", "0.4895", "0.10925", "0.4895", "0.12354", "0.12006", "0.12354" ], [ "0.553", "0.10925", "0.882", "0.10925", "0.882", "0.1223", "0.553", "0.1223" ], [ "0.12006", "0.1259", "0.4878", "0.1259", "0.4878", "0.1395", "0.12006", "0.1395" ], [ "0.552", "0.1259", "0.882", "0.1259", "0.882", "0.1395", "0.552", "0.1395" ], [ "0.12006", "0.1401", "0.486", "0.1401", "0.486", "0.1538", "0.12006", "0.1538" ], [ "0.552", "0.1414", "0.8833", "0.1414", "0.8833", "0.155", "0.552", "0.155" ], [ "0.1184", "0.1567", "0.293", "0.1567", "0.293", "0.1704", "0.1184", "0.1704" ], [ "0.552", "0.1567", "0.8228", "0.1561", "0.8228", "0.1704", "0.552", "0.1716" ], [ "0.1301", "0.1776", "0.486", "0.1799", "0.486", "0.1948", "0.1301", "0.1924" ], [ "0.5264", "0.1729", "0.882", "0.1733", "0.882", "0.1882", "0.5264", "0.187" ], [ "0.55", "0.1882", "0.881", "0.1895", "0.881", "0.2037", "0.55", "0.2031" ], [ "0.1561", "0.1965", "0.4878", "0.1965", "0.4878", "0.2102", "0.1561", "0.2102" ], [ "0.552", "0.206", "0.882", "0.206", "0.882", "0.2197", "0.552", "0.2197" ], [ "0.1561", "0.2131", "0.1973", "0.2131", "0.1973", "0.2245", "0.1561", "0.2245" ], [ "0.552", "0.2214", "0.7866", "0.2214", "0.7866", "0.2363", "0.552", "0.2363" ], [ "0.1293", "0.228", "0.4836", "0.228", "0.4836", "0.2411", "0.1293", "0.2411" ], [ "0.1545", "0.2446", "0.465", "0.2446", "0.465", "0.2583", "0.1545", "0.2583" ], [ "0.5156", "0.2517", "0.529", "0.2517", "0.529", "0.2625", "0.5156", "0.2625" ], [ "0.55", "0.2494", "0.733", "0.2494", "0.733", "0.2642", "0.55", "0.2642" ], [ "0.12006", "0.2673", "0.486", "0.2673", "0.486", "0.2808", "0.12006", "0.2808" ], [ "0.5146", "0.2766", "0.565", "0.2766", "0.565", "0.288", "0.5146", "0.288" ], [ "0.5557", "0.275", "0.7153", "0.2756", "0.7153", "0.2903", "0.5557", "0.2893" ], [ "0.12177", "0.2832", "0.4878", "0.2832", "0.4878", "0.2964", "0.12177", "0.2964" ], [ "0.12177", "0.3005", "0.486", "0.3005", "0.486", "0.3113", "0.12177", "0.3113" ], [ "0.5156", "0.2976", "0.882", "0.2976", "0.882", "0.3113", "0.5156", "0.3113" ], [ "0.12006", "0.3154", "0.4844", "0.3154", "0.4844", "0.326", "0.12006", "0.326" ], [ "0.5146", "0.3123", "0.882", "0.3135", "0.882", "0.328", "0.5146", "0.3267" ], [ "0.12006", "0.33", "0.4878", "0.33", "0.4878", "0.3433", "0.12006", "0.3433" ], [ "0.513", "0.326", "0.8833", "0.328", "0.8833", "0.3457", "0.513", "0.3433" ], [ "0.1184", "0.3467", "0.4895", "0.3457", "0.4895", "0.3599", "0.1184", "0.361" ], [ "0.517", "0.3457", "0.882", "0.3457", "0.882", "0.3594", "0.517", "0.3594" ], [ "0.1184", "0.3623", "0.4895", "0.361", "0.4895", "0.376", "0.1184", "0.3772" ], [ "0.5146", "0.3599", "0.881", "0.3599", "0.881", "0.3735", "0.5146", "0.3735" ], [ "0.12177", "0.379", "0.4878", "0.379", "0.4878", "0.3926", "0.12177", "0.3926" ], [ "0.511", "0.3772", "0.882", "0.3772", "0.882", "0.3901", "0.511", "0.3901" ], [ "0.1184", "0.3943", "0.4878", "0.3943", "0.4878", "0.408", "0.1184", "0.408" ], [ "0.5146", "0.3943", "0.882", "0.3943", "0.882", "0.408", "0.5146", "0.408" ], [ "0.12177", "0.411", "0.4895", "0.4104", "0.4895", "0.4246", "0.12177", "0.4258" ], [ "0.513", "0.4104", "0.881", "0.4104", "0.881", "0.4233", "0.513", "0.4233" ], [ "0.12006", "0.4275", "0.4878", "0.4275", "0.4878", "0.4412", "0.12006", "0.4412" ], [ "0.5156", "0.427", "0.882", "0.427", "0.882", "0.4407", "0.5156", "0.4407" ], [ "0.12006", "0.4436", "0.4895", "0.4436", "0.4895", "0.4573", "0.12006", "0.4573" ], [ "0.5156", "0.4436", "0.882", "0.4436", "0.882", "0.4573", "0.5156", "0.4573" ], [ "0.12006", "0.4602", "0.4878", "0.4602", "0.4878", "0.4739", "0.12006", "0.4739" ], [ "0.519", "0.459", "0.882", "0.459", "0.882", "0.4727", "0.519", "0.4727" ], [ "0.12177", "0.4756", "0.4878", "0.4756", "0.4878", "0.4893", "0.12177", "0.4893" ], [ "0.5146", "0.4744", "0.882", "0.4744", "0.882", "0.488", "0.5146", "0.488" ], [ "0.12177", "0.4922", "0.4878", "0.4922", "0.4878", "0.506", "0.12177", "0.506" ], [ "0.5146", "0.4905", "0.882", "0.4893", "0.882", "0.5044", "0.5146", "0.505" ], [ "0.12006", "0.5083", "0.4878", "0.5083", "0.4878", "0.5215", "0.12006", "0.5215" ], [ "0.5156", "0.5073", "0.8833", "0.5073", "0.8833", "0.521", "0.5156", "0.521" ], [ "0.12006", "0.5254", "0.4844", "0.5254", "0.4844", "0.536", "0.12006", "0.536" ], [ "0.5146", "0.5225", "0.754", "0.5225", "0.754", "0.536", "0.5146", "0.536" ], [ "0.12006", "0.5405", "0.4895", "0.5405", "0.4895", "0.554", "0.12006", "0.554" ], [ "0.534", "0.539", "0.8833", "0.539", "0.8833", "0.5527", "0.534", "0.5527" ], [ "0.12006", "0.557", "0.4902", "0.557", "0.4902", "0.571", "0.12006", "0.571" ], [ "0.5146", "0.5557", "0.882", "0.5557", "0.882", "0.5693", "0.5146", "0.5693" ], [ "0.12006", "0.572", "0.4895", "0.572", "0.4895", "0.586", "0.12006", "0.586" ], [ "0.513", "0.571", "0.879", "0.571", "0.879", "0.5845", "0.513", "0.5845" ], [ "0.12006", "0.5884", "0.4895", "0.5884", "0.4895", "0.6016", "0.12006", "0.6016" ], [ "0.513", "0.586", "0.882", "0.5874", "0.882", "0.6016", "0.513", "0.601" ], [ "0.12006", "0.6025", "0.4895", "0.604", "0.4895", "0.618", "0.12006", "0.6177" ], [ "0.513", "0.6016", "0.5996", "0.604", "0.5996", "0.618", "0.513", "0.616" ], [ "0.12177", "0.6206", "0.3418", "0.6206", "0.3418", "0.6343", "0.12177", "0.6343" ], [ "0.1377", "0.637", "0.4878", "0.637", "0.4878", "0.651", "0.1377", "0.651" ], [ "0.636", "0.633", "0.756", "0.63", "0.756", "0.6445", "0.636", "0.648" ], [ "0.6455", "0.641", "0.6753", "0.6436", "0.6724", "0.6577", "0.642", "0.655" ], [ "0.858", "0.6455", "0.8833", "0.6455", "0.8833", "0.66", "0.858", "0.66" ], [ "0.12006", "0.653", "0.4878", "0.653", "0.4878", "0.6665", "0.12006", "0.6665" ], [ "0.531", "0.66", "0.839", "0.66", "0.839", "0.6733", "0.531", "0.6733" ], [ "0.12006", "0.6694", "0.4878", "0.6694", "0.4878", "0.683", "0.12006", "0.683" ], [ "0.12006", "0.684", "0.486", "0.684", "0.486", "0.6978", "0.12006", "0.6978" ], [ "0.513", "0.6895", "0.882", "0.6904", "0.882", "0.705", "0.513", "0.7036" ], [ "0.1368", "0.7017", "0.4878", "0.7017", "0.4878", "0.7153", "0.1368", "0.7153" ], [ "0.5146", "0.707", "0.882", "0.707", "0.882", "0.72", "0.5146", "0.72" ], [ "0.12177", "0.7188", "0.4878", "0.7188", "0.4878", "0.7324", "0.12177", "0.7324" ], [ "0.513", "0.7227", "0.882", "0.7227", "0.882", "0.7363", "0.513", "0.7363" ], [ "0.12177", "0.734", "0.4878", "0.734", "0.4878", "0.7476", "0.12177", "0.7476" ], [ "0.513", "0.7393", "0.882", "0.7393", "0.882", "0.753", "0.513", "0.753" ], [ "0.12177", "0.7505", "0.4844", "0.7505", "0.4844", "0.764", "0.12177", "0.764" ], [ "0.5146", "0.7544", "0.881", "0.7544", "0.881", "0.767", "0.5146", "0.767" ], [ "0.12006", "0.7666", "0.4878", "0.7666", "0.4878", "0.78", "0.12006", "0.78" ], [ "0.5146", "0.7715", "0.8833", "0.7715", "0.8833", "0.785", "0.5146", "0.785" ], [ "0.1184", "0.781", "0.486", "0.781", "0.486", "0.7944", "0.1184", "0.7944" ], [ "0.5146", "0.7876", "0.882", "0.7876", "0.882", "0.802", "0.5146", "0.802" ], [ "0.1184", "0.7964", "0.4878", "0.7974", "0.4878", "0.8125", "0.1184", "0.811" ], [ "0.5156", "0.804", "0.882", "0.804", "0.882", "0.818", "0.5156", "0.818" ], [ "0.12006", "0.814", "0.4878", "0.814", "0.4878", "0.8276", "0.12006", "0.8276" ], [ "0.5156", "0.8193", "0.882", "0.8193", "0.882", "0.833", "0.5156", "0.833" ], [ "0.12006", "0.8306", "0.3855", "0.8306", "0.3855", "0.844", "0.12006", "0.844" ], [ "0.5156", "0.836", "0.553", "0.836", "0.553", "0.8486", "0.5156", "0.8486" ], [ "0.1318", "0.847", "0.4836", "0.847", "0.4836", "0.861", "0.1318", "0.861" ], [ "0.1528", "0.8623", "0.4878", "0.8633", "0.4878", "0.8774", "0.1528", "0.877" ], [ "0.532", "0.858", "0.7617", "0.86", "0.7617", "0.8745", "0.532", "0.8726" ], [ "0.513", "0.8716", "0.763", "0.8716", "0.763", "0.8853", "0.513", "0.8853" ], [ "0.1561", "0.88", "0.4895", "0.88", "0.4895", "0.8945", "0.1561", "0.8945" ], [ "0.532", "0.883", "0.7617", "0.8843", "0.7617", "0.8984", "0.532", "0.898" ], [ "0.157", "0.898", "0.4878", "0.898", "0.4878", "0.9077", "0.157", "0.9077" ], [ "0.513", "0.898", "0.684", "0.8965", "0.684", "0.9077", "0.513", "0.909" ] ]

[ "在这项工作中，我们对词汇选择问题进行了", "sification tasks.", "首次系统的研究。我们的研究将基于文本分类", "2. We propose a novel vocabulary selection", "任务，一个广泛的NLP任务系列，包括文档", "algorithm based on variational dropout", "分类（DC），自然语言推理（NLI），对话系统", "by re-formulating text classification under", "中的自然语言理解（NLU）等。具体来说，我", "the Bayesian inference framework. The", "们旨在回答以下问题：", "code will be released in Github 1", "1. How important a role does the vocabulary", "3.We conduct comprehensive experiments", "to demonstrate the superiority of the pro-", "selection algorithm play in text classifica-", "posed vocabulary selection algorithm over", "tion?", "a number of strong baselines.", "2. How to dramatically reduce the vocabu", "lary size while retaining the accuracy?", "2", "Vocabulary Selection", "本文的其余部分组织如下：我们首先正式定义", "2.1", "Problem Definition", "了词汇选择问题（subsection2.1），并对具", "有不同词汇选择的分类准确性进行了定量研", "我们现在正式定义问题设置并介绍问题的符", "究，以展示其在最终任务中的重要性（subsec-", "号。通常，我们假设神经分类模型通过嵌入", "tion2.2）。我们还提出了两个用于评估文本", "矩阵WERV*D将离散语言输人向量化为", "分类任务中词汇选择性能的新指标（subsec-", "向量表示，其中V表示词汇表的大小，D表", "tion2.3）。然后，我们提出了一种新颖的，具", "示向量维。嵌人与预定义的单词到索引字典", "有任务意识的词汇选择算法，称为变体词汇", "V=fwi:il<i<Vi相关联，其中wi表", "Dropout（VVD）（section3），该算法借鉴", "示与嵌入矩阵中的 ith行相对应的文字单词。", "了变体dropout(Kingmaetal.，2015）的思", "嵌入矩阵W涵盖了特定NLP任务感兴趣的", "想：如果我们在给定任务的模型训练期间学", "词汇表的子集，请注意，由于人类语言的丰", "习了词汇表V中每个给定单词W的退出概率", "富变化，V的值非常大。在这里，我们展示", "XMATHX结束XX的重要性，。我们建议在", "了Table1中AG-news数据集（Zhangetal.,", "贝叶斯推理框架下推断潜在的辙学概率。在测", "2015）上流行的文本分类模型2的嵌入矩阵大", "试期间，我们通过仅保留丢失概率低于某个阈", "小。从中我们可以很容易地观察到嵌人矩阵通", "值的单词来选择子词汇V。对于使用VVD取", "常占据了大部分参数容量，这可能是许多计算", "消选择的任何单词，我们将简单地将它们视为", "资源有限的实际应用中的瓶颈", "具有空向量表示[0,0,··,0]的特殊标记。请", "为了减轻这种余问题并使嵌入矩阵尽可能", "注意，我们提出的算法需要重新训练单词嵌人", "高效，我们特别感兴趣的是发现最小行大小的", "矩阵，因此它与Word2VecE(Mikolovetal.,", "嵌入W，以实现与使用全行大小的嵌入W", "2013)或GloveE(Pennington etal.,2014)等", "几乎有希望的性能。更正式地说，我们将问题", "预训练单词嵌入的研究相切，尽管我们可以使", "定义如下：", "用它们来初始化我们的嵌入", "我们进行了全面的实验，以评估VVD（sec-", "argmin#Row(W)", "w.", "(1)", "tion4）在不同末端分类任务上的性能。具体", "s.t. Acc(fe(α; W),y) -Acc(fe(a; W),y) ≤e", "来说，我们与一系列强基线选择算法进行了比", "较，包括基于频率的算法E(Luongetal，2015)", "其中#Row是矩阵W中的行数，fe是具", "TF-IDF算法E(Ramosetal.，2003）和结构", "有参数θ的学习神经模型，用于在给定输入", "套索算法E(Friedmanetal.2010），并证明它", "α的情况下预测类别，Acc是测量模型预测和", "可以始终如一地优于这些竞争算法。为了表明", "XMATHX之间。值得注意的是，这里θ包含", "结论被广泛接受，我们的评估基于各种文本分", "了神经网络除嵌人矩阵W之外的所有参数集。", "类任务和具有不同神经网络的数据集，包括卷", "对于每种词汇选择算法A，我们建议绘制其", "积神经网络（CNN)E(Kim，2014），双向长短", "特征曲线Acc(fa(c;W),y)=gA(#Row(W)）", "期记忆(BiLSTM)E(Bahdanau etal.,2014)", "以了解词汇容量与分类准确性之间的关系，我", "和增强的LSTM(ESIM)E(Chenetal.,2017)", "们在整个论文中将其称为（特征）准确性-词汇", "。总之，我们的贡献有三个方面：", "曲线", "1. We formally define the vocabulary se", "lection problem, demonstrate its impor-", "\"https://github.com/wenhuchen/", "Variational-Vocabulary-Selection.git", "tance, and propose new evaluation met-", "2https://github.com/dennybritz/", "rics for vocabulary selection in text clas-", "cnn-text-classification-tf" ]

{ "arxiv_id": "1902.10339", "language": "zh", "page_idx": 1 }

[ [ "0.2301", "0.076", "0.3032", "0.0766", "0.3032", "0.0849", "0.2301", "0.08435" ], [ "0.4358", "0.076", "0.575", "0.0778", "0.575", "0.0849", "0.4358", "0.0831" ], [ "0.705", "0.0748", "0.784", "0.076", "0.784", "0.0861", "0.705", "0.0849" ], [ "0.7188", "0.0944", "0.7446", "0.0944", "0.7446", "0.0986", "0.7188", "0.0986" ], [ "0.2065", "0.0766", "0.2275", "0.0766", "0.2275", "0.08435", "0.2065", "0.08435" ], [ "0.69", "0.0956", "0.708", "0.0956", "0.708", "0.0986", "0.69", "0.0986" ], [ "0.55", "0.1526", "0.571", "0.1526", "0.571", "0.1561", "0.55", "0.1561" ], [ "0.51", "0.1644", "0.5615", "0.1644", "0.5615", "0.1687", "0.51", "0.1687" ], [ "0.12006", "0.1996", "0.881", "0.1996", "0.881", "0.2131", "0.12006", "0.2131" ], [ "0.1184", "0.2144", "0.8833", "0.2144", "0.8833", "0.228", "0.1184", "0.228" ], [ "0.12006", "0.228", "0.2737", "0.228", "0.2737", "0.2411", "0.12006", "0.2411" ], [ "0.1621", "0.2683", "0.4568", "0.2683", "0.4568", "0.282", "0.1621", "0.282" ], [ "0.5156", "0.2683", "0.797", "0.2683", "0.797", "0.282", "0.5156", "0.282" ], [ "0.12006", "0.2915", "0.4878", "0.2915", "0.4878", "0.306", "0.12006", "0.306" ], [ "0.12006", "0.2695", "0.1696", "0.2695", "0.1696", "0.2808", "0.12006", "0.2808" ], [ "0.644", "0.2944", "0.806", "0.2944", "0.806", "0.3088", "0.644", "0.3088" ], [ "0.588", "0.3005", "0.6514", "0.2986", "0.6533", "0.3135", "0.5884", "0.3154" ], [ "0.858", "0.3018", "0.8833", "0.3018", "0.8833", "0.3164", "0.858", "0.3164" ], [ "0.12006", "0.3088", "0.486", "0.3088", "0.486", "0.3225", "0.12006", "0.3225" ], [ "0.6616", "0.3113", "0.7925", "0.31", "0.7925", "0.3247", "0.6616", "0.326" ], [ "0.12177", "0.3247", "0.4878", "0.3247", "0.4878", "0.3384", "0.12177", "0.3384" ], [ "0.1184", "0.339", "0.486", "0.339", "0.486", "0.3528", "0.1184", "0.3528" ], [ "0.534", "0.3362", "0.882", "0.3362", "0.882", "0.3499", "0.534", "0.3499" ], [ "0.12006", "0.357", "0.4878", "0.357", "0.4878", "0.3706", "0.12006", "0.3706" ], [ "0.5156", "0.3516", "0.882", "0.3516", "0.882", "0.3652", "0.5156", "0.3652" ], [ "0.5146", "0.3677", "0.881", "0.3677", "0.881", "0.3813", "0.5146", "0.3813" ], [ "0.1184", "0.3735", "0.4844", "0.3723", "0.4844", "0.3872", "0.1184", "0.3884" ], [ "0.4802", "0.3777", "0.4844", "0.3777", "0.4844", "0.383", "0.4802", "0.383" ], [ "0.5146", "0.3843", "0.6753", "0.3843", "0.6753", "0.398", "0.5146", "0.398" ], [ "0.12006", "0.3901", "0.4878", "0.3896", "0.4878", "0.4038", "0.12006", "0.405" ], [ "0.12177", "0.4067", "0.4878", "0.4067", "0.4878", "0.4204", "0.12177", "0.4204" ], [ "0.5146", "0.4133", "0.5283", "0.4133", "0.5283", "0.4216", "0.5146", "0.4216" ], [ "0.552", "0.412", "0.66", "0.412", "0.66", "0.4233", "0.552", "0.4233" ], [ "0.12006", "0.4229", "0.4878", "0.4229", "0.4878", "0.4358", "0.12006", "0.4358" ], [ "0.1184", "0.4382", "0.4878", "0.437", "0.4878", "0.452", "0.1184", "0.4531" ], [ "0.5146", "0.4353", "0.882", "0.4353", "0.882", "0.4495", "0.5146", "0.4495" ], [ "0.12177", "0.4548", "0.4878", "0.4548", "0.4878", "0.4697", "0.12177", "0.4697" ], [ "0.5146", "0.452", "0.882", "0.452", "0.882", "0.466", "0.5146", "0.466" ], [ "0.12177", "0.4702", "0.4895", "0.4702", "0.4895", "0.485", "0.12177", "0.485" ], [ "0.5156", "0.4685", "0.882", "0.4685", "0.882", "0.4822", "0.5156", "0.4822" ], [ "0.12006", "0.4868", "0.4878", "0.4868", "0.4878", "0.5005", "0.12006", "0.5005" ], [ "0.5156", "0.484", "0.882", "0.484", "0.882", "0.4976", "0.5156", "0.4976" ], [ "0.12177", "0.503", "0.4878", "0.503", "0.4878", "0.5166", "0.12177", "0.5166" ], [ "0.5146", "0.5005", "0.881", "0.5005", "0.881", "0.514", "0.5146", "0.514" ], [ "0.12006", "0.5186", "0.486", "0.5186", "0.486", "0.532", "0.12006", "0.532" ], [ "0.5156", "0.5166", "0.882", "0.5166", "0.882", "0.53", "0.5156", "0.53" ], [ "0.12177", "0.535", "0.4844", "0.535", "0.4844", "0.549", "0.12177", "0.549" ], [ "0.5156", "0.533", "0.881", "0.533", "0.881", "0.5464", "0.5156", "0.5464" ], [ "0.12006", "0.552", "0.4844", "0.552", "0.4844", "0.5625", "0.12006", "0.5625" ], [ "0.517", "0.5503", "0.881", "0.5503", "0.881", "0.561", "0.517", "0.561" ], [ "0.12006", "0.5684", "0.4844", "0.5684", "0.4844", "0.579", "0.12006", "0.579" ], [ "0.5156", "0.5654", "0.882", "0.5654", "0.882", "0.58", "0.5156", "0.58" ], [ "0.12006", "0.5845", "0.4878", "0.5845", "0.4878", "0.597", "0.12006", "0.597" ], [ "0.5156", "0.5806", "0.882", "0.5806", "0.882", "0.594", "0.5156", "0.594" ], [ "0.12177", "0.6016", "0.4844", "0.6016", "0.4844", "0.6123", "0.12177", "0.6123" ], [ "0.5156", "0.597", "0.882", "0.597", "0.882", "0.611", "0.5156", "0.611" ], [ "0.12177", "0.615", "0.4878", "0.615", "0.4878", "0.63", "0.12177", "0.63" ], [ "0.5156", "0.6133", "0.881", "0.6133", "0.881", "0.627", "0.5156", "0.627" ], [ "0.1184", "0.6313", "0.3955", "0.632", "0.3955", "0.6465", "0.1184", "0.6455" ], [ "0.517", "0.629", "0.881", "0.629", "0.881", "0.6426", "0.517", "0.6426" ], [ "0.5156", "0.6436", "0.882", "0.6445", "0.882", "0.659", "0.5156", "0.6577" ], [ "0.12006", "0.663", "0.1696", "0.663", "0.1696", "0.675", "0.12006", "0.675" ], [ "0.1628", "0.663", "0.3208", "0.663", "0.3208", "0.6733", "0.1628", "0.6733" ], [ "0.5146", "0.6577", "0.6094", "0.66", "0.6094", "0.676", "0.5146", "0.6733" ], [ "0.12006", "0.686", "0.4895", "0.686", "0.4895", "0.6997", "0.12006", "0.6997" ], [ "0.5283", "0.6953", "0.5786", "0.6953", "0.5786", "0.707", "0.5283", "0.707" ], [ "0.733", "0.6987", "0.748", "0.6987", "0.748", "0.703", "0.733", "0.703" ], [ "0.525", "0.709", "0.5728", "0.707", "0.574", "0.7227", "0.5264", "0.7246" ], [ "0.1184", "0.6997", "0.4878", "0.7017", "0.4878", "0.7163", "0.1184", "0.7144" ], [ "0.1234", "0.7188", "0.4844", "0.7188", "0.4844", "0.7324", "0.1234", "0.7324" ], [ "0.729", "0.7144", "0.75", "0.7144", "0.75", "0.7217", "0.729", "0.7217" ], [ "0.812", "0.72", "0.839", "0.72", "0.839", "0.728", "0.812", "0.728" ], [ "0.8364", "0.7217", "0.858", "0.7217", "0.858", "0.7285", "0.8364", "0.7285" ], [ "0.797", "0.7227", "0.8096", "0.7227", "0.8096", "0.727", "0.797", "0.727" ], [ "0.531", "0.727", "0.58", "0.727", "0.58", "0.737", "0.531", "0.737" ], [ "0.6206", "0.727", "0.6274", "0.731", "0.621", "0.7363", "0.614", "0.7324" ], [ "0.12177", "0.734", "0.4878", "0.734", "0.4878", "0.7476", "0.12177", "0.7476" ], [ "0.529", "0.741", "0.5728", "0.7393", "0.574", "0.7505", "0.531", "0.753" ], [ "0.6987", "0.742", "0.7515", "0.742", "0.7515", "0.75", "0.6987", "0.75" ], [ "0.771", "0.742", "0.796", "0.742", "0.796", "0.75", "0.771", "0.75" ], [ "0.6724", "0.7437", "0.6978", "0.7437", "0.6978", "0.7505", "0.6724", "0.7505" ], [ "0.657", "0.7437", "0.6753", "0.7437", "0.6753", "0.7505", "0.657", "0.7505" ], [ "0.8076", "0.7437", "0.8564", "0.7437", "0.8564", "0.7476", "0.8076", "0.7476" ], [ "0.614", "0.745", "0.629", "0.745", "0.629", "0.75", "0.614", "0.75" ], [ "0.12177", "0.7505", "0.4844", "0.7505", "0.4844", "0.764", "0.12177", "0.764" ], [ "0.5996", "0.745", "0.607", "0.745", "0.607", "0.749", "0.5996", "0.749" ], [ "0.12177", "0.7666", "0.4844", "0.7666", "0.4844", "0.78", "0.12177", "0.78" ], [ "0.5156", "0.7754", "0.5938", "0.7754", "0.5938", "0.7876", "0.5156", "0.7876" ], [ "0.602", "0.775", "0.882", "0.775", "0.882", "0.788", "0.602", "0.788" ], [ "0.12006", "0.782", "0.4878", "0.781", "0.4878", "0.796", "0.12006", "0.7964" ], [ "0.513", "0.789", "0.8833", "0.789", "0.8833", "0.8027", "0.513", "0.8027" ], [ "0.12006", "0.7974", "0.4878", "0.7974", "0.4878", "0.8125", "0.12006", "0.8125" ], [ "0.513", "0.804", "0.882", "0.804", "0.882", "0.818", "0.513", "0.818" ], [ "0.12006", "0.814", "0.4878", "0.814", "0.4878", "0.8276", "0.12006", "0.8276" ], [ "0.511", "0.8154", "0.601", "0.818", "0.601", "0.832", "0.511", "0.83" ], [ "0.12006", "0.8306", "0.4878", "0.8306", "0.4878", "0.844", "0.12006", "0.844" ], [ "0.1184", "0.8457", "0.4878", "0.8467", "0.4878", "0.861", "0.1184", "0.86" ], [ "0.513", "0.8594", "0.5474", "0.8594", "0.5474", "0.8706", "0.513", "0.8706" ], [ "0.544", "0.861", "0.561", "0.861", "0.561", "0.868", "0.544", "0.868" ], [ "0.561", "0.858", "0.705", "0.858", "0.705", "0.8716", "0.561", "0.8716" ], [ "0.12177", "0.879", "0.4878", "0.879", "0.4878", "0.8926", "0.12177", "0.8926" ], [ "0.12006", "0.8623", "0.4895", "0.8623", "0.4895", "0.877", "0.12006", "0.877" ], [ "0.5146", "0.879", "0.881", "0.879", "0.881", "0.8926", "0.5146", "0.8926" ], [ "0.12006", "0.8955", "0.4895", "0.8955", "0.4895", "0.909", "0.12006", "0.909" ], [ "0.5156", "0.8955", "0.8833", "0.8955", "0.8833", "0.909", "0.5156", "0.909" ] ]

[ "nentClassificatior", "NaturallanguageUnderstanding", "Evaluation Metrics", "eaho.", "Docu", "", "racv", "Accuracv.Banr", "Figure 1: Monte-Carlo simulation on vocabulary selection. Left: CNN-based document classification", "on AG-news dataset. Middle: Natural language understanding on Snips dataset. Right: Metrics for", "vocabulary selection.", "Importance of Vocabulary Selection", "过以下方式计算归一化的封闭区域：", "为了研究词汇选择算法所起作用的重要性，我", "2.2", "Jv Acc(log(V)d log(V)", "AUC=", "(2)", "们设计了一种蒙特卡洛模拟策略，以近似估计", "Je Acc(V)d log(V)", "可能的选择算法A所达到的给定词汇大小的", "准确性下限和上限。更具体地说，对于√的", "值得注意的是，Vocab@-X%从范围[O,V", "给定词汇量，存在（）算法，该算法可以从", "中取值较小的值表示性能更好。由于AUC由", "Acc（V）归一化，因此无论分类错误如何，它", "完整词汇表V中选择不同的词汇表子集立", "", "都从范围[0,1]取值", "直接列举这些可能性是不可能的，我们建议使", "用蒙特卡洛词汇选择策略，该策略可以随机选", "3", "Our Method", "择词汇子集V，通过运行N次来模拟可能的", "选择算法。仿真后，我们在每个给定的√处", "受 DNNdropoutE(Srivastava et al.,2014", "获得了各种点估计（Acc1，·，AccNV），并", "WangandManning，2013）的启发，我们建", "在EFigure1中描绘了点估计，以近似可视化", "议从单词级dropout的角度解决词汇选择问", "精度-vocab曲线的上限和下限。从EFigure1", "题，其中我们假设每个单词wi（整数索引）与", "中，我们可以很容易地观察到，在有限的词汇", "其特征丢失率Pi相关联，代表被空占位符替", "量下精度范围极大，当预算√增加时，差距", "换的可能性，具体来说，较高的丢失概率表示", "逐渐缩小。例如，对于预算为1000的文档分", "从词汇表中删除它所遭受的损失较少。因此", "类，选择算法A可以产生从42.5到85.1的", "可以将Equation1中的原始优化问题视为推", "潜在准确性，而对干预算为27的自然语言理", "断潜在的丢失概率向量P=[P1，··，Pv]。我", "解任务，选择算法A可以产生从33.2到80.1", "们的哲学概述在Figure2中进行了描述，我", "的潜在准确性。这样的蒙特卡洛模拟研究证明", "们将嵌人矩阵的每一行与一个丢失概率相关", "了NLP任务中词汇选择策略的重要性，也暗", "联，然后重新训练整个系统，该系统掌握词汇", "示了最佳词汇选择算法的巨大潜力", "表中每个单词对最终NLP任务的贡献，并从", "词汇表中删除那些“贡献较小”的单词，而不", "2.3", "Evaluation Metrics", "会影响性能", "为了评估给定的选择算法A的性能如何，我", "XWhat", "1R0", "×New", "们通过定量研究其特征精度-vocab曲线，提出", "了EFigure1中描述的评估指标。这些指标", "0.98", "", "", "", "Genre", "福", "即曲线下面积（AUC）和Vocab@-X%分别", "Band", "P4=0.20", "", "", "", "", "", "测量全局和局部的词汇选择性能。具体来说", "", "AUC通过曲线计算封闭区域，该曲线概述了", "Figure 2:", "Variational dropout in classification", "词汇选择算法的执行情况。相比之下，如果允", "models, \"New\" and \"What\" can be safely removed", "许X%性能下降，Vocab@-X%会计算所需的", "without harming performance due to large dropout", "最小词汇量，这直接表示实现给定精度所需", "probability.", "的词汇量。对于本地评估指标，我们主要考虑", "Vocab@-3%和Vocab@-5%。但是，我们观察", "3.1", "", "Bernouli Dropout", "此无法表示低词汇量条件下算法的选择能力。", "到直接计算AUC过于强调大词汇量区域，因", "这里我们首先假设神经网络使用嵌入矩阵W", "因此，我们建议取词汇表大小的对数，然后通", "对离散输人进行失量化，以将给定的单词&" ]

{ "arxiv_id": "1902.10339", "language": "zh", "page_idx": 2 }

[ [ "0.12006", "0.0766", "0.4895", "0.0766", "0.4895", "0.0903", "0.12006", "0.0903" ], [ "0.4878", "0.0778", "0.881", "0.0778", "0.881", "0.09143", "0.4878", "0.09143" ], [ "0.12006", "0.0932", "0.4878", "0.0932", "0.4878", "0.1069", "0.12006", "0.1069" ], [ "0.4963", "0.0986", "0.5054", "0.0968", "0.5073", "0.0998", "0.4978", "0.10156" ], [ "0.5146", "0.0944", "0.614", "0.0944", "0.614", "0.1057", "0.5146", "0.1057" ], [ "0.1184", "0.10925", "0.4836", "0.10925", "0.4836", "0.1223", "0.1184", "0.1223" ], [ "0.12177", "0.1259", "0.4878", "0.1259", "0.4878", "0.1395", "0.12177", "0.1395" ], [ "0.5576", "0.12054", "0.8867", "0.12354", "0.8867", "0.1478", "0.5576", "0.1443" ], [ "0.12177", "0.1414", "0.4878", "0.1414", "0.4878", "0.155", "0.12177", "0.155" ], [ "0.1184", "0.1561", "0.486", "0.1561", "0.486", "0.1704", "0.1184", "0.1704" ], [ "0.5146", "0.1567", "0.882", "0.158", "0.882", "0.1729", "0.5146", "0.1716" ], [ "0.12006", "0.1733", "0.4895", "0.1733", "0.4895", "0.187", "0.12006", "0.187" ], [ "0.5146", "0.1733", "0.881", "0.1733", "0.881", "0.187", "0.5146", "0.187" ], [ "0.12006", "0.1907", "0.4878", "0.1907", "0.4878", "0.2037", "0.12006", "0.2037" ], [ "0.5146", "0.1907", "0.763", "0.1907", "0.763", "0.2037", "0.5146", "0.2037" ], [ "0.12177", "0.206", "0.4878", "0.206", "0.4878", "0.2197", "0.12177", "0.2197" ], [ "0.12006", "0.2203", "0.4878", "0.2203", "0.4878", "0.2351", "0.12006", "0.2351" ], [ "0.5356", "0.2203", "0.748", "0.2227", "0.748", "0.2369", "0.5356", "0.2351" ], [ "0.742", "0.2197", "0.86", "0.2173", "0.86", "0.2351", "0.742", "0.2369" ], [ "0.12006", "0.2382", "0.4878", "0.2382", "0.4878", "0.2517", "0.12006", "0.2517" ], [ "0.552", "0.2405", "0.8447", "0.2411", "0.8447", "0.2559", "0.552", "0.2546" ], [ "0.1184", "0.253", "0.461", "0.2542", "0.461", "0.2683", "0.1184", "0.2673" ], [ "0.858", "0.257", "0.882", "0.257", "0.882", "0.2708", "0.858", "0.2708" ], [ "0.12006", "0.2832", "0.3342", "0.2832", "0.3342", "0.2964", "0.12006", "0.2964" ], [ "0.5146", "0.284", "0.881", "0.284", "0.881", "0.2976", "0.5146", "0.2976" ], [ "0.12177", "0.3047", "0.4895", "0.3047", "0.4895", "0.3184", "0.12177", "0.3184" ], [ "0.5156", "0.3018", "0.882", "0.3018", "0.882", "0.3154", "0.5156", "0.3154" ], [ "0.12177", "0.3196", "0.4895", "0.3196", "0.4895", "0.3342", "0.12177", "0.3342" ], [ "0.5146", "0.3164", "0.881", "0.3164", "0.881", "0.33", "0.5146", "0.33" ], [ "0.1184", "0.335", "0.4878", "0.3362", "0.4878", "0.3508", "0.1184", "0.3499" ], [ "0.5156", "0.3333", "0.882", "0.3333", "0.882", "0.3467", "0.5156", "0.3467" ], [ "0.12177", "0.354", "0.486", "0.354", "0.486", "0.3677", "0.12177", "0.3677" ], [ "0.5156", "0.3499", "0.8774", "0.3499", "0.8774", "0.3599", "0.5156", "0.3599" ], [ "0.5156", "0.3635", "0.881", "0.3652", "0.881", "0.3801", "0.5156", "0.3777" ], [ "0.12006", "0.3694", "0.4148", "0.3694", "0.4148", "0.383", "0.12006", "0.383" ], [ "0.5146", "0.3813", "0.881", "0.3813", "0.881", "0.3943", "0.5146", "0.3943" ], [ "0.2024", "0.402", "0.4106", "0.402", "0.4106", "0.412", "0.2024", "0.412" ], [ "0.461", "0.402", "0.4895", "0.402", "0.4895", "0.4146", "0.461", "0.4146" ], [ "0.513", "0.3967", "0.7446", "0.398", "0.7446", "0.412", "0.513", "0.411" ], [ "0.419", "0.4038", "0.4534", "0.4038", "0.4534", "0.411", "0.419", "0.411" ], [ "0.11755", "0.4287", "0.4785", "0.43", "0.4785", "0.4478", "0.11755", "0.4465" ], [ "0.5146", "0.4246", "0.7314", "0.4246", "0.7314", "0.4382", "0.5146", "0.4382" ], [ "0.12006", "0.4478", "0.4954", "0.4448", "0.4954", "0.462", "0.12006", "0.4656" ], [ "0.5156", "0.4453", "0.881", "0.4453", "0.881", "0.459", "0.5156", "0.459" ], [ "0.1184", "0.4656", "0.4895", "0.4656", "0.4895", "0.4785", "0.1184", "0.4785" ], [ "0.5146", "0.4614", "0.882", "0.4614", "0.882", "0.4744", "0.5146", "0.4744" ], [ "0.12006", "0.4822", "0.4895", "0.4822", "0.4895", "0.4954", "0.12006", "0.4954" ], [ "0.5146", "0.478", "0.882", "0.478", "0.882", "0.4912", "0.5146", "0.4912" ], [ "0.12006", "0.4976", "0.4895", "0.4976", "0.4895", "0.511", "0.12006", "0.511" ], [ "0.5146", "0.4946", "0.559", "0.4946", "0.559", "0.506", "0.5146", "0.506" ], [ "0.1184", "0.514", "0.4895", "0.513", "0.4895", "0.528", "0.1184", "0.5293" ], [ "0.6187", "0.5225", "0.777", "0.524", "0.777", "0.538", "0.6187", "0.5376" ], [ "0.1184", "0.5293", "0.4878", "0.53", "0.4878", "0.5444", "0.1184", "0.5435" ], [ "0.858", "0.524", "0.882", "0.524", "0.882", "0.5376", "0.858", "0.5376" ], [ "0.1184", "0.5444", "0.4878", "0.546", "0.4878", "0.56", "0.1184", "0.5586" ], [ "0.513", "0.552", "0.882", "0.5527", "0.882", "0.568", "0.513", "0.5664" ], [ "0.12006", "0.5625", "0.4895", "0.5625", "0.4895", "0.576", "0.12006", "0.576" ], [ "0.5146", "0.568", "0.8613", "0.5684", "0.8613", "0.583", "0.5146", "0.582" ], [ "0.1184", "0.5776", "0.2377", "0.5776", "0.2377", "0.5913", "0.1184", "0.5913" ], [ "0.511", "0.5957", "0.6616", "0.597", "0.6616", "0.6123", "0.511", "0.61" ], [ "0.2225", "0.611", "0.3862", "0.611", "0.3862", "0.6216", "0.2225", "0.6216" ], [ "0.4626", "0.611", "0.4878", "0.611", "0.4878", "0.6235", "0.4626", "0.6235" ], [ "0.5146", "0.6226", "0.882", "0.6226", "0.882", "0.6357", "0.5146", "0.6357" ], [ "0.1184", "0.637", "0.4895", "0.637", "0.4895", "0.655", "0.1184", "0.655" ], [ "0.5146", "0.6396", "0.846", "0.6396", "0.846", "0.653", "0.5146", "0.653" ], [ "0.1184", "0.6567", "0.4895", "0.6567", "0.4895", "0.6714", "0.1184", "0.6714" ], [ "0.513", "0.663", "0.7676", "0.6646", "0.7676", "0.6787", "0.513", "0.6777" ], [ "0.12006", "0.6733", "0.4895", "0.6733", "0.4895", "0.687", "0.12006", "0.687" ], [ "0.12006", "0.6904", "0.3284", "0.6904", "0.3284", "0.7036", "0.12006", "0.7036" ], [ "0.5156", "0.686", "0.882", "0.686", "0.882", "0.6997", "0.5156", "0.6997" ], [ "0.517", "0.703", "0.8774", "0.703", "0.8774", "0.7163", "0.517", "0.7163" ], [ "0.1335", "0.7227", "0.2805", "0.7227", "0.2805", "0.7363", "0.1335", "0.7363" ], [ "0.293", "0.7227", "0.476", "0.7227", "0.476", "0.737", "0.293", "0.737" ], [ "0.5156", "0.7188", "0.882", "0.7188", "0.882", "0.7324", "0.5156", "0.7324" ], [ "0.5156", "0.7354", "0.881", "0.7354", "0.881", "0.749", "0.5156", "0.749" ], [ "0.1184", "0.7466", "0.4895", "0.742", "0.4895", "0.7627", "0.1184", "0.7666" ], [ "0.5156", "0.7505", "0.882", "0.7505", "0.882", "0.764", "0.5156", "0.764" ], [ "0.1184", "0.7695", "0.1838", "0.7695", "0.1838", "0.7837", "0.1184", "0.7837" ], [ "0.5156", "0.767", "0.882", "0.767", "0.882", "0.781", "0.5156", "0.781" ], [ "0.5156", "0.783", "0.881", "0.783", "0.881", "0.7964", "0.5156", "0.7964" ], [ "0.1184", "0.799", "0.4878", "0.799", "0.4878", "0.8125", "0.1184", "0.8125" ], [ "0.513", "0.799", "0.881", "0.799", "0.881", "0.8125", "0.513", "0.8125" ], [ "0.12006", "0.8154", "0.4895", "0.8154", "0.4895", "0.829", "0.12006", "0.829" ], [ "0.5156", "0.8164", "0.881", "0.8164", "0.881", "0.8267", "0.5156", "0.8267" ], [ "0.1184", "0.83", "0.4895", "0.8306", "0.4895", "0.8457", "0.1184", "0.844" ], [ "0.5146", "0.832", "0.882", "0.832", "0.882", "0.8457", "0.5146", "0.8457" ], [ "0.12006", "0.847", "0.4895", "0.847", "0.4895", "0.861", "0.12006", "0.861" ], [ "0.513", "0.847", "0.882", "0.847", "0.882", "0.861", "0.513", "0.861" ], [ "0.12006", "0.8633", "0.4895", "0.8633", "0.4895", "0.877", "0.12006", "0.877" ], [ "0.4963", "0.8677", "0.5146", "0.8677", "0.5146", "0.8716", "0.4963", "0.8716" ], [ "0.511", "0.8623", "0.882", "0.8633", "0.882", "0.8774", "0.511", "0.877" ], [ "0.12006", "0.8774", "0.4895", "0.879", "0.4895", "0.8936", "0.12006", "0.8926" ], [ "0.4978", "0.8843", "0.511", "0.8843", "0.511", "0.887", "0.4978", "0.887" ], [ "0.5146", "0.88", "0.882", "0.88", "0.882", "0.8936", "0.5146", "0.8936" ], [ "0.12006", "0.8945", "0.4902", "0.8955", "0.4902", "0.91", "0.12006", "0.909" ], [ "0.4937", "0.8955", "0.882", "0.8955", "0.882", "0.909", "0.4937", "0.909" ] ]

[ "投影到向量空间RD中，然后建议将随机丢", "α，i。e。与μ无关。形式上，我们将先验分", "失噪声添加到嵌入输入中，以模拟如下所示", "", "布写成如下：", "的丢失过程：。在采样的Bernouli向量b下", "使用给定的嵌人矩阵W计算嵌入输出向量", "P(log|B=l) = const→P(Bul) αB_ (5)", "E(cb)。为了在贝叶斯框架下推断具有参数", "p的潜在Bernouli 分布，其中给出了训练对", "由于这种KL散度不存在封闭形式的表达式，", "(α=ai·.an，y）作为证据，我们首先将目标", "因此我们遵循ELouizosetal.（2017）通过以", "函数定义为C(fe(a),y），然后得出其下界如", "下最小方差公式对其进行近似：", "下（使用p=1一p):。这里我们使用E(α）作", "为{E(ai)，·，E(αn))的简化形式，我们将文", "DkL =-kio(k2 + ks loga) +", " log(1 + ) + ki", "本分类模型的参数θ与嵌人参数W分开，并", "ki=0.63576k2=1.87320ks=1.48695", "假设分类模型fe直接将嵌人E作为输入。", "(6)", "3.2Gaussian Relaxation", "通过采用不适当的对数均匀先验，更多的权重", "但是，Bernouli分布很难重新参数化，我们", "被压缩为零，并且KL-散度与较学率α负相", "需要枚举2V个不同的值来计算随机丢失向", "关。直观上，丢失率αi是词汇表中ith单词", "量b上的期望值。因此，我们遵循EWang", "的穴余指标，较大的Qi表示因删除ith单词", "andManning（2013）使用连续高斯近似，其中", "而导致的性能损失较小。在训练过程中，我们", "使用重新参数化技巧E(KingmaandWelling，", "Bernouli噪声b被高斯噪声z代替：", "2013）从正态分布中采样嵌入权重，以减少贝", "E(αz)= (z@ OneHot(r))-W", "(3)", "叶斯训练中的蒙特卡洛方差。", "", "其中RV服从高斯分布～N(1,i", "3.3Vocabulary Selection", "-p）。值得注意的是，α和p是一一对应的，", "优化后，我们可以获得与每个单词wi关联的", "Q是P的单调递增函数。有关更多详细信息，", "退出率αi。我们建议使用阈值QT根据较学", "请参阅WangandManning（2013）。基于这", "率选择词汇子集。因此，剩余的词汇子集描述", "种近似，我们可以使用α作为退出标准，例如。", "如下：", "g。扔掉α高于某个给定值QT的单词。我", "V= (wiE Vlai<ar)", "们进一步遵循ELouizosetal.（2017)；Kingma", "(7)", "et al.(2015);Molchanov etal.(2017)将输人", "其中我们使用来表示感兴趣的子集词汇，通", "噪声重新解释为嵌入权重B本身的内在随机", "过调整QT，我们能够控制所选词汇的大小", "性，如下所示：", "4Experiments", "E(z) =OneHot(r)B", "(4)", "我们将所提出的词汇选择算法与多种文本分类", "其中BERV*D遵循多变量高斯分布Bii~", "任务和数据集上的几种强基线进行了比较", "N(uii=Wi,o,=aiWa），其中每行中的随", "4.1Datasets & Architectures", "机权重具有固定的方差/均值比MAX9。因此，", "我们将证据下界重写如下：", "我们正在使用的主要数据集在ETable 2中列", "出，它提供了其描述和容量的概述。具体来说", "log C(fe(α), y) = log", "C(fe(E(az),y)P(B)dB", "我们按照（Zhangetal.,2015;Gooetal.,2018;", "Williamsetal.，2018）对文档分类数据集、自", "≥B~μ,。)[log C(fo(E(a|2),y)] - KL(N(u,0)P(B))", "然语言理解数据集和自然语言推理数据集进行", "=C(B,0)", "预处理。我们完全复制他们的实验设置，使我", "们的方法与他们的方法相当。我们的模型是用", "其中P(B）是先验分布，N(μu,α）表示具有参", "TensorFlow（Abadietal.,2015）实现的为了", "数μ和α的高斯近似后验。L(B,の)被用作", "评估VVD选择算法在深度学习体系结构中的", "边际对数似然logC(fe(α),y)）的松弛证据下", "泛化能力，我们研究了其在不同的已建立体系", "界。在这里，我们遵循EKingmaetal.（2015)；", "结构下的性能（在EFigure3中进行了描述）。", "Louizosetal.（2017）选择先验分布P(B）作", "", "在自然语言理解中，我们使用最新的基于注", "为“不适当的对数缩放均匀分布”，以保证正", "", "意力的模型进行意图跟踪E(Go0etal.，2018)", "则化项DKL(N(μ,)I/P(B)）仅取决于丢失率", "，该模型首先使用BLSTM递归网络利用从" ]

{ "arxiv_id": "1902.10339", "language": "zh", "page_idx": 3 }

[ [ "0.126", "0.0748", "0.1914", "0.0766", "0.1897", "0.09143", "0.12427", "0.08905" ], [ "0.419", "0.076", "0.504", "0.0778", "0.504", "0.09265", "0.419", "0.0903" ], [ "0.7314", "0.0766", "0.79", "0.0766", "0.79", "0.09143", "0.7314", "0.09143" ], [ "0.802", "0.0766", "0.8623", "0.0748", "0.864", "0.08905", "0.8037", "0.09143" ], [ "0.8774", "0.0766", "0.9326", "0.0766", "0.9326", "0.09143", "0.8774", "0.09143" ], [ "0.1276", "0.0932", "0.3267", "0.0932", "0.3267", "0.1069", "0.1276", "0.1069" ], [ "0.3687", "0.0778", "0.4072", "0.0778", "0.4072", "0.08905", "0.3687", "0.08905" ], [ "0.4224", "0.0944", "0.6245", "0.0956", "0.6245", "0.10925", "0.4224", "0.1081" ], [ "0.7305", "0.0956", "0.7515", "0.0956", "0.7515", "0.10925", "0.7305", "0.10925" ], [ "0.806", "0.0956", "0.8447", "0.0956", "0.8447", "0.1081", "0.806", "0.1081" ], [ "0.879", "0.0956", "0.9077", "0.0956", "0.9077", "0.1081", "0.879", "0.1081" ], [ "0.3687", "0.10156", "0.4055", "0.10156", "0.4055", "0.1134", "0.3687", "0.1134" ], [ "0.4224", "0.1057", "0.7095", "0.1069", "0.7095", "0.12054", "0.4224", "0.1194" ], [ "0.806", "0.1081", "0.8506", "0.1081", "0.8506", "0.1194", "0.806", "0.1194" ], [ "0.879", "0.1081", "0.9077", "0.1081", "0.9077", "0.12054", "0.879", "0.12054" ], [ "0.7314", "0.10925", "0.7446", "0.10925", "0.7446", "0.1194", "0.7314", "0.1194" ], [ "0.126", "0.1069", "0.3123", "0.1069", "0.3123", "0.12054", "0.126", "0.12054" ], [ "0.126", "0.12354", "0.3333", "0.1247", "0.3333", "0.1395", "0.126", "0.1383" ], [ "0.4224", "0.1247", "0.6323", "0.12354", "0.6323", "0.1383", "0.4224", "0.1395" ], [ "0.804", "0.1247", "0.8613", "0.1247", "0.8613", "0.1395", "0.804", "0.1395" ], [ "0.8774", "0.1247", "0.922", "0.1247", "0.922", "0.1395", "0.8774", "0.1395" ], [ "0.1293", "0.1372", "0.3525", "0.1372", "0.3525", "0.1509", "0.1293", "0.1509" ], [ "0.7305", "0.1271", "0.7466", "0.1271", "0.7466", "0.1372", "0.7305", "0.1372" ], [ "0.424", "0.1372", "0.7173", "0.1372", "0.7173", "0.1509", "0.424", "0.1509" ], [ "0.7305", "0.1383", "0.7515", "0.1383", "0.7515", "0.1509", "0.7305", "0.1509" ], [ "0.804", "0.1383", "0.86", "0.1383", "0.86", "0.1497", "0.804", "0.1497" ], [ "0.879", "0.1383", "0.9253", "0.1383", "0.9253", "0.1497", "0.879", "0.1497" ], [ "0.1276", "0.1497", "0.3494", "0.1497", "0.3494", "0.1633", "0.1276", "0.1633" ], [ "0.3704", "0.1443", "0.3972", "0.1443", "0.3972", "0.1561", "0.3704", "0.1561" ], [ "0.4224", "0.1497", "0.665", "0.1497", "0.665", "0.1633", "0.4224", "0.1633" ], [ "0.7305", "0.1509", "0.7446", "0.1509", "0.7446", "0.1621", "0.7305", "0.1621" ], [ "0.806", "0.1509", "0.86", "0.1509", "0.86", "0.1621", "0.806", "0.1621" ], [ "0.879", "0.1509", "0.927", "0.1509", "0.927", "0.1621", "0.879", "0.1621" ], [ "0.729", "0.1609", "0.7466", "0.1609", "0.7466", "0.1758", "0.729", "0.1758" ], [ "0.8037", "0.1603", "0.86", "0.1621", "0.858", "0.177", "0.802", "0.1746" ], [ "0.8774", "0.1609", "0.9277", "0.1609", "0.9277", "0.1758", "0.8774", "0.1758" ], [ "0.424", "0.1621", "0.6533", "0.1621", "0.6533", "0.1758", "0.424", "0.1758" ], [ "0.1301", "0.1633", "0.3477", "0.1633", "0.3477", "0.1733", "0.1301", "0.1733" ], [ "0.126", "0.1776", "0.3208", "0.1776", "0.3208", "0.1912", "0.126", "0.1912" ], [ "0.804", "0.1799", "0.86", "0.1799", "0.86", "0.1912", "0.804", "0.1912" ], [ "0.881", "0.1799", "0.9277", "0.1799", "0.9277", "0.1912", "0.881", "0.1912" ], [ "0.7305", "0.1812", "0.7446", "0.1812", "0.7446", "0.1924", "0.7305", "0.1924" ], [ "0.3687", "0.1859", "0.4014", "0.1859", "0.4014", "0.1978", "0.3687", "0.1978" ], [ "0.4248", "0.1812", "0.69", "0.1812", "0.69", "0.1912", "0.4248", "0.1912" ], [ "0.1276", "0.1912", "0.3284", "0.1912", "0.3284", "0.2048", "0.1276", "0.2048" ], [ "0.424", "0.1912", "0.588", "0.1924", "0.588", "0.2037", "0.424", "0.2031" ], [ "0.806", "0.1924", "0.858", "0.1924", "0.858", "0.2037", "0.806", "0.2037" ], [ "0.881", "0.1924", "0.927", "0.1924", "0.927", "0.2037", "0.881", "0.2037" ], [ "0.1184", "0.2173", "0.8833", "0.2185", "0.8833", "0.2328", "0.1184", "0.2322" ], [ "0.7305", "0.1936", "0.7446", "0.1936", "0.7446", "0.2031", "0.7305", "0.2031" ], [ "0.1184", "0.2328", "0.157", "0.2351", "0.1561", "0.2477", "0.11755", "0.2452" ], [ "0.1855", "0.2642", "0.367", "0.2642", "0.367", "0.2778", "0.1855", "0.2778" ], [ "0.4282", "0.2642", "0.561", "0.2642", "0.561", "0.275", "0.4282", "0.275" ], [ "0.644", "0.2642", "0.797", "0.2642", "0.797", "0.2778", "0.644", "0.2778" ], [ "0.3494", "0.279", "0.3938", "0.279", "0.3938", "0.2893", "0.3494", "0.2893" ], [ "0.4592", "0.2766", "0.525", "0.279", "0.525", "0.2903", "0.4592", "0.288" ], [ "0.644", "0.279", "0.8423", "0.2798", "0.8423", "0.2944", "0.644", "0.2935" ], [ "0.2141", "0.2986", "0.293", "0.2998", "0.293", "0.3113", "0.2141", "0.31" ], [ "0.451", "0.2976", "0.531", "0.2976", "0.531", "0.3088", "0.451", "0.3088" ], [ "0.1814", "0.3247", "0.1897", "0.3247", "0.1897", "0.33", "0.1814", "0.33" ], [ "0.4358", "0.3623", "0.5425", "0.3623", "0.5425", "0.3735", "0.4358", "0.3735" ], [ "0.756", "0.3599", "0.7866", "0.3599", "0.7866", "0.3677", "0.756", "0.3677" ], [ "0.742", "0.3694", "0.75", "0.3694", "0.75", "0.3723", "0.742", "0.3723" ], [ "0.5054", "0.376", "0.5146", "0.376", "0.5146", "0.3801", "0.5054", "0.3801" ], [ "0.727", "0.3813", "0.7515", "0.3813", "0.7515", "0.3896", "0.727", "0.3896" ], [ "0.748", "0.3801", "0.812", "0.3801", "0.812", "0.3901", "0.748", "0.3901" ], [ "0.8096", "0.3813", "0.8237", "0.3813", "0.8237", "0.3896", "0.8096", "0.3896" ], [ "0.69", "0.3818", "0.6987", "0.3818", "0.6987", "0.3872", "0.69", "0.3872" ], [ "0.2301", "0.3818", "0.2377", "0.3855", "0.2351", "0.3884", "0.2275", "0.386" ], [ "0.635", "0.383", "0.642", "0.383", "0.642", "0.386", "0.635", "0.386" ], [ "0.66", "0.383", "0.6724", "0.383", "0.6724", "0.3872", "0.66", "0.3872" ], [ "0.5474", "0.3843", "0.5557", "0.3843", "0.5557", "0.3884", "0.5474", "0.3884" ], [ "0.272", "0.3843", "0.2795", "0.3843", "0.2795", "0.3872", "0.272", "0.3872" ], [ "0.4626", "0.3855", "0.4744", "0.3855", "0.4744", "0.3896", "0.4626", "0.3896" ], [ "0.2258", "0.4067", "0.777", "0.4067", "0.777", "0.4204", "0.2258", "0.4204" ], [ "0.12006", "0.4478", "0.4895", "0.4478", "0.4895", "0.4614", "0.12006", "0.4614" ], [ "0.5225", "0.4478", "0.882", "0.4478", "0.882", "0.4614", "0.5225", "0.4614" ], [ "0.12006", "0.4644", "0.4878", "0.4644", "0.4878", "0.478", "0.12006", "0.478" ], [ "0.5146", "0.4631", "0.882", "0.4631", "0.882", "0.478", "0.5146", "0.478" ], [ "0.12006", "0.481", "0.4878", "0.481", "0.4878", "0.4946", "0.12006", "0.4946" ], [ "0.5156", "0.4797", "0.882", "0.4797", "0.882", "0.4946", "0.5156", "0.4946" ], [ "0.1184", "0.4954", "0.4895", "0.4963", "0.4895", "0.511", "0.1184", "0.5103" ], [ "0.5146", "0.4963", "0.882", "0.4963", "0.882", "0.5103", "0.5146", "0.5103" ], [ "0.12177", "0.5127", "0.486", "0.5127", "0.486", "0.5254", "0.12177", "0.5254" ], [ "0.513", "0.511", "0.8486", "0.5127", "0.8486", "0.527", "0.513", "0.5254" ], [ "0.12177", "0.5293", "0.4895", "0.5293", "0.4895", "0.542", "0.12177", "0.542" ], [ "0.5728", "0.5337", "0.797", "0.542", "0.796", "0.5684", "0.571", "0.56" ], [ "0.771", "0.535", "0.8237", "0.5415", "0.8193", "0.557", "0.7656", "0.5503" ], [ "0.12177", "0.5444", "0.486", "0.5444", "0.486", "0.558", "0.12177", "0.558" ], [ "0.8564", "0.546", "0.882", "0.546", "0.882", "0.558", "0.8564", "0.558" ], [ "0.12006", "0.56", "0.4878", "0.56", "0.4878", "0.5737", "0.12006", "0.5737" ], [ "0.12177", "0.5767", "0.4878", "0.5767", "0.4878", "0.5903", "0.12177", "0.5903" ], [ "0.5146", "0.576", "0.881", "0.576", "0.881", "0.589", "0.5146", "0.589" ], [ "0.12177", "0.5933", "0.4878", "0.5933", "0.4878", "0.607", "0.12177", "0.607" ], [ "0.5156", "0.5933", "0.882", "0.5933", "0.882", "0.607", "0.5156", "0.607" ], [ "0.12006", "0.608", "0.486", "0.608", "0.486", "0.6216", "0.12006", "0.6216" ], [ "0.5146", "0.608", "0.882", "0.6094", "0.882", "0.6235", "0.5146", "0.6226" ], [ "0.12006", "0.6235", "0.486", "0.6235", "0.486", "0.637", "0.12006", "0.637" ], [ "0.5146", "0.6235", "0.617", "0.626", "0.617", "0.64", "0.5146", "0.638" ], [ "0.1184", "0.64", "0.486", "0.641", "0.486", "0.6562", "0.1184", "0.655" ], [ "0.5205", "0.6494", "0.881", "0.6494", "0.881", "0.663", "0.5205", "0.663" ], [ "0.12006", "0.6567", "0.4878", "0.6567", "0.4878", "0.6714", "0.12006", "0.6714" ], [ "0.5156", "0.6665", "0.882", "0.6665", "0.882", "0.6797", "0.5156", "0.6797" ], [ "0.1184", "0.673", "0.4878", "0.673", "0.4878", "0.686", "0.1184", "0.686" ], [ "0.513", "0.681", "0.882", "0.681", "0.882", "0.695", "0.513", "0.695" ], [ "0.1184", "0.6885", "0.486", "0.6885", "0.486", "0.7017", "0.1184", "0.7017" ], [ "0.513", "0.6987", "0.882", "0.6987", "0.882", "0.712", "0.513", "0.712" ], [ "0.11755", "0.705", "0.486", "0.705", "0.486", "0.7188", "0.11755", "0.7188" ], [ "0.5146", "0.7134", "0.6514", "0.7144", "0.6514", "0.7285", "0.5146", "0.728" ], [ "0.11755", "0.72", "0.287", "0.7217", "0.287", "0.7363", "0.11755", "0.7354" ], [ "0.623", "0.7437", "0.771", "0.7407", "0.771", "0.7583", "0.623", "0.761" ], [ "0.858", "0.749", "0.8833", "0.749", "0.8833", "0.763", "0.858", "0.763" ], [ "0.12006", "0.7573", "0.1486", "0.7573", "0.1486", "0.7686", "0.12006", "0.7686" ], [ "0.1663", "0.7573", "0.2418", "0.7573", "0.2418", "0.7686", "0.1663", "0.7686" ], [ "0.12006", "0.782", "0.3652", "0.782", "0.3652", "0.796", "0.12006", "0.796" ], [ "0.5146", "0.782", "0.882", "0.783", "0.882", "0.7974", "0.5146", "0.7964" ], [ "0.5156", "0.8", "0.882", "0.8", "0.882", "0.8135", "0.5156", "0.8135" ], [ "0.126", "0.814", "0.486", "0.814", "0.486", "0.8276", "0.126", "0.8276" ], [ "0.5146", "0.8154", "0.8833", "0.8154", "0.8833", "0.83", "0.5146", "0.83" ], [ "0.1184", "0.829", "0.4878", "0.83", "0.4878", "0.844", "0.1184", "0.8433" ], [ "0.5146", "0.8306", "0.5635", "0.8306", "0.5635", "0.8457", "0.5146", "0.8457" ], [ "0.12006", "0.8467", "0.4878", "0.8467", "0.4878", "0.86", "0.12006", "0.86" ], [ "0.5156", "0.8594", "0.5615", "0.8594", "0.5615", "0.8706", "0.5156", "0.8706" ], [ "0.5576", "0.858", "0.739", "0.858", "0.739", "0.8716", "0.5576", "0.8716" ], [ "0.5156", "0.879", "0.8833", "0.879", "0.8833", "0.8926", "0.5156", "0.8926" ], [ "0.12006", "0.8633", "0.2568", "0.8633", "0.2568", "0.8774", "0.12006", "0.8774" ], [ "0.1351", "0.8926", "0.3938", "0.8936", "0.3938", "0.911", "0.1351", "0.91" ], [ "0.517", "0.8945", "0.882", "0.8945", "0.882", "0.909", "0.517", "0.909" ] ]

[ "Datasets", "Description", "# Class", "#Train", "# Test", "ATIS-fight (Tur et al.. 2010)", "Task", "Classify Airline Travel dialog", "21", "4,478", "893", "NLU", "Classify inputs to personal voice assistant", "13,084", "700", "7", "Snips (Coucke et al., 2018)", "AG-news (Zhang et al., 2015)", "Categories: World, Sports, etc", "120,000", "7,600", "DBPedia (Lehmann et al., 2015)", "4", "Categories: Company, Athlete, Album, etc", "14", "560,000", "70,000", "Sogou-news (Zhang et al., 2015)", "DC", "Categories: Sports, Technology, etc", "5", "450,000", "60,000", "5", "650,000", "50,000", "Categories: Review Ratings (1-5)", "Yelp-review (Zhang et al., 2015)", "SNLI (Bowman et al., 2015)", "550,152", "10.000", "3", "NLI", "Entailment: Contradict,Neutral,Entail", "MNLI (Williams et al., 2018)", "Multi-Genre Entailment", "392,702", "10,000", "Table 2: An overview of different datasets under different classification tasks including description and", "3", "sizes", "Natural Language Understanding", "Document Classification", "Natural Language Inference", "Intenty", "Categoryy", "Inference CompositionEntailmenty", "Intent Attention", "Pooling/Flatten", "11", "Convolution Network", "1", "", "+", "pi", "P2P3", "pa", "h", "X", "", "ho", "Y", ".", "X", "Figure 3: The neural network architecture overview of different NLP tasks.", "左到右和从右到左的上下文信息来形成隐藏", "TF-IDF（task-agnostic）该算法将词汇选择", "表示，然后计算自我注意权重以聚合隐藏表示", "视为检索问题E(Ramos etal.,2003），其中", "并预测用户意图。在文档分类中，我们主要遵", "术语频率被视为单词频率，文档频率被视为出", "循CNN架构E(Kim,2014)来提取n-gram特", "现该单词的句子数。在这里，我们遵循规范的", "征，然后聚合这些特征以预测文档类别。在自", "TF-IDF方法来计算检索分数，如下所示：", "然语言推理中，我们遵循流行的ESIM架构E", "tfid(u, D) =t(u)**(log )", "N1-x", "(Williams et al., 2018; Chen et al.,2017)使用", "(8)", "Github实现3。在此结构中，使用三个主要组", "件输入编码，局部推理建模和推理组合来执行", "其中tf(w）表示单词频率，入是平衡因子，N", "顺序推理和组合，以模拟前提与假设之间的交", "表示句子数，nw表示出现w的句子数。我们", "互。请注意，我们不应用（Chenetal.，2017）", "根据tfidf对整个词汇进行排名，并在给定的", "中提出的基于语法树的LSTM，因为我们在词", "值处截止。", "汇压缩后丢失了解析树（Klein and Manning", "GroupLasso（task-aware）此基线旨在通过", "2003），相反，我们遵循更简单的顺序LSTM", "对每行单词嵌入进行分组来找到内在的稀疏结", "框架，没有任何语法解析作为输入。此外，准", "构E(Liu et al., 2015: Park et al., 2016: Wen", "确性曲线是使用公开可用的测试拆分而不是官", "etal.，2016）。正则化目标描述如下，旨在找", "方在线评估获得的，因为我们需要在不同的词", "到行式稀疏结构：", "汇量下进行多次评估", "Creg =E(we)2", "(9)", "4.2", "Baselines", "在这里我们主要考虑以下基线：", "经过上述正则化优化后，我们在嵌人矩阵的行", "范数上使用基于阈值的选择策略，所选词汇描", "Frequency-based（task-agnostic）这种方法", "述为V=[wiEVillWil2>βr}，其中βT是", "已经在Esection1中广泛讨论过，其基本思想", "阈值。", "是根据其频率对单词进行排序，然后设置阈值", "4.3", "Experimental Results", "在这里，我们分别在XX5中展示了我们在", "以切断长尾分布。", "3https://github.com/coetaur0/ESIM", "自然语言理解，文档分类和从这些表中，首" ]

{ "arxiv_id": "1902.10339", "language": "zh", "page_idx": 4 }

[ [ "0.2834", "0.1338", "0.6895", "0.1338", "0.6895", "0.1483", "0.2834", "0.1483" ], [ "0.0948", "0.2615", "0.1152", "0.262", "0.11273", "0.2986", "0.09235", "0.2974" ], [ "0.10864", "0.262", "0.1283", "0.262", "0.1283", "0.2966", "0.10864", "0.2966" ], [ "0.09314", "0.3018", "0.11273", "0.3018", "0.11273", "0.3806", "0.09314", "0.3806" ], [ "0.11273", "0.3113", "0.1356", "0.3113", "0.1356", "0.375", "0.11273", "0.375" ], [ "0.0801", "0.385", "0.1405", "0.3838", "0.1438", "0.4734", "0.0825", "0.4749" ], [ "0.859", "0.464", "0.8735", "0.464", "0.8735", "0.4749", "0.859", "0.4749" ], [ "0.1111", "0.5024", "0.1299", "0.5024", "0.1323", "0.5747", "0.1136", "0.5747" ], [ "0.09314", "0.512", "0.11273", "0.512", "0.11273", "0.5645", "0.09314", "0.5645" ], [ "0.1854", "0.589", "0.2115", "0.589", "0.2115", "0.601", "0.1854", "0.601" ], [ "0.7656", "0.5903", "0.8105", "0.5903", "0.8105", "0.6016", "0.7656", "0.6016" ], [ "0.8545", "0.589", "0.876", "0.589", "0.876", "0.6016", "0.8545", "0.6016" ], [ "0.0809", "0.616", "0.4722", "0.616", "0.4722", "0.627", "0.0809", "0.627" ], [ "0.492", "0.6187", "0.6235", "0.6187", "0.6235", "0.6255", "0.492", "0.6255" ], [ "0.626", "0.616", "0.8906", "0.616", "0.8906", "0.627", "0.626", "0.627" ], [ "0.0809", "0.629", "0.8906", "0.629", "0.8906", "0.6426", "0.0809", "0.6426" ], [ "0.0809", "0.6426", "0.892", "0.6426", "0.892", "0.656", "0.0809", "0.656" ], [ "0.0809", "0.657", "0.8555", "0.657", "0.8555", "0.6704", "0.0809", "0.6704" ], [ "0.2361", "0.69", "0.3137", "0.6924", "0.3137", "0.7085", "0.2361", "0.7056" ], [ "0.5024", "0.6934", "0.892", "0.6934", "0.892", "0.707", "0.5024", "0.707" ], [ "0.5024", "0.709", "0.892", "0.707", "0.892", "0.72", "0.5024", "0.7227" ], [ "0.5024", "0.7227", "0.8896", "0.7227", "0.8896", "0.7363", "0.5024", "0.7363" ], [ "0.0989", "0.7373", "0.468", "0.7373", "0.468", "0.7505", "0.0989", "0.7505" ], [ "0.5024", "0.7373", "0.892", "0.738", "0.892", "0.7524", "0.5024", "0.752" ], [ "0.0809", "0.752", "0.468", "0.752", "0.468", "0.765", "0.0809", "0.765" ], [ "0.505", "0.755", "0.8906", "0.755", "0.8906", "0.765", "0.505", "0.765" ], [ "0.0809", "0.7676", "0.468", "0.7676", "0.468", "0.781", "0.0809", "0.781" ], [ "0.5034", "0.768", "0.8906", "0.768", "0.8906", "0.782", "0.5034", "0.782" ], [ "0.0825", "0.784", "0.4656", "0.784", "0.4656", "0.7944", "0.0825", "0.7944" ], [ "0.5024", "0.784", "0.892", "0.782", "0.892", "0.797", "0.5024", "0.799" ], [ "0.0801", "0.7974", "0.4673", "0.7974", "0.4673", "0.811", "0.0801", "0.811" ], [ "0.5034", "0.8003", "0.8906", "0.8003", "0.8906", "0.811", "0.5034", "0.811" ], [ "0.0809", "0.8125", "0.4697", "0.8125", "0.4697", "0.8257", "0.0809", "0.8257" ], [ "0.501", "0.8125", "0.8906", "0.813", "0.8906", "0.8276", "0.501", "0.827" ], [ "0.0809", "0.827", "0.468", "0.827", "0.468", "0.8403", "0.0809", "0.8403" ], [ "0.501", "0.829", "0.71", "0.8276", "0.71", "0.8423", "0.501", "0.8433" ], [ "0.0809", "0.8423", "0.4673", "0.8423", "0.4673", "0.857", "0.0809", "0.857" ], [ "0.0809", "0.858", "0.4697", "0.858", "0.4697", "0.872", "0.0809", "0.872" ], [ "0.0809", "0.8726", "0.468", "0.8726", "0.468", "0.886", "0.0809", "0.886" ], [ "0.5195", "0.875", "0.892", "0.8755", "0.892", "0.89", "0.5195", "0.8896" ], [ "0.0784", "0.888", "0.4697", "0.886", "0.4697", "0.901", "0.0784", "0.903" ], [ "0.5024", "0.8916", "0.584", "0.8916", "0.584", "0.9014", "0.5024", "0.9014" ] ]

[ "Audio-Driven Emotional Video Portraits", "du", "pn", "e", "dar", "Iro l pal", "[a]", "", "n", "Sad", "Happy", "(b)", "Figurel:Audio-DrivenEmotionalVideoPortraits.Given", "audiocip anda:tar", "getvideo,ourEmotionalVideoPortraits", "(EVP) approach is capable of generating emotion-controllable talking portraits and change the emotion of them smoothly", "by interpolating at the latent space. (a) Generated video portraits with the same speech content but different emotions", "(i.e., contempt and sad). (b) Linear interpolation of the learned latent representation of emotions from sad to happy.", "Abstract", "coupled spaces, i.e., a duration-independent emotion", "space and a duration-dependent content space.With", "the disentangled features, dynamic 2D emotional fa-", "Despite previous success in generating audio-driven", "cial landmarks can be deduced. Then we propose the", "talking heads, most of the previous studies focus on", "Target-AdaptiveFaceSynthesistechniguetogenerate", "the correlation between speech content and the mouth", "the final high-quality video portraits, by bridging the", "shape.Facial emotion,whichis oneofthemostim-", "gap between the deduced landmarks and the natural", "portant features on natural human faces, is always ne-", "head poses of target videos.Extensive experiments", "glected in their methods.In this work, we present", "demonstrate the effectiveness of our method both qual-", "Emotional Video Portraits (EVP), a system for syn-", "itativelyandquantitatively.", "thesizing high-quality video portraits with vivid emo-", "tional dynamics driven by audios. Specifically, we", "propose the Cross-Reconstructed Emotion Disentan-", "1All materials are available at https://jixinya.github.io/", "glement technique to decompose speech into two de-", "projects/evp/" ]

{ "arxiv_id": "2104.07452", "language": "zh", "page_idx": 1 }

[ [ "0.0809", "0.09216", "0.2157", "0.09216", "0.2157", "0.1061", "0.0809", "0.1061" ], [ "0.505", "0.09344", "0.8906", "0.09344", "0.8906", "0.1067", "0.505", "0.1067" ], [ "0.5063", "0.10986", "0.888", "0.10986", "0.888", "0.1206", "0.5063", "0.1206" ], [ "0.1013", "0.11615", "0.4673", "0.11615", "0.4673", "0.13", "0.1013", "0.13" ], [ "0.5034", "0.1225", "0.888", "0.1237", "0.888", "0.137", "0.5034", "0.1364" ], [ "0.0825", "0.132", "0.4697", "0.132", "0.4697", "0.1459", "0.0825", "0.1459" ], [ "0.5034", "0.137", "0.8906", "0.1383", "0.8906", "0.1528", "0.5034", "0.1515" ], [ "0.0809", "0.1477", "0.468", "0.1477", "0.468", "0.161", "0.0809", "0.161" ], [ "0.505", "0.154", "0.8906", "0.154", "0.8906", "0.1674", "0.505", "0.1674" ], [ "0.0809", "0.1622", "0.468", "0.1622", "0.468", "0.1755", "0.0809", "0.1755" ], [ "0.505", "0.1686", "0.8906", "0.1686", "0.8906", "0.1818", "0.505", "0.1818" ], [ "0.08417", "0.1787", "0.4656", "0.1787", "0.4656", "0.1895", "0.08417", "0.1895" ], [ "0.5063", "0.1849", "0.8896", "0.1849", "0.8896", "0.1957", "0.5063", "0.1957" ], [ "0.08417", "0.1925", "0.468", "0.1925", "0.468", "0.2058", "0.08417", "0.2058" ], [ "0.5024", "0.1989", "0.732", "0.1989", "0.732", "0.2122", "0.5024", "0.2122" ], [ "0.08417", "0.209", "0.4673", "0.209", "0.4673", "0.2197", "0.08417", "0.2197" ], [ "0.523", "0.214", "0.6787", "0.214", "0.6787", "0.2279", "0.523", "0.2279" ], [ "0.0809", "0.2229", "0.4673", "0.2229", "0.4673", "0.2361", "0.0809", "0.2361" ], [ "0.09644", "0.2393", "0.4673", "0.2393", "0.4673", "0.25", "0.09644", "0.25" ], [ "0.5186", "0.2412", "0.888", "0.2412", "0.888", "0.2551", "0.5186", "0.2551" ], [ "0.0809", "0.2524", "0.3472", "0.2524", "0.3472", "0.2664", "0.0809", "0.2664" ], [ "0.08905", "0.2437", "0.0972", "0.2437", "0.0972", "0.2468", "0.08905", "0.2468" ], [ "0.5337", "0.2556", "0.8857", "0.2556", "0.8857", "0.2695", "0.5337", "0.2695" ], [ "0.1013", "0.2683", "0.468", "0.2683", "0.468", "0.2822", "0.1013", "0.2822" ], [ "0.536", "0.2695", "0.8896", "0.2695", "0.8896", "0.2827", "0.536", "0.2827" ], [ "0.08417", "0.2827", "0.4673", "0.2827", "0.4673", "0.2966", "0.08417", "0.2966" ], [ "0.535", "0.2874", "0.681", "0.2842", "0.681", "0.2974", "0.535", "0.3005" ], [ "0.0825", "0.2998", "0.4656", "0.2998", "0.4656", "0.31", "0.0825", "0.31" ], [ "0.0825", "0.3132", "0.4673", "0.3132", "0.4673", "0.327", "0.0825", "0.327" ], [ "0.517", "0.3125", "0.8896", "0.3125", "0.8896", "0.3257", "0.517", "0.3257" ], [ "0.0825", "0.3289", "0.4673", "0.3289", "0.4673", "0.3423", "0.0825", "0.3423" ], [ "0.536", "0.327", "0.8906", "0.327", "0.8906", "0.3403", "0.536", "0.3403" ], [ "0.0809", "0.3423", "0.4656", "0.3423", "0.4656", "0.356", "0.0809", "0.356" ], [ "0.536", "0.3435", "0.7715", "0.3435", "0.7715", "0.3567", "0.536", "0.3567" ], [ "0.0809", "0.3591", "0.4673", "0.3591", "0.4673", "0.3726", "0.0809", "0.3726" ], [ "0.5186", "0.3694", "0.8896", "0.3694", "0.8896", "0.3833", "0.5186", "0.3833" ], [ "0.09644", "0.3757", "0.4656", "0.3757", "0.4656", "0.3865", "0.09644", "0.3865" ], [ "0.08417", "0.3901", "0.4656", "0.3901", "0.4656", "0.401", "0.08417", "0.401" ], [ "0.536", "0.385", "0.8896", "0.3838", "0.8896", "0.3977", "0.536", "0.3984" ], [ "0.0825", "0.4048", "0.4207", "0.4048", "0.4207", "0.4153", "0.0825", "0.4153" ], [ "0.536", "0.401", "0.8896", "0.3997", "0.8896", "0.4136", "0.536", "0.414" ], [ "0.1013", "0.4185", "0.4673", "0.4185", "0.4673", "0.432", "0.1013", "0.432" ], [ "0.536", "0.4153", "0.7656", "0.4153", "0.7656", "0.4287", "0.536", "0.4287" ], [ "0.0825", "0.4343", "0.4673", "0.4343", "0.4673", "0.4475", "0.0825", "0.4475" ], [ "0.5024", "0.4443", "0.6504", "0.4443", "0.6504", "0.4583", "0.5024", "0.4583" ], [ "0.08496", "0.4507", "0.4656", "0.4507", "0.4656", "0.461", "0.08496", "0.461" ], [ "0.0809", "0.464", "0.2607", "0.464", "0.2607", "0.4778", "0.0809", "0.4778" ], [ "0.5024", "0.4702", "0.8384", "0.4702", "0.8384", "0.4841", "0.5024", "0.4841" ], [ "0.10046", "0.4792", "0.468", "0.4792", "0.468", "0.4924", "0.10046", "0.4924" ], [ "0.0809", "0.4944", "0.4673", "0.4944", "0.4673", "0.5083", "0.0809", "0.5083" ], [ "0.524", "0.4937", "0.8896", "0.4937", "0.8896", "0.507", "0.524", "0.507" ], [ "0.0809", "0.5103", "0.4673", "0.5103", "0.4673", "0.5234", "0.0809", "0.5234" ], [ "0.5034", "0.5083", "0.8896", "0.5083", "0.8896", "0.5215", "0.5034", "0.5215" ], [ "0.08496", "0.5264", "0.4673", "0.5264", "0.4673", "0.537", "0.08496", "0.537" ], [ "0.5024", "0.523", "0.8896", "0.5234", "0.8896", "0.537", "0.5024", "0.536" ], [ "0.0825", "0.5415", "0.4673", "0.5415", "0.4673", "0.552", "0.0825", "0.552" ], [ "0.5034", "0.539", "0.888", "0.539", "0.888", "0.553", "0.5034", "0.553" ], [ "0.0801", "0.5547", "0.468", "0.5537", "0.468", "0.5684", "0.0801", "0.5693" ], [ "0.5034", "0.5537", "0.8906", "0.5537", "0.8906", "0.5674", "0.5034", "0.5674" ], [ "0.08417", "0.5713", "0.452", "0.5713", "0.452", "0.582", "0.08417", "0.582" ], [ "0.4495", "0.5747", "0.4631", "0.5747", "0.4631", "0.579", "0.4495", "0.579" ], [ "0.5063", "0.571", "0.732", "0.571", "0.732", "0.5806", "0.5063", "0.5806" ], [ "0.0825", "0.587", "0.4656", "0.587", "0.4656", "0.5977", "0.0825", "0.5977" ], [ "0.5034", "0.5854", "0.8906", "0.5854", "0.8906", "0.5986", "0.5034", "0.5986" ], [ "0.0825", "0.5996", "0.468", "0.601", "0.468", "0.6157", "0.0825", "0.6143" ], [ "0.505", "0.5996", "0.8906", "0.5996", "0.8906", "0.613", "0.505", "0.613" ], [ "0.0825", "0.6143", "0.4656", "0.6143", "0.4656", "0.6274", "0.0825", "0.6274" ], [ "0.505", "0.6143", "0.8896", "0.613", "0.8896", "0.627", "0.505", "0.6274" ], [ "0.0809", "0.631", "0.406", "0.631", "0.406", "0.6445", "0.0809", "0.6445" ], [ "0.5034", "0.6274", "0.8906", "0.629", "0.8906", "0.643", "0.5034", "0.6426" ], [ "0.10297", "0.6445", "0.4656", "0.6445", "0.4656", "0.6577", "0.10297", "0.6577" ], [ "0.5063", "0.6445", "0.8906", "0.6445", "0.8906", "0.6577", "0.5063", "0.6577" ], [ "0.0825", "0.662", "0.4673", "0.662", "0.4673", "0.6724", "0.0825", "0.6724" ], [ "0.5034", "0.66", "0.8906", "0.66", "0.8906", "0.674", "0.5034", "0.674" ], [ "0.08417", "0.678", "0.4673", "0.678", "0.4673", "0.688", "0.08417", "0.688" ], [ "0.505", "0.675", "0.8906", "0.675", "0.8906", "0.6895", "0.505", "0.6895" ], [ "0.0809", "0.6914", "0.4673", "0.6914", "0.4673", "0.705", "0.0809", "0.705" ], [ "0.5024", "0.6895", "0.8896", "0.6895", "0.8896", "0.7026", "0.5024", "0.7026" ], [ "0.0825", "0.7085", "0.4673", "0.7085", "0.4673", "0.7183", "0.0825", "0.7183" ], [ "0.5034", "0.705", "0.8906", "0.705", "0.8906", "0.7197", "0.5034", "0.7197" ], [ "0.0825", "0.7217", "0.468", "0.7217", "0.468", "0.735", "0.0825", "0.735" ], [ "0.5034", "0.72", "0.846", "0.72", "0.846", "0.7344", "0.5034", "0.7344" ], [ "0.0825", "0.7363", "0.4673", "0.7363", "0.4673", "0.75", "0.0825", "0.75" ], [ "0.527", "0.7373", "0.888", "0.7373", "0.888", "0.7476", "0.527", "0.7476" ], [ "0.0825", "0.752", "0.468", "0.752", "0.468", "0.765", "0.0825", "0.765" ], [ "0.5024", "0.75", "0.8906", "0.752", "0.8906", "0.765", "0.5024", "0.763" ], [ "0.0809", "0.7676", "0.4673", "0.7676", "0.4673", "0.781", "0.0809", "0.781" ], [ "0.505", "0.7666", "0.8906", "0.7666", "0.8906", "0.78", "0.505", "0.78" ], [ "0.0809", "0.782", "0.4673", "0.782", "0.4673", "0.7954", "0.0809", "0.7954" ], [ "0.5034", "0.781", "0.8906", "0.781", "0.8906", "0.7944", "0.5034", "0.7944" ], [ "0.08417", "0.799", "0.4673", "0.799", "0.4673", "0.8096", "0.08417", "0.8096" ], [ "0.505", "0.797", "0.8906", "0.797", "0.8906", "0.81", "0.505", "0.81" ], [ "0.08417", "0.813", "0.4673", "0.813", "0.4673", "0.8237", "0.08417", "0.8237" ], [ "0.505", "0.813", "0.888", "0.813", "0.888", "0.8237", "0.505", "0.8237" ], [ "0.0809", "0.827", "0.468", "0.827", "0.468", "0.8403", "0.0809", "0.8403" ], [ "0.505", "0.8276", "0.5522", "0.8276", "0.5522", "0.839", "0.505", "0.839" ], [ "0.0809", "0.8423", "0.468", "0.8423", "0.468", "0.856", "0.0809", "0.856" ], [ "0.5063", "0.8433", "0.886", "0.8433", "0.886", "0.854", "0.5063", "0.854" ], [ "0.0825", "0.857", "0.468", "0.857", "0.468", "0.8706", "0.0825", "0.8706" ], [ "0.505", "0.8594", "0.886", "0.8594", "0.886", "0.869", "0.505", "0.869" ], [ "0.0809", "0.872", "0.468", "0.8706", "0.468", "0.8853", "0.0809", "0.886" ], [ "0.505", "0.8726", "0.8906", "0.8726", "0.8906", "0.886", "0.505", "0.886" ], [ "0.0809", "0.886", "0.468", "0.887", "0.468", "0.9014", "0.0809", "0.901" ], [ "0.505", "0.887", "0.8896", "0.887", "0.8896", "0.901", "0.505", "0.901" ] ]

[ "1.Introduction", "因此在生成的地标与目标视频中头部姿势和运动的大", "差异之间存在差距。为此，我们提出了自标自适应人脸", "生成音频驱动的照片般逼真的肖像视频对于多媒体", "合成技术来弥合3D空间中推断的地标和目标视频肖", "应用程序非常需要，例如电影制作E[22]，远程呈现E", "像之间的姿态差距。通过精心设计的3D感知关键点", "[2]和数字人类动画E[25，14，48]。以前的作品已经探", "对齐算法，我们能够将2D地标投影到目标视频中。最", "索了生成会说话的头部或肖像，其嘴唇运动与输入的", "后，我们训练一个边缘到视频翻译网络，以生成最终的", "语音内容同步。通常，这些技术可以分为两类：1）基于", "高质量情感视频肖像。大量实验证明了我们方法的优", "图像的方法，可以对一顿或几顿裁剪的人脸进行动画", "越性能和几个关键部件的有效性", "处理E11，45，33，9，29，47，71，以及2基于视频的编", "我们的贡献总结如下：", "辑方法，可以直接编辑目标视频剪辑E[34，32，35，42]", "尽管如此，以前的大多数研究都没有对情感进行建", ": We propose the Emotional Video Portraits (EVP)", "模，情感是肖像自然主义的关键因素。", "", "system, which is the first attempt to achieve emo", "只有少数基于图像的作品讨论了会说话的头部生成", "tional control in video-based editing talking face", "中的情感信息。由于缺乏带有情感注释的适当视听数", "generation methods.", "据集，Vougioukas等人。36不明确建模情绪。只需", "将情感和音频内容信息编码为单个功能，它们就会产", ": We introduce Cross-Reconstructed Emotion Dis-", "生低质量的初步结果。最近，Wang等人。[37]收集", "entanglement technique, to distill content-agnostic", "MEAD数据集，其中包含高质量的会说话的头部视频", "emotion features for free control.", "带有情感类别和强度的注释。然后，他们将情绪设置为", ". We introduce Target-Adaptive Face Synthesis, to", "热条件，以控制生成的面孔。然而，所有这些基于图", "像的方法仅谊染具有固定甚至没有背景的轻微头部运", "synthesize high quality portrait by making the", "动，使得它们在大多数现实世界场景中不切实际", "generated face adapt to the target video with nat-", "然而，基于视频的编辑方法，如[34，15，32,35，42]", "ural head poses and movements.", "中讨论的更适用，没有考虑情绪控制。他们中的大多数", "2. Related Work", "人只编辑嘴色，并保持视频肖像的上半部分不变，从而", "无法获得自由的情绪控制", "2.1. Audio-Driven Talking Face Generation", "在这项研究中，我们提出了一种新颖的算法，称为", "情感视频肖像（EVP），旨在使基于视频的编辑说话人", "音频驱动的说话人生成任务旨在合成由音频驱动的", "脸生成具有从音频中进行情感控制的能力。我们使用", "说话人的口型同步视频。它是娱乐领域要求很高的话", "与语音语调更好匹配的情感动态来制作全像动画，从", "题，因此长期以来一直是计算机视觉和计算机图形学", "而获得更生动的效果。然而，实现这一一自标是不平凡", "领域的研究兴趣[4，21，11，38，34，45，33，49，8，9，35", "的。存在几个复杂的挑战：1）从音频中提取情感相当", "29，47，37,6，46，421。我们可以根据这些方法在结果", "困难，因为情感信息与语音内容等其他因素纠缠在", "", "可视化方面的差异将其分为两类", "起。2）在合成高保真度结果的同时，难以将编辑后的", "基于图像的方法。一种类型的模型侧重于以一个或多", "人脸与目标视频混合。音频不为头部姿势和头部的全", "个顿作为身份参考来驱动裁剪的面部区域。Chung等", "局运动提供任何提示，因此从音频推断的编辑头部可", "人。11]首次提出以图像到图像转换的方式生成口型", "能与目标视频具有较大的头部姿势和运动差异", "同步视频[19]方式。ThenZhouetal。[45]和Song等", "为了应对上述挑战，我们设法在拟议的情感视频肖", "人。E33使用解开的视听表示和递归神经网络来改善", "像系统中实现了基于音频的情感控制，该系统具有两", "其结果。此外，Chen等人。[9]利用地标作为中间表", "个关键组件，即交叉重建的情感消除纠缠和自标自适", "示，并将流程分为两个阶段。但是，这些方法只能保证", "应面部合成。为了对生成的肖像进行情感控制，我们首", "生成的嘴巴和音频之间的同步。结果几乎没有任何表", "先提出了音频上的交叉重建情感去纠缠技术，以提取", "情或头部动作。Zhouetal。47]成功地产生了与身份", "两个单独的潜在空间：i）与持续时间无关的空间，它", "相关的头部运动，但他们的模型也未能控制情绪", "是情感的内容不可知编码；i）与持续时间相关的空间", "至于情感说话的面孔，Vougioukas等人。36采用", "它对音频的语音内容进行编码。一旦提取，来自这些潜", "三个分离的鉴别器，分别增强合成细节、同步和真实", "在空间的特征将被重新组合以产生新的音频表示，从", "表达。然而，他们的实验是在有限的规模上进行的。最", "而允许计算和优化交叉重建损失。但是，为了实现交叉", "近，Wang等人。[37]提出了MEAD数据集，并建议", "重建训练，应提供具有相同内容但具有相同长度的不", "通过分别对面部的上部和下部进行拆分操作来生成情", "同情绪的成对句子。这在现实世界的场景中儿乎是无", "绪说话的面孔。然而，他们的结果不太现实，仅限于面", "法实现的。为此，我们采用时间序列分析中的经典算法", "部区域", "动态时间规整(DTW)E[3]来帮助形成对齐的不均匀", "基于视频的编辑方法。全顿视频肖像不仅包含面部区", "长度语音语料库的伪训练对遵循先前的方法E[34，9]", "域，还包含人的颈部和肩部以及背景。毫无疑问，这利", "然后引人具有分解特征的音频到地标动画模块，以推", "设置更现实，但更难重建。因此，大多数方法仅编辑嘴", "导情感2D地标动态。由于音频中没有提供姿势信息，", "部区域。Suwajanakorn等人。[34]通过将音频训练到具" ]

{ "arxiv_id": "2104.07452", "language": "zh", "page_idx": 2 }

[ [ "0.10297", "0.0947", "0.4656", "0.0947", "0.4656", "0.1048", "0.10297", "0.1048" ], [ "0.505", "0.0947", "0.786", "0.0947", "0.786", "0.1048", "0.505", "0.1048" ], [ "0.0825", "0.108", "0.4673", "0.108", "0.4673", "0.1218", "0.0825", "0.1218" ], [ "0.0825", "0.125", "0.4673", "0.125", "0.4673", "0.1351", "0.0825", "0.1351" ], [ "0.528", "0.1206", "0.8906", "0.1218", "0.8906", "0.1364", "0.528", "0.1351" ], [ "0.0825", "0.1383", "0.468", "0.1383", "0.468", "0.1515", "0.0825", "0.1515" ], [ "0.0825", "0.154", "0.468", "0.154", "0.468", "0.1674", "0.0825", "0.1674" ], [ "0.5034", "0.1509", "0.847", "0.1509", "0.847", "0.161", "0.5034", "0.161" ], [ "0.0825", "0.1686", "0.4673", "0.1686", "0.4673", "0.1818", "0.0825", "0.1818" ], [ "0.5024", "0.1736", "0.7837", "0.1755", "0.7837", "0.1895", "0.5024", "0.1875" ], [ "0.0809", "0.1831", "0.468", "0.1831", "0.468", "0.1963", "0.0809", "0.1963" ], [ "0.0809", "0.1989", "0.25", "0.1989", "0.25", "0.2122", "0.0809", "0.2122" ], [ "0.524", "0.1989", "0.8896", "0.1989", "0.8896", "0.2122", "0.524", "0.2122" ], [ "0.505", "0.2134", "0.8896", "0.2134", "0.8896", "0.2267", "0.505", "0.2267" ], [ "0.183", "0.226", "0.3635", "0.226", "0.3635", "0.2437", "0.183", "0.2437" ], [ "0.4468", "0.2343", "0.468", "0.2343", "0.468", "0.2468", "0.4468", "0.2468" ], [ "0.5034", "0.2291", "0.8906", "0.2291", "0.8906", "0.2424", "0.5034", "0.2424" ], [ "0.2157", "0.2443", "0.2615", "0.2443", "0.2615", "0.2588", "0.2157", "0.2588" ], [ "0.505", "0.2437", "0.8906", "0.2437", "0.8906", "0.2568", "0.505", "0.2568" ], [ "0.505", "0.2615", "0.8896", "0.2615", "0.8896", "0.2715", "0.505", "0.2715" ], [ "0.0809", "0.2683", "0.468", "0.2695", "0.468", "0.2827", "0.0809", "0.2822" ], [ "0.5024", "0.2734", "0.8896", "0.2734", "0.8896", "0.2874", "0.5024", "0.2874" ], [ "0.0825", "0.286", "0.4673", "0.286", "0.4673", "0.2966", "0.0825", "0.2966" ], [ "0.5024", "0.2886", "0.8896", "0.289", "0.8896", "0.303", "0.5024", "0.3018" ], [ "0.0825", "0.2998", "0.4673", "0.2986", "0.4673", "0.3125", "0.0825", "0.3132" ], [ "0.5034", "0.3037", "0.8906", "0.3037", "0.8906", "0.3176", "0.5034", "0.3176" ], [ "0.0809", "0.3145", "0.4697", "0.3145", "0.4697", "0.3276", "0.0809", "0.3276" ], [ "0.505", "0.3193", "0.8906", "0.3193", "0.8906", "0.3333", "0.505", "0.3333" ], [ "0.0809", "0.33", "0.1969", "0.33", "0.1969", "0.3435", "0.0809", "0.3435" ], [ "0.505", "0.3352", "0.8906", "0.3352", "0.8906", "0.3484", "0.505", "0.3484" ], [ "0.08496", "0.3447", "0.468", "0.3447", "0.468", "0.358", "0.08496", "0.358" ], [ "0.5034", "0.3499", "0.8896", "0.3499", "0.8896", "0.363", "0.5034", "0.363" ], [ "0.0809", "0.3599", "0.468", "0.3599", "0.468", "0.3738", "0.0809", "0.3738" ], [ "0.5024", "0.363", "0.8896", "0.3643", "0.8896", "0.3774", "0.5024", "0.377" ], [ "0.0801", "0.3738", "0.468", "0.375", "0.468", "0.3894", "0.0801", "0.3882" ], [ "0.5034", "0.3801", "0.892", "0.3801", "0.892", "0.3933", "0.5034", "0.3933" ], [ "0.0809", "0.3901", "0.468", "0.3901", "0.468", "0.404", "0.0809", "0.404" ], [ "0.505", "0.3945", "0.846", "0.3945", "0.846", "0.4077", "0.505", "0.4077" ], [ "0.0801", "0.404", "0.4673", "0.406", "0.4673", "0.4204", "0.0801", "0.4185" ], [ "0.505", "0.4104", "0.8896", "0.4104", "0.8896", "0.4236", "0.505", "0.4236" ], [ "0.0809", "0.4204", "0.464", "0.4204", "0.464", "0.4343", "0.0809", "0.4343" ], [ "0.5024", "0.4236", "0.8896", "0.4248", "0.8896", "0.4382", "0.5024", "0.4375" ], [ "0.0825", "0.4363", "0.4673", "0.4363", "0.4673", "0.4495", "0.0825", "0.4495" ], [ "0.505", "0.4426", "0.8896", "0.4426", "0.8896", "0.4526", "0.505", "0.4526" ], [ "0.0825", "0.4507", "0.468", "0.4507", "0.468", "0.4653", "0.0825", "0.4653" ], [ "0.5034", "0.455", "0.8906", "0.455", "0.8906", "0.4697", "0.5034", "0.4697" ], [ "0.0825", "0.4673", "0.4673", "0.4673", "0.4673", "0.4778", "0.0825", "0.4778" ], [ "0.5034", "0.4702", "0.8677", "0.4702", "0.8677", "0.4841", "0.5034", "0.4841" ], [ "0.0825", "0.4817", "0.3733", "0.4817", "0.3733", "0.4924", "0.0825", "0.4924" ], [ "0.505", "0.4873", "0.8906", "0.4873", "0.8906", "0.4976", "0.505", "0.4976" ], [ "0.1013", "0.4956", "0.4673", "0.4956", "0.4673", "0.509", "0.1013", "0.509" ], [ "0.5024", "0.5", "0.8906", "0.5005", "0.8906", "0.515", "0.5024", "0.5146" ], [ "0.0809", "0.511", "0.468", "0.511", "0.468", "0.5244", "0.0809", "0.5244" ], [ "0.5034", "0.5166", "0.539", "0.5166", "0.539", "0.528", "0.5034", "0.528" ], [ "0.0809", "0.523", "0.4714", "0.5264", "0.4714", "0.5435", "0.0809", "0.539" ], [ "0.5254", "0.5327", "0.8896", "0.5327", "0.8896", "0.5435", "0.5254", "0.5435" ], [ "0.0809", "0.5415", "0.282", "0.5415", "0.282", "0.5547", "0.0809", "0.5547" ], [ "0.5034", "0.547", "0.722", "0.547", "0.722", "0.56", "0.5034", "0.56" ], [ "0.1274", "0.5674", "0.4167", "0.5674", "0.4167", "0.582", "0.1274", "0.582" ], [ "0.5825", "0.5747", "0.7983", "0.5747", "0.7983", "0.589", "0.5825", "0.589" ], [ "0.8677", "0.5757", "0.8906", "0.5757", "0.8906", "0.589", "0.8677", "0.589" ], [ "0.1724", "0.5854", "0.4265", "0.586", "0.4265", "0.601", "0.1724", "0.5996" ], [ "0.437", "0.5776", "0.4673", "0.5776", "0.4673", "0.589", "0.437", "0.589" ], [ "0.5215", "0.6016", "0.8896", "0.604", "0.8896", "0.617", "0.5215", "0.6157" ], [ "0.1013", "0.613", "0.468", "0.613", "0.468", "0.627", "0.1013", "0.627" ], [ "0.505", "0.6187", "0.8906", "0.6187", "0.8906", "0.632", "0.505", "0.632" ], [ "0.0825", "0.631", "0.3325", "0.631", "0.3325", "0.6416", "0.0825", "0.6416" ], [ "0.5063", "0.6353", "0.8896", "0.6353", "0.8896", "0.646", "0.5063", "0.646" ], [ "0.505", "0.6484", "0.888", "0.648", "0.888", "0.661", "0.505", "0.662" ], [ "0.1299", "0.655", "0.4182", "0.656", "0.4182", "0.6704", "0.1299", "0.6694" ], [ "0.4436", "0.6665", "0.4673", "0.6665", "0.4673", "0.678", "0.4436", "0.678" ], [ "0.505", "0.6636", "0.8896", "0.6636", "0.8896", "0.677", "0.505", "0.677" ], [ "0.1683", "0.674", "0.4126", "0.675", "0.4126", "0.6895", "0.1683", "0.688" ], [ "0.505", "0.6787", "0.8906", "0.6787", "0.8906", "0.6924", "0.505", "0.6924" ], [ "0.505", "0.696", "0.8906", "0.696", "0.8906", "0.7056", "0.505", "0.7056" ], [ "0.1013", "0.702", "0.4673", "0.702", "0.4673", "0.7153", "0.1013", "0.7153" ], [ "0.5034", "0.709", "0.8906", "0.709", "0.8906", "0.7227", "0.5034", "0.7227" ], [ "0.08417", "0.7183", "0.4656", "0.7183", "0.4656", "0.729", "0.08417", "0.729" ], [ "0.5024", "0.7236", "0.892", "0.7246", "0.892", "0.738", "0.5024", "0.7373" ], [ "0.0809", "0.732", "0.4429", "0.732", "0.4429", "0.7456", "0.0809", "0.7456" ], [ "0.5034", "0.738", "0.892", "0.7393", "0.892", "0.754", "0.5034", "0.7524" ], [ "0.2566", "0.7583", "0.2712", "0.7583", "0.2712", "0.7676", "0.2566", "0.7676" ], [ "0.505", "0.756", "0.8364", "0.756", "0.8364", "0.7666", "0.505", "0.7666" ], [ "0.183", "0.7676", "0.3652", "0.7676", "0.3652", "0.7886", "0.183", "0.7886" ], [ "0.4468", "0.7715", "0.468", "0.7715", "0.468", "0.785", "0.4468", "0.785" ], [ "0.5933", "0.781", "0.8", "0.781", "0.8", "0.7954", "0.5933", "0.7954" ], [ "0.8677", "0.7827", "0.8906", "0.7827", "0.8906", "0.797", "0.8677", "0.797" ], [ "0.0809", "0.81", "0.468", "0.81", "0.468", "0.8237", "0.0809", "0.8237" ], [ "0.2507", "0.786", "0.2769", "0.786", "0.2769", "0.7974", "0.2507", "0.7974" ], [ "0.5034", "0.811", "0.8906", "0.811", "0.8906", "0.8247", "0.5034", "0.8247" ], [ "0.08417", "0.827", "0.4673", "0.827", "0.4673", "0.837", "0.08417", "0.837" ], [ "0.5034", "0.832", "0.8965", "0.8257", "0.8965", "0.8496", "0.5034", "0.856" ], [ "0.509", "0.829", "0.585", "0.829", "0.585", "0.839", "0.509", "0.839" ], [ "0.0809", "0.8403", "0.2507", "0.8403", "0.2507", "0.854", "0.0809", "0.854" ], [ "0.505", "0.857", "0.8896", "0.857", "0.8896", "0.8706", "0.505", "0.8706" ], [ "0.1438", "0.866", "0.4346", "0.866", "0.4346", "0.884", "0.1438", "0.884" ], [ "0.4468", "0.872", "0.4673", "0.872", "0.4673", "0.882", "0.4468", "0.882" ], [ "0.505", "0.8726", "0.8906", "0.8726", "0.8906", "0.886", "0.505", "0.886" ], [ "0.1978", "0.884", "0.2296", "0.886", "0.2279", "0.8984", "0.1969", "0.8965" ], [ "0.505", "0.887", "0.652", "0.887", "0.652", "0.901", "0.505", "0.901" ] ]

[ "具体来说，我们使用梅尔频率倒谱系数（MFCC）26", "将这四个项相加：我们得到总损失函数：", "作为音频表示，并使用动态时序扭曲（DTW)E[3]算", "法通过沿时间维度拉伸或收缩MFCC特征向量来扭", "Ldis=Leross +Lself +AclaLcla +AconLcon,(6)", "曲它们。给定两个内容相同但长度不同的MFCC序列", "S。和Sb，DTW通过动态编程计算一组索引坐标对", "其中入clo和入con分别是分类和内容损失的权重", "[(i,j)，..]，以强制Sa[i]和Sb[i] 相似。给定序列之间", "3.3. Target-Adaptive Face Synthesis", "的最佳匹配是通过最小化对齐的MFCC特征之间的距", "离成本的总和来实现的：", "为了生成输人肖像的照片般逼真的面部动画，我们", "首先在[34，9]之后引人了一个音频到地标网络，该网", "min d(Sa[i], Sb[i]),", "(1)", "络可以从解开的音频嵌入中预测地标运动。之后，基于", "(ij)eP", "视频的编辑方法通常对记录在目标人物上的目标视频", "剪辑执行面部编辑。但是，这将在我们的环境中带来两", "其中d是距离成本，P是对齐的路径。路径约束是，在", "个挑战，这些挑战会改变整个脸部，而不仅仅是嘴巴", "（），有效的步骤是（+1.），（i1+1）和（+1.+1）", "1）头部姿势的错位。预测的地标的头部运动可能与目", "，确保每次至少有一个信号的对齐总是向前移动。然", "标视频严重不同，并且在音频中几乎没有提供任何姿", "后，这些对齐的音频样本可以用作解纠缠网络的输入，", "势信息。2）在合成高保真度结果的同时，难以将编辑", "以进行交叉重建", "后的人脸与目标视频混合。为了应对此类挑战，我们提", "Cross-ReconstructedTraining：交叉重建的训练过程", "出了3D感知关键点对齐算法，以将我们生成的地标与", "如图。3。为了独立提取情感和内容信息位于具有内", "3D空间中的引导地标对齐。然后，我们建议将生成的", "容i和情感m的音频剪辑ai,m中，利用两个编码器", "地标与目标图像的边缘图合并。他们两个一起可以作", "E。和E。分别嵌人这两个信息。直观地，当两个表示", "为指导，以训练边缘到视频翻译网络的最终结果", "完全解开时，我们可以使用来自音频剪辑aim和ai,n", "音频到地标模块。我们的第一个目标是从情感音频剪", "的内容嵌人Ec（ci.m）和情感嵌人Ec（ci.n）中的信息", "辑中学习地标位移，因此需要面部形状或来自对齐的", "从解码器D重建剪辑&i,n。通过利用我们之前构建的", "地标的身份信息保持不变。因此，我们使用多层感", "伪训练对，我们引人了两个新的样本ai,n，i,m作为重", "知器（MLP）提取嵌入fa的地标标识，如图E2所示。", "建过程的监督。由于每个样本只能提供一种有利于交", "然后将fa与两个解开的音频嵌入Ec(α）和Ee(c)", "重建的信息，因此最终可以实现解纠缠", "起发送到音频到地标模块。音频到地标模块通过长短", "我们使用损失函数来监督训练过程，该损失函数包", "期记忆（LSTM）网络预测地标位移ld，然后是两层", "括四个部分：交叉重建损失，自我重建损失，分类损失", "MLP", "和内容损失。给定四个音频样本ai,m，Lj,n，aj.m,Li,n，", "就损失函数而言，我们将参考地标与以下定义的", "我们将交叉重建损失表示为", "预测地标1之间的距离最小化：", "Lcross = ID(Ec(i,m), Ee(j,n)) -Li,nll2", "La =- 2 = Ila + la- l2.", "(7)", "+D(E(aj,n),Ee(i,m))-j,mll2.", "(2)", "???D感知关键点对齐。为了对齐头部姿势，我们首", "此外，我们还能够通过使用编码器和解码器来重建", "先使用现成的方法对目标视频进行地标检测E[43]。然", "原始输人，即定义为的自重建损失：", "后我们在3D空间中操作，其中明确定义了姿势信息。", "利用参数化三维人脸模型[5]解决非线性优化问题，从", "Lself = D(E(cim),Ee(ri,m)) -Ci,mll2", "(3)", "二维地标的三维参数恢复。利用3D几何和表达式参", "+ ID(Ee(ajn),Ee(ajn)) -Cj,nll2.", "数，我们得到了一组姿态不变的3D地标L3d（有关", "详细信息：请参阅补充）。姿态参数P包含一个3×3", "为了鼓励E。将具有相同情感类型的样本映射到潜", "旋转矩阵R、2个平移系数t和1个缩放系数s。通", "在空间中的聚类组中，我们为情感嵌人添加了一个分", "过将预测地标的姿态参数替换为目标视频（Rt,tt，St）", "类器Ce，并添加了一个额外的分类损失，定义为：", "中检测到的姿态参数，我们获得了适应的3D关键点，", "N", "然后将它们投影到具有比例正投影的图像平面：", "Lela = - (pis * og qk).", "(4)", "Ip = St * Pr *Rt* L3d + tt,", "(8)", "这里，N表示不同情感类型的数量，Pk表示样本是否", "k=1", "其中lp是投影的2D地标，Pr是正投影矩阵", "采用4情感。此外，我们还约束具有相同话语的样本", "（61）。由于几何和表情参数保持不变，因此", "7100)", "以共享相似的内容嵌人：", "投影的地标lp自然地与预测的地标共享一致的身份和", "Lcon = IEe(rkm)-E(ck,n)ll1:", "(5)", "面部表情。同时将头部姿势、比例和位置设置为与目标", "k=i,j", "视频顿中的面部相同" ]

{ "arxiv_id": "2104.07452", "language": "zh", "page_idx": 4 }

[ [ "0.3342", "0.1092", "0.6626", "0.1092", "0.6626", "0.1237", "0.3342", "0.1237" ], [ "0.2852", "0.1477", "0.3857", "0.1477", "0.3857", "0.1622", "0.2852", "0.1622" ], [ "0.589", "0.1465", "0.739", "0.1465", "0.739", "0.161", "0.589", "0.161" ], [ "0.1978", "0.1654", "0.4722", "0.1666", "0.4722", "0.1812", "0.1978", "0.1799" ], [ "0.524", "0.1654", "0.8", "0.1666", "0.8", "0.1812", "0.524", "0.1799" ], [ "0.4102", "0.209", "0.591", "0.209", "0.591", "0.2235", "0.4102", "0.2235" ], [ "0.361", "0.226", "0.636", "0.2267", "0.636", "0.2412", "0.361", "0.2405" ], [ "0.468", "0.2747", "0.5317", "0.2747", "0.5317", "0.286", "0.468", "0.286" ], [ "0.1708", "0.2942", "0.8506", "0.2942", "0.8506", "0.308", "0.1708", "0.308" ], [ "0.152", "0.31", "0.8486", "0.31", "0.8486", "0.3208", "0.152", "0.3208" ], [ "0.1503", "0.3225", "0.8447", "0.3225", "0.8447", "0.3333", "0.1503", "0.3333" ], [ "0.1478", "0.336", "0.851", "0.336", "0.851", "0.351", "0.1478", "0.351" ], [ "0.1487", "0.351", "0.8486", "0.3499", "0.8486", "0.363", "0.1487", "0.3643" ], [ "0.1487", "0.3643", "0.851", "0.3643", "0.851", "0.3774", "0.1487", "0.3774" ], [ "0.1478", "0.377", "0.8486", "0.377", "0.8486", "0.3901", "0.1478", "0.3901" ], [ "0.1478", "0.3914", "0.8486", "0.3926", "0.8486", "0.406", "0.1478", "0.4048" ], [ "0.1487", "0.406", "0.8506", "0.406", "0.8506", "0.4192", "0.1487", "0.4192" ], [ "0.1487", "0.4192", "0.8486", "0.4192", "0.8486", "0.433", "0.1487", "0.433" ], [ "0.1478", "0.432", "0.5107", "0.4343", "0.5107", "0.4475", "0.1478", "0.4458" ], [ "0.10455", "0.4685", "0.2998", "0.4697", "0.2998", "0.4873", "0.10455", "0.486" ], [ "0.10706", "0.502", "0.8896", "0.502", "0.8896", "0.515", "0.10706", "0.515" ], [ "0.10864", "0.5176", "0.8906", "0.5176", "0.8906", "0.531", "0.10864", "0.531" ], [ "0.1095", "0.532", "0.859", "0.532", "0.859", "0.5454", "0.1095", "0.5454" ], [ "0.1356", "0.549", "0.8896", "0.549", "0.8896", "0.5596", "0.1356", "0.5596" ], [ "0.10864", "0.5625", "0.888", "0.5625", "0.888", "0.5757", "0.10864", "0.5757" ], [ "0.1168", "0.5806", "0.125", "0.5806", "0.125", "0.586", "0.1168", "0.586" ], [ "0.1209", "0.577", "0.8906", "0.577", "0.8906", "0.5903", "0.1209", "0.5903" ], [ "0.1095", "0.5923", "0.789", "0.5923", "0.789", "0.606", "0.1095", "0.606" ], [ "0.1316", "0.607", "0.8906", "0.607", "0.8906", "0.6206", "0.1316", "0.6206" ], [ "0.1095", "0.6206", "0.8896", "0.6206", "0.8896", "0.634", "0.1095", "0.634" ], [ "0.1095", "0.636", "0.8896", "0.636", "0.8896", "0.6494", "0.1095", "0.6494" ], [ "0.1095", "0.654", "0.4224", "0.654", "0.4224", "0.664", "0.1095", "0.664" ], [ "0.1316", "0.6685", "0.8906", "0.6685", "0.8906", "0.6816", "0.1316", "0.6816" ], [ "0.1095", "0.683", "0.8896", "0.683", "0.8896", "0.6978", "0.1095", "0.6978" ], [ "0.10706", "0.696", "0.8906", "0.6978", "0.8906", "0.712", "0.10706", "0.7104" ], [ "0.1095", "0.714", "0.2059", "0.714", "0.2059", "0.726", "0.1095", "0.726" ], [ "0.1323", "0.729", "0.8906", "0.729", "0.8906", "0.742", "0.1323", "0.742" ], [ "0.1095", "0.7437", "0.8906", "0.7437", "0.8906", "0.757", "0.1095", "0.757" ], [ "0.1095", "0.757", "0.8896", "0.757", "0.8896", "0.771", "0.1095", "0.771" ], [ "0.1095", "0.7744", "0.8896", "0.7744", "0.8896", "0.785", "0.1095", "0.785" ], [ "0.1095", "0.7886", "0.8906", "0.7886", "0.8906", "0.802", "0.1095", "0.802" ], [ "0.1095", "0.8047", "0.888", "0.8047", "0.888", "0.815", "0.1095", "0.815" ], [ "0.1095", "0.8184", "0.888", "0.8184", "0.888", "0.832", "0.1095", "0.832" ], [ "0.10864", "0.834", "0.8906", "0.834", "0.8906", "0.8477", "0.10864", "0.8477" ], [ "0.1095", "0.851", "0.687", "0.851", "0.687", "0.8613", "0.1095", "0.8613" ], [ "0.125", "0.8706", "0.8906", "0.872", "0.8906", "0.8853", "0.125", "0.884" ], [ "0.10864", "0.884", "0.8906", "0.884", "0.8906", "0.898", "0.10864", "0.898" ], [ "0.10864", "0.898", "0.2115", "0.898", "0.2115", "0.9077", "0.10864", "0.9077" ], [ "0.492", "0.94", "0.5073", "0.94", "0.5073", "0.955", "0.492", "0.955" ] ]

[ "积极的因果结构学习与建议", "Davin Choo", "Themis Gouleakis", "National University of Singapore", "National University of Singapore", "Arnab Bhattacharyya", "National University of Singapore", "Abstract", "We introduce the problem of active causal structure learning with advice. In the typical well-studied", "setting,thelearningalgorithmisgiventheessentialgraphfortheobservationaldistributionandisaskedto", "recovertheunderlying causaldirectedacyclicgraph(DAG)G*whileminimizingthenumber ofintervention", "made. In our setting, we are additionally given side information about G* as advice, e.g.a DAG G", "purported to be G*.We ask whether the learning algorithm can benefit from the advice when it is close to", "being correct, while still having worst-case guarantees even when the advice is arbitrarily bad. Our work", "is in the same space as the growing body of research on algorithms with predictions . When the advice", "is a DAG G , we design an adaptive search algorithm to recover G*whose intervention cost is at most", "O(max[1,log ]) times the cost for verifying G*; here, is a distance measure between G and G* that is", "upper bounded by the number of variables n , and is exactly 0 when G = G* : Our approximation factor", "matches the state-of-the-art for the advice-less setting.", "1Introduction", "n变量的集合V上的因果有向无环图是贝叶斯网络，其中边模型直接因果效应。因果DAG不仅可以用来推", "断V的观测分布，还可以用来推断对变量V'V的任何子集进行任何干预的结果。在这项工作中，我们将", "自已限制在因果关系充分的环境中，在该环境中，没有潜在的混杂因素，没有选择偏差，也没有数据缺失", "因果结构学习的自标是从数据中恢复底层DAG。这是一个重要的问题，在多个领域的应用，包括哲学，医", "学，生物学，遗传学和计量经济学[Rei56，Hoo90，KWJ+04，Woo05，RW06，ES07，SC17，RHT+17，POS+18", "", "不幸的是，一般来说，众所周知，观测数据只能将因果DAG恢复到等价类[Pea09，SGSH00]。因此，如果", "要避免对因果机制进行参数假设，唯一的办法是从干预[EGS05，EGS06，Ebe10]获得实验数据。", "这样的考虑激发了介人设计的问题，其中任务是找到一组足以恢复因果DAG的最佳成本的介人。最近", "有一系列的工作在各种假设下研究这个问题[HG08HLV14，SKDV15，KDV17.LKDV18，GKS+19,SMG+20.", "CSB22，CS23]。特别是，假设因果充分，[CSB22]给出了一种自适应算法，该算法主动生成有界大小的干预序", "列，因此干预总数最多是①（logn）倍的最佳", "通常，在因果结构学习的大多数应用中，都有领域专家和从业者可以提供有关因果关系的其他“建议”。实际", "上，研究如何将专家建议纳人因果图发现过程的工作很长；e。g。请参见[Mee95a，SSG+98，DCJ11，FNB+11", "LB18，ASC20，FH201。在这项工作中，我们以一种有原则的方式研究了如何使用所谓的专家建议可以改进介", "入设计的算法", "在讨论我们的具体贡献之前，让我们以具有实际重要性的具体问题为基础进行上述讨论。在现代虚拟化基", "础架构中，将应用程序模块化为大量相互依赖的微服务越来越普遍。这些微服务以依赖于应用程序代码和触发", "用户流的方式相互通信。至关重要的是，平台提供商通常不知道微服务之间的通信图，因为应用程序代码可能", "是私有的并且属于不同的实体。但是，了解该图对于各种关键的平台级任务很有用，例如故障定位ZPX+191", "主动探测[TJG+19]，测试[JBT+19]和污点分析【CLO07]。最近，【WAJ+23]和[ICM+22]建议将微服务通", "信图视为稀疏因果DAG。特别是，WAJ+23表明，可以在暂存环境中将任意于预实现为故障注人，因此可以", "部署因果结构学习算法来生成足以学习底层通信图的干预序列。在这样的设置中，很自然地假设平台提供商已", "经对图形有了近似的猜测，例如。g。在先前运行的算法中发现的图或由公共元数据标记微服务代码建议的图。", "我们提出的研究计划是设计因果结构学习算法，可以利用这种可能不完美的建议", "1Note however that the system in [WAJ+23] is not causally sufficient due to confounding user behavior and [ICM+22] does", "not actively perform interventions: So, the algorithm proposed in this work cannot be used directly for the microservices graph", "learningproblem", "2" ]

{ "arxiv_id": "2305.19588", "language": "zh", "page_idx": 1 }

[ [ "0.1095", "0.05997", "0.1602", "0.05997", "0.1602", "0.07135", "0.1095", "0.07135" ], [ "0.1609", "0.05936", "0.335", "0.05936", "0.335", "0.0726", "0.1609", "0.0726" ], [ "0.10864", "0.0833", "0.8896", "0.0833", "0.8896", "0.0966", "0.10864", "0.0966" ], [ "0.10864", "0.09784", "0.1438", "0.09784", "0.1438", "0.113", "0.10864", "0.113" ], [ "0.1299", "0.1237", "0.8906", "0.1256", "0.8906", "0.1401", "0.1299", "0.1383" ], [ "0.1478", "0.1414", "0.8906", "0.1414", "0.8906", "0.1547", "0.1478", "0.1547" ], [ "0.1478", "0.1559", "0.8906", "0.1559", "0.8906", "0.1692", "0.1478", "0.1692" ], [ "0.1487", "0.1718", "0.892", "0.1718", "0.892", "0.1849", "0.1487", "0.1849" ], [ "0.1487", "0.1863", "0.892", "0.1863", "0.892", "0.1995", "0.1487", "0.1995" ], [ "0.1478", "0.202", "0.8906", "0.202", "0.8906", "0.2153", "0.1478", "0.2153" ], [ "0.1462", "0.2166", "0.785", "0.2166", "0.785", "0.2299", "0.1462", "0.2299" ], [ "0.134", "0.2412", "0.892", "0.2412", "0.892", "0.2551", "0.134", "0.2551" ], [ "0.1478", "0.2568", "0.892", "0.2568", "0.892", "0.2708", "0.1478", "0.2708" ], [ "0.1487", "0.2715", "0.892", "0.2715", "0.892", "0.2854", "0.1487", "0.2854" ], [ "0.1478", "0.286", "0.892", "0.2854", "0.892", "0.2998", "0.1478", "0.3005" ], [ "0.1478", "0.3018", "0.892", "0.3018", "0.892", "0.315", "0.1478", "0.315" ], [ "0.1478", "0.3176", "0.892", "0.3176", "0.892", "0.3308", "0.1478", "0.3308" ], [ "0.1487", "0.332", "0.5923", "0.332", "0.5923", "0.3452", "0.1487", "0.3452" ], [ "0.1365", "0.358", "0.8906", "0.358", "0.8906", "0.3718", "0.1365", "0.3718" ], [ "0.1462", "0.3726", "0.892", "0.3726", "0.892", "0.3865", "0.1462", "0.3865" ], [ "0.1478", "0.3865", "0.892", "0.387", "0.892", "0.4016", "0.1478", "0.401" ], [ "0.1487", "0.4016", "0.8906", "0.4016", "0.8906", "0.4153", "0.1487", "0.4153" ], [ "0.1478", "0.4167", "0.886", "0.4172", "0.886", "0.432", "0.1478", "0.4312" ], [ "0.1487", "0.433", "0.892", "0.433", "0.892", "0.4463", "0.1487", "0.4463" ], [ "0.1487", "0.4475", "0.892", "0.4475", "0.892", "0.461", "0.1487", "0.461" ], [ "0.1487", "0.4634", "0.8906", "0.4634", "0.8906", "0.4766", "0.1487", "0.4766" ], [ "0.1478", "0.476", "0.697", "0.4778", "0.697", "0.4924", "0.1478", "0.4905" ], [ "0.1323", "0.506", "0.8896", "0.506", "0.8896", "0.5195", "0.1323", "0.5195" ], [ "0.10864", "0.521", "0.8896", "0.521", "0.8896", "0.534", "0.10864", "0.534" ], [ "0.10864", "0.536", "0.8906", "0.536", "0.8906", "0.55", "0.10864", "0.55" ], [ "0.1054", "0.5684", "0.3447", "0.5684", "0.3447", "0.5835", "0.1054", "0.5835" ], [ "0.10864", "0.5923", "0.8906", "0.5923", "0.8906", "0.606", "0.10864", "0.606" ], [ "0.1095", "0.6094", "0.8906", "0.6094", "0.8906", "0.6235", "0.1095", "0.6235" ], [ "0.10706", "0.6226", "0.1772", "0.625", "0.1772", "0.6396", "0.10706", "0.637" ], [ "0.10706", "0.662", "0.5293", "0.662", "0.5293", "0.6797", "0.10706", "0.6797" ], [ "0.1095", "0.6924", "0.8896", "0.6924", "0.8896", "0.7056", "0.1095", "0.7056" ], [ "0.1095", "0.709", "0.888", "0.709", "0.888", "0.7236", "0.1095", "0.7236" ], [ "0.10864", "0.7236", "0.8906", "0.7236", "0.8906", "0.7373", "0.10864", "0.7373" ], [ "0.10864", "0.738", "0.892", "0.738", "0.892", "0.7524", "0.10864", "0.7524" ], [ "0.10864", "0.754", "0.8906", "0.754", "0.8906", "0.7676", "0.10864", "0.7676" ], [ "0.10864", "0.768", "0.892", "0.768", "0.892", "0.7827", "0.10864", "0.7827" ], [ "0.10864", "0.784", "0.888", "0.784", "0.888", "0.7974", "0.10864", "0.7974" ], [ "0.1095", "0.8", "0.8896", "0.8", "0.8896", "0.813", "0.1095", "0.813" ], [ "0.1095", "0.8145", "0.8613", "0.8145", "0.8613", "0.8276", "0.1095", "0.8276" ], [ "0.1095", "0.83", "0.8896", "0.83", "0.8896", "0.8433", "0.1095", "0.8433" ], [ "0.10864", "0.8447", "0.8906", "0.8447", "0.8906", "0.8594", "0.10864", "0.8594" ], [ "0.10706", "0.86", "0.581", "0.86", "0.581", "0.8735", "0.10706", "0.8735" ], [ "0.492", "0.9395", "0.5073", "0.9395", "0.5073", "0.954", "0.492", "0.954" ] ]

[ "1.1", "Our contributions", "在这项工作中，我们研究了自适应干预设计，用于在专家的建议下恢复非参数因果图。具体而言，我们的贡献", "如下", ". Problem Formulation : Our work connects the causal structure learning problem with the burgeoning", "research area of algorithms with predictions or learning-augmented algorithms [MV22] where the goal", "is to design algorithms that bypass worst-case behavior by taking advantage of (possibly erroneous)", "advice or predictions about the problem instance. Most work in this area has been restricted to online", "algorithms, data structure design, or optimization, as described later in Section 2.5 : However, as we", "motivated above, expert advice is highly relevant for causal discovery, and to the best of our knowlege,", "ours is the first attempt to formally address the issue of imperfect advice in this context.", "Adaptive Search Algorithm . We consider the setting where the advice is a DAG G purported to be the", "orientations of all the edges in the graph. We define a distance measure which is always bounded by n ,", "the number of variables, and equals 0 when G = G*. For any integer k ≥ 1 , we propose an adaptive", "algorithm to generate a sequence of interventions of size at most k that recovers the true DAG G* , such", "that the total number of interventions is O(log b(G, G*) -log k) times the optimal number of interventions", "of size k : Thus, our approximation factor is never worse than the factor for the advice-less setting in", "[CSB22] . Our search algorithm also runs in polynomial time.", ".Verification Cost Approximation . For a given upper bound k ≥ 1 , a verifying intervention set for a DAG", "G* is a set of interventions of size at most k that, together with knowledge of the Markov equivalence", "class of G* , determines the orientations of all edges in G* . The minimum size of a verifying intervention", "set for G* , denoted vs(G*) , is clearly a lower bound for the number of interventions required to learn", "G* (regardless of the advice graph G ). One of our key technical results is a structural result about V1", "We prove that for any two DAGs G and G' within the same Markov equivalence class, we always have", "Vi(G) ≤ 2 vi(G') and that this is tight in the worst case. Beyond an improved structural understanding", "of minimum verifying intervention sets, which we believe is of independent interest, this enables us to", "\"blindly trust\" the information provided by imperfect advice to some extent.", "类似于以前的作品（e。g。[SMG+20，CSB22，CS23]），我们在使用理想干预的同时假设因果充分和忠诚。", "在这些假设下，运行标准因果发现算法（例如。g。PC[SGSH00l，GES[Chi02]）将始终成功地从数据中恢复", "正确的基本图形。我们还假设给定的专家建议与观察本质图一致。。有关我们假设的讨论，请参见AppendixA", "1.2Paper organization", "在Section2中，我们将初步概念与相关工作穿插在一起。我们的主要结果在Section3中给出，具有Section4", "中给出的高级技术思想和直觉。Section5提供了一些经验验证。有关完整的证明，源代码和实验详细信息，请", "参阅附录", "2Preliminaries and Related Work", "关于图和因果模型的基本概念在AppendixB中定义。简而言之，如果G=（V.E）是IVI=n节点/顶点上的", "图，其中V(G)，E(G)和A(G)CE(G)分别表示G的节点，边和弧，我们将～表示两个节点u,UEV", "在G中连接，并在IVI=n中指定某个方向。骨架skel(G)指的是基础图，其中所有边都是无向的。G中的", "v结构是指三个不同顶点u,u,wEV的集合，例如u→U←w和uw。让G=(V,E）完全无方向。对", "于顶点u,uEV，顶点VV的子集和整数r≥O，我们将distG(u,u）定义为u和之间的最短路径长", "度，并将N(V'）=（uEV：minueVdistG(u,）≤r)V定义为r的顶点集-跳远离XMATHX_。有向无", "环图（DAG）是没有有向循环的完全定向图。对于任何DAGG，我们用[GI表示其马尔可夫等价类（MEC）", "用E(G表示基本图。同一MEC中的DAG具有相同的骨架，并且基本图是部分有向图，因此如果MEC[G]", "中的每个DAG中的u→都指向弧u→U，并且如果在MAXXXXXMAX3_4中存在两个DAGu~", "已知当且仅当两个图具有相同的骨架和v结构【VP90，AMP97并且可以通过在skel(G)中定向V结构并应用", "Meek规则从G计算基本图E(G）时，它们才是马尔可夫等效的（请参阅AppendixD）。在DAGG中，如果", "Pa(u)=Pa(u)/(u)。我们使用C(G)CE(G)表示G的覆盖边集", "3" ]

{ "arxiv_id": "2305.19588", "language": "zh", "page_idx": 2 }

[ [ "0.10706", "0.05807", "0.343", "0.05936", "0.343", "0.07385", "0.10706", "0.0726" ], [ "0.10864", "0.0821", "0.888", "0.0821", "0.888", "0.09534", "0.10864", "0.09534" ], [ "0.10864", "0.09784", "0.8906", "0.09784", "0.8906", "0.1111", "0.10864", "0.1111" ], [ "0.10864", "0.113", "0.8774", "0.113", "0.8774", "0.127", "0.10864", "0.127" ], [ "0.10706", "0.127", "0.8906", "0.1276", "0.8906", "0.1427", "0.10706", "0.1414" ], [ "0.1111", "0.1459", "0.8896", "0.1459", "0.8896", "0.1559", "0.1111", "0.1559" ], [ "0.1095", "0.159", "0.621", "0.159", "0.621", "0.1724", "0.1095", "0.1724" ], [ "0.1323", "0.1749", "0.3962", "0.1749", "0.3962", "0.1881", "0.1323", "0.1881" ], [ "0.10864", "0.1963", "0.8896", "0.1963", "0.8896", "0.2108", "0.10864", "0.2108" ], [ "0.1054", "0.2108", "0.8906", "0.2122", "0.8906", "0.2267", "0.1054", "0.226" ], [ "0.10706", "0.2279", "0.8857", "0.2279", "0.8857", "0.2412", "0.10706", "0.2412" ], [ "0.10864", "0.2507", "0.892", "0.2507", "0.892", "0.2644", "0.10864", "0.2644" ], [ "0.1111", "0.2664", "0.8906", "0.2664", "0.8906", "0.2798", "0.1111", "0.2798" ], [ "0.10864", "0.2798", "0.4004", "0.2798", "0.4004", "0.2935", "0.10864", "0.2935" ], [ "0.134", "0.3062", "0.8896", "0.3062", "0.8896", "0.3164", "0.134", "0.3164" ], [ "0.1095", "0.3193", "0.8906", "0.3193", "0.8906", "0.334", "0.1095", "0.334" ], [ "0.10864", "0.3333", "0.8906", "0.334", "0.8906", "0.3484", "0.10864", "0.348" ], [ "0.1095", "0.3499", "0.785", "0.3499", "0.785", "0.363", "0.1095", "0.363" ], [ "0.1316", "0.363", "0.8906", "0.3643", "0.8906", "0.379", "0.1316", "0.3774" ], [ "0.10864", "0.3801", "0.8906", "0.3801", "0.8906", "0.3933", "0.10864", "0.3933" ], [ "0.10706", "0.3945", "0.892", "0.3933", "0.892", "0.4077", "0.10706", "0.4092" ], [ "0.1095", "0.4104", "0.8906", "0.4104", "0.8906", "0.4236", "0.1095", "0.4236" ], [ "0.1095", "0.4255", "0.8896", "0.4255", "0.8896", "0.4395", "0.1095", "0.4395" ], [ "0.10864", "0.44", "0.888", "0.44", "0.888", "0.4539", "0.10864", "0.4539" ], [ "0.10864", "0.4558", "0.638", "0.4558", "0.638", "0.4697", "0.10864", "0.4697" ], [ "0.1095", "0.4792", "0.706", "0.4792", "0.706", "0.4924", "0.1095", "0.4924" ], [ "0.2834", "0.5024", "0.714", "0.504", "0.714", "0.518", "0.2834", "0.5176" ], [ "0.10864", "0.53", "0.8906", "0.53", "0.8906", "0.5435", "0.10864", "0.5435" ], [ "0.1054", "0.5625", "0.2932", "0.5625", "0.2932", "0.577", "0.1054", "0.577" ], [ "0.10864", "0.5884", "0.8896", "0.5884", "0.8896", "0.6016", "0.10864", "0.6016" ], [ "0.1095", "0.603", "0.8906", "0.603", "0.8906", "0.616", "0.1095", "0.616" ], [ "0.10864", "0.6187", "0.744", "0.6187", "0.744", "0.632", "0.10864", "0.632" ], [ "0.1054", "0.64", "0.8896", "0.638", "0.8896", "0.655", "0.1054", "0.657" ], [ "0.1054", "0.655", "0.728", "0.656", "0.728", "0.6704", "0.1054", "0.6694" ], [ "0.134", "0.6787", "0.8896", "0.6787", "0.8896", "0.6924", "0.134", "0.6924" ], [ "0.10864", "0.6943", "0.8906", "0.6943", "0.8906", "0.7085", "0.10864", "0.7085" ], [ "0.1095", "0.7104", "0.8906", "0.7104", "0.8906", "0.7236", "0.1095", "0.7236" ], [ "0.10864", "0.7246", "0.3677", "0.7246", "0.3677", "0.738", "0.10864", "0.738" ], [ "0.10864", "0.7485", "0.8906", "0.7485", "0.8906", "0.762", "0.10864", "0.762" ], [ "0.1095", "0.763", "0.8906", "0.763", "0.8906", "0.778", "0.1095", "0.778" ], [ "0.10706", "0.778", "0.8906", "0.779", "0.8906", "0.7935", "0.10706", "0.7925" ], [ "0.10706", "0.7944", "0.2263", "0.7944", "0.2263", "0.808", "0.10706", "0.808" ], [ "0.1323", "0.8164", "0.8906", "0.8164", "0.8906", "0.8306", "0.1323", "0.8306" ], [ "0.10864", "0.8335", "0.367", "0.8335", "0.367", "0.8467", "0.10864", "0.8467" ], [ "0.1217", "0.854", "0.8906", "0.855", "0.8906", "0.869", "0.1217", "0.8687" ], [ "0.10706", "0.866", "0.4192", "0.867", "0.4192", "0.881", "0.10706", "0.88" ], [ "0.125", "0.879", "0.8906", "0.879", "0.8906", "0.893", "0.125", "0.893" ], [ "0.10864", "0.8926", "0.3203", "0.8945", "0.3203", "0.905", "0.10864", "0.903" ], [ "0.1259", "0.904", "0.501", "0.904", "0.501", "0.9175", "0.1259", "0.9175" ], [ "0.492", "0.94", "0.509", "0.94", "0.509", "0.955", "0.492", "0.955" ] ]

[ "2.1Ideal interventions", "于预SCV是一个实验，其中所有变量SES被强制设置为某个值，与潜在的因果结构无关。对于某些", "k≥1，如果IS=1，则干预是原子的，如果S≤k，则干预是有界的；观测数据是S=の的特殊情", "况。Pearl的do-calculus[Pea09]正式捕获了干预措施的效果。我们称任何工C2为干预集：干预集是", "组干预，其中每个干预对应于变量的子集。在G中对SCV的理想干预会产生一个干预图Gs，其中所", "有进入顶点ES的弧都被删除EGSO5。众所周知，通过对S进行十预，我们可以推断出S和VS", "[Ebe07，HEH13，HLV14，SKDV15，KDV17]切割的任何边缘的边缘方向", "我们现在给出图分隔符的定义和结果", "Definition 1 (α -separator and Q -clique separator, Definition 19 from [CSB22]). Let A, B,C be a partition", "of the vertices V of a graph G = (V, E) . We say that C is an α -separator if no edge joins a vertex in A with", "a vertex in B and Al, Bl≤ q -V- We call C is an α -clique separator if it is an α -separator and a clique.", "Theorem 2 ([GRE84l, instantiated for unweighted graphs). Let G = (V,E) be a chordal graph with [V|≥ 2", "and p vertices in its largest clique. There erists a 1/2 -clique-separator C involving at most p -1 vertices. The", "clique C can be computed in O(IEl) time", "对于理想的干预措施，G的工基本图&（G）是表示图的马尔可夫等价类的基本图，其每个干预措施的干预", "图的马尔可夫等效于任何干预措施SE工的Gs。关于工-基本图属性[HB12,HB14]有几个已知的属性：每", "个工-基本图是一个带有弦2链组件的链图3。这包括工=の的情况。一个链组件中的方向不会影响其他组件", "中的方向。换句话说，要完全定向任何基本图S（G*），必须并充分定向8（G*）中的每个链组件", "对于任何干预集工C2V，我们编写R(G,I)=A(&r(G))CE来表示DAGG的工基本图中的定向弧集。", "对于更清晰的表示法，我们为某些ICV编写用于单干预的R(G,I)I={I}，为某些EV编写用于单原子", "千预的工={{u})。对于任何介人集工C2V，将GI=G[E R(G,工)]定义为由&z(G中的无向弧诱导的完", "全有向子图DAG，其中G是由于v结构而在观察本质图中去除所有定向弧后获得的图。有关示例，请参见", "Fig.1。在R(）的表示法中，以下结果证明仅在没有V结构的DAG上通过理想干预来研究验证和自适应搜", "索是合理的，即。e。道德DAG（Definition 4)：自R(G,L)=R(Go,工)UR(G,の）以来，在执行任何干预之", "前，可以删除观察图中的任何定向弧，因为解决方案的最优性不受影响。4", "Theorem 3 ([CS23]). For any DAG G =(V,E) and intervention sets A,B 2V,", "R(G,AU B) = R(GA, B) U R(GB,A) U(R(G,A)N R(G,B))", "Definition 4 (Moral DAG). A DAG G is called a moral DAG if it has no v-structures. So, &(G) = skel(G) .", "2.2Verifying sets", "DAGGE[G*的验证集工是一个干预集，它从C(G*）完全定向G，可能重复应用Meek规则（请参阅", "AppendixD），i。e。(G*）=G*。此外，如果工是G*的验证集，则IUS对于任何其他干预SCV也", "是如此。虽然通常可能有多个验证集，但我们通常有兴趣找到一个具有最小大小的验证集", "Definition 5 (Minimum size verifying set). An intervention set I 2V is called a verifying set for a DAG G*", "if &r(G*)= G*I is a minimum size verifying set if &r (G*) G* for any [T'|<|| :", "对于有界大小的干预，最小验证数Vs（G)表示任何DAGGE[G*1的最小大小验证集的大小；我们为原子", "干预编写Vi(G）。也就是说，在&(G*）上执行千预时，任何显示的弧方向都会尊重G。[CSB22]告诉我们，为", "了验证DAGG，有必要和充分地干预覆盖边缘C(G）的最小顶点覆盖，并且Vi(G)在给定G的情况下是可", "有效计算的，因为C(G)诱导了森林", "Theorem 6 ([CSB22l). Fic an essential graph &(G*) and G E [G*l : An atomic intervention set I is a", "minimal sized verifying set for G if and only if I is a minimum verter cover of covered edges C(G) of G .", "A minimal sized atomic verifying set can be computed in polynomial time since the edge-induced subgraph on", "C(G) is a forest.", "对于任何DAGG，我们使用V(G)2V表示G的所有原子验证集的集合。也就是说，V(G)中的每个原", "子干预集都是C(G)的最小顶点覆盖", "2A chordal graph is a graph where every cycle of length at least 4 has an edge that is not part of the cycle but connects two", "vertices of the cycle; see [BP93] for an introduction", "3A partially directed graph is a chain graph if it does not contain any partially directed cycles where all directed arcs point in", "the same direction along the cycle.", "4The notation AU B denotes disjoint union of sets A and B .", "4" ]

{ "arxiv_id": "2305.19588", "language": "zh", "page_idx": 3 }

[ [ "0.4395", "0.1925", "0.4697", "0.1925", "0.4697", "0.209", "0.4395", "0.209" ], [ "0.638", "0.1925", "0.675", "0.1925", "0.675", "0.209", "0.638", "0.209" ], [ "0.8364", "0.1925", "0.8716", "0.1925", "0.8716", "0.209", "0.8364", "0.209" ], [ "0.10706", "0.2437", "0.892", "0.2437", "0.892", "0.2568", "0.10706", "0.2568" ], [ "0.2427", "0.1932", "0.2646", "0.1932", "0.2646", "0.2084", "0.2427", "0.2084" ], [ "0.10864", "0.2588", "0.4102", "0.2588", "0.4102", "0.2727", "0.10864", "0.2727" ], [ "0.407", "0.2588", "0.892", "0.2588", "0.892", "0.2727", "0.407", "0.2727" ], [ "0.10864", "0.2734", "0.892", "0.2734", "0.892", "0.2886", "0.10864", "0.2886" ], [ "0.1095", "0.289", "0.8906", "0.289", "0.8906", "0.303", "0.1095", "0.303" ], [ "0.1095", "0.3037", "0.892", "0.3037", "0.892", "0.3176", "0.1095", "0.3176" ], [ "0.10864", "0.318", "0.822", "0.318", "0.822", "0.3333", "0.10864", "0.3333" ], [ "0.10706", "0.3599", "0.1609", "0.3599", "0.1609", "0.375", "0.10706", "0.375" ], [ "0.1609", "0.3599", "0.5605", "0.3599", "0.5605", "0.375", "0.1609", "0.375" ], [ "0.1111", "0.3838", "0.8906", "0.3838", "0.8906", "0.3984", "0.1111", "0.3984" ], [ "0.1095", "0.3997", "0.3635", "0.3997", "0.3635", "0.4136", "0.1095", "0.4136" ], [ "0.1323", "0.4153", "0.8164", "0.4153", "0.8164", "0.4287", "0.1323", "0.4287" ], [ "0.1095", "0.44", "0.8896", "0.44", "0.8896", "0.4539", "0.1095", "0.4539" ], [ "0.10864", "0.455", "0.892", "0.455", "0.892", "0.4697", "0.10864", "0.4697" ], [ "0.1095", "0.4697", "0.4805", "0.4697", "0.4805", "0.4841", "0.1095", "0.4841" ], [ "0.1095", "0.4849", "0.339", "0.4849", "0.339", "0.4988", "0.1095", "0.4988" ], [ "0.10706", "0.5", "0.411", "0.4988", "0.411", "0.513", "0.10706", "0.5146" ], [ "0.134", "0.5244", "0.8906", "0.5244", "0.8906", "0.539", "0.134", "0.539" ], [ "0.10706", "0.539", "0.761", "0.5405", "0.761", "0.5547", "0.10706", "0.5537" ], [ "0.10864", "0.565", "0.8906", "0.565", "0.8906", "0.58", "0.10864", "0.58" ], [ "0.10706", "0.5806", "0.8906", "0.5806", "0.8906", "0.5947", "0.10706", "0.5947" ], [ "0.10706", "0.5967", "0.8906", "0.5967", "0.8906", "0.61", "0.10706", "0.61" ], [ "0.1323", "0.636", "0.8896", "0.636", "0.8896", "0.6494", "0.1323", "0.6494" ], [ "0.10864", "0.651", "0.4028", "0.651", "0.4028", "0.6655", "0.10864", "0.6655" ], [ "0.10864", "0.6753", "0.589", "0.6753", "0.589", "0.69", "0.10864", "0.69" ], [ "0.581", "0.677", "0.892", "0.677", "0.892", "0.69", "0.581", "0.69" ], [ "0.10706", "0.6914", "0.8813", "0.69", "0.8813", "0.705", "0.10706", "0.7056" ], [ "0.1095", "0.707", "0.8936", "0.707", "0.8936", "0.72", "0.1095", "0.72" ], [ "0.10706", "0.72", "0.4592", "0.72", "0.4592", "0.735", "0.10706", "0.735" ], [ "0.10864", "0.7363", "0.4265", "0.7363", "0.4265", "0.7505", "0.10864", "0.7505" ], [ "0.10706", "0.7505", "0.492", "0.752", "0.492", "0.7666", "0.10706", "0.765" ], [ "0.134", "0.778", "0.6587", "0.778", "0.6587", "0.7915", "0.134", "0.7915" ], [ "0.10864", "0.8125", "0.1602", "0.8125", "0.1602", "0.8237", "0.10864", "0.8237" ], [ "0.1569", "0.81", "0.546", "0.811", "0.546", "0.8257", "0.1569", "0.8247" ], [ "0.10864", "0.834", "0.681", "0.8354", "0.681", "0.8496", "0.10864", "0.8486" ], [ "0.1274", "0.8613", "0.3816", "0.8613", "0.3816", "0.875", "0.1274", "0.875" ], [ "0.1299", "0.8853", "0.8896", "0.8853", "0.8896", "0.8984", "0.1299", "0.8984" ], [ "0.10706", "0.896", "0.4968", "0.8965", "0.4968", "0.91", "0.10706", "0.909" ], [ "0.4927", "0.94", "0.5063", "0.94", "0.5063", "0.952", "0.4927", "0.952" ] ]

[ "(II)", "(III)", "(IV)", "Figure 1: (I) Ground truth DAG G* ; (II) Observational essential graph &(G*) where C → E D is a", "(I)", "v-structure and Meek rules orient arcs D", "→ F and E→ F ; (III) G = G[ER(G,O)l where oriented arcs", "in (G*) are removed from G* ; (IV) MPDAG G E [G*] incorporating the following partial order advice", "(Si = B), S2 = {A,D],S3 = {C,E,F}) , which can be converted to required arcs B -→ A and B → D .", "Observe that A → C is oriented by Meek Rl via B → A ~ C,the arc A~ D is still unoriented, the arc", "B -→ A disagrees with G* , and there are two possible DAGs consistent with the resulting MPDAG.", "2.3", "Adaptive search using ideal interventions", "自适应搜索算法已在认真的[HG08，HB14SKDV15,SMG+20，CSB22，CS23]中进行了研究，因为它们可以使", "用比非自适应对应物少得多的干预。", "最近，[CSB22]给出了一种有效的算法，用于在一般图上计算具有可证明近似保证的自适应干预", "Theorem 7 ([CSB22l). Fia an unknown underlying DAGG* : Given an essential graph &(G*) and intervention", "set bound k ≥ 1 , there is a deterministic polynomial time algorithm that computes an intervention set I", "adaptively such that &r(G*) = G* , and [| has size", "1. O(log(n) ·Vi(G*)) when k =", "2. O(log(n) - log(k) Vk(G*)) when k > 1 .", "同时，在局部因果图发现的上下文中，人们只对学习因果关系的子集感兴趣，[CS23]的SubsetSearch算", "法会产生乘法开销，当在节点诱导的子图中定向边时，该开销与相关节点的数量成对数比例", "Definition 8 (Relevant nodes). Fix a DAG G* = (V,E) and arbitrary subset V' C V . For any intervention", "set I C 2V and resulting interventional essential graph &r(G*), we define the relevant nodes p(L, V') C V! as", "the set of nodes within V' that is adjacent to some unoriented arc within the node-induced subgraph &r(G*)[v'l", "有关相关节点的示例，请参见Fig.1：对于（I中的子集V'={A,C,D,E,F)，只有A,C,D是相关", "的，因为E和F的人射边缘都是定向的", "Theorem 9 ([CS23). Fir an unknown underlying DAG G*", "Given an interventional essential graph", "&r(G*) , node-induced subgraph H with relevant nodes p(,V(H)) and intervention set bound k ≥ 1", "there is a deterministic polynomial time algorithm that computes an intervention set I adaptively such that", "&rui'(G*)[V(H)] = G*[V(H)l , and [I'I has size", "1. O(log(le(L, V(H))I) - Vi(G*)) when k = 1", "2. O(log(le(I, V(H)D) - log(k) - Vk(G*)) when k > 1 ", "请注意，k=1指的是原子干预的设置，我们总是有0≤lp(I,V(H))≤n", "2.4", "Expert advice in causal graph discovery", "领域专家可以提供三种主要类型的信息（e。g。请参见Section1中给出的引用）：", "(I) Required parental arcs: X -→ Y", "If the essential graph E(G*) is a path of n nodes, then non-adaptive algorithms need 2(n) atomic interventions to recover G*", "while O(logn) atomic interventions suffices for adaptive search.", "5" ]

{ "arxiv_id": "2305.19588", "language": "zh", "page_idx": 4 }

[ [ "0.1511", "0.0873", "0.8447", "0.0885", "0.8447", "0.1027", "0.1511", "0.10156" ], [ "0.4685", "0.11755", "0.531", "0.11755", "0.531", "0.1318", "0.4685", "0.1318" ], [ "0.6885", "0.11755", "0.81", "0.11755", "0.81", "0.1318", "0.6885", "0.1318" ], [ "0.1957", "0.1342", "0.3064", "0.1342", "0.3064", "0.1484", "0.1957", "0.1484" ], [ "0.1838", "0.11816", "0.318", "0.11816", "0.318", "0.133", "0.1838", "0.133" ], [ "0.444", "0.1342", "0.555", "0.1342", "0.555", "0.1484", "0.444", "0.1484" ], [ "0.6963", "0.1342", "0.804", "0.1342", "0.804", "0.1484", "0.6963", "0.1484" ], [ "0.1394", "0.152", "0.361", "0.1509", "0.361", "0.165", "0.1394", "0.1663" ], [ "0.424", "0.152", "0.5728", "0.1509", "0.5728", "0.165", "0.424", "0.1663" ], [ "0.6396", "0.1526", "0.86", "0.1526", "0.86", "0.1663", "0.6396", "0.1663" ], [ "0.294", "0.1812", "0.3704", "0.1841", "0.3687", "0.1996", "0.293", "0.1965" ], [ "0.6113", "0.1799", "0.721", "0.1829", "0.721", "0.199", "0.6113", "0.1953" ], [ "0.2913", "0.199", "0.372", "0.199", "0.372", "0.2131", "0.2913", "0.2131" ], [ "0.613", "0.1996", "0.721", "0.1996", "0.721", "0.2144", "0.613", "0.2144" ], [ "0.241", "0.2173", "0.4224", "0.2173", "0.4224", "0.231", "0.241", "0.231" ], [ "0.575", "0.2173", "0.7583", "0.2173", "0.7583", "0.231", "0.575", "0.231" ], [ "0.2627", "0.2542", "0.34", "0.2542", "0.34", "0.2683", "0.2627", "0.2683" ], [ "0.513", "0.2546", "0.881", "0.2546", "0.881", "0.2683", "0.513", "0.2683" ], [ "0.513", "0.2695", "0.879", "0.2695", "0.879", "0.2832", "0.513", "0.2832" ], [ "0.1453", "0.282", "0.4568", "0.282", "0.4568", "0.2957", "0.1453", "0.2957" ], [ "0.513", "0.2861", "0.881", "0.2861", "0.881", "0.2998", "0.513", "0.2998" ], [ "0.1453", "0.2964", "0.4568", "0.2964", "0.4568", "0.31", "0.1453", "0.31" ], [ "0.513", "0.3027", "0.881", "0.3027", "0.881", "0.3164", "0.513", "0.3164" ], [ "0.147", "0.31", "0.4575", "0.31", "0.4575", "0.3237", "0.147", "0.3237" ], [ "0.513", "0.3196", "0.881", "0.3196", "0.881", "0.3333", "0.513", "0.3333" ], [ "0.1486", "0.326", "0.455", "0.326", "0.455", "0.3362", "0.1486", "0.3362" ], [ "0.147", "0.3384", "0.4575", "0.3384", "0.4575", "0.3516", "0.147", "0.3516" ], [ "0.5146", "0.335", "0.881", "0.335", "0.881", "0.3486", "0.5146", "0.3486" ], [ "0.1453", "0.3528", "0.4568", "0.3528", "0.4568", "0.3665", "0.1453", "0.3665" ], [ "0.513", "0.3516", "0.881", "0.3516", "0.881", "0.3652", "0.513", "0.3652" ], [ "0.1453", "0.3652", "0.4575", "0.3677", "0.4575", "0.3813", "0.1453", "0.379" ], [ "0.5146", "0.3677", "0.881", "0.3677", "0.881", "0.3813", "0.5146", "0.3813" ], [ "0.1453", "0.3801", "0.4568", "0.3813", "0.4568", "0.3943", "0.1453", "0.3938" ], [ "0.513", "0.3843", "0.879", "0.3843", "0.879", "0.398", "0.513", "0.398" ], [ "0.147", "0.3955", "0.4592", "0.3955", "0.4592", "0.4104", "0.147", "0.4104" ], [ "0.5146", "0.3997", "0.881", "0.3997", "0.881", "0.4133", "0.5146", "0.4133" ], [ "0.147", "0.4092", "0.4575", "0.4104", "0.4575", "0.4233", "0.147", "0.4229" ], [ "0.513", "0.4163", "0.8774", "0.4163", "0.8774", "0.43", "0.513", "0.43" ], [ "0.1453", "0.4233", "0.4592", "0.4233", "0.4592", "0.437", "0.1453", "0.437" ], [ "0.513", "0.4316", "0.881", "0.4316", "0.881", "0.4453", "0.513", "0.4453" ], [ "0.1453", "0.4382", "0.4568", "0.4382", "0.4568", "0.452", "0.1453", "0.452" ], [ "0.513", "0.4478", "0.881", "0.4478", "0.881", "0.4614", "0.513", "0.4614" ], [ "0.1486", "0.4536", "0.455", "0.4536", "0.455", "0.4644", "0.1486", "0.4644" ], [ "0.147", "0.466", "0.455", "0.4656", "0.455", "0.4785", "0.147", "0.4797" ], [ "0.513", "0.4644", "0.879", "0.4644", "0.879", "0.478", "0.513", "0.478" ], [ "0.1436", "0.4785", "0.4568", "0.4797", "0.4568", "0.4946", "0.1436", "0.4934" ], [ "0.511", "0.4797", "0.879", "0.4797", "0.879", "0.4934", "0.511", "0.4934" ], [ "0.1453", "0.4954", "0.4575", "0.4946", "0.4575", "0.5083", "0.1453", "0.509" ], [ "0.5146", "0.4963", "0.881", "0.4963", "0.881", "0.5103", "0.5146", "0.5103" ], [ "0.1453", "0.511", "0.4568", "0.509", "0.4568", "0.5195", "0.1453", "0.5215" ], [ "0.513", "0.5127", "0.881", "0.5127", "0.881", "0.5254", "0.513", "0.5254" ], [ "0.1486", "0.525", "0.4534", "0.525", "0.4534", "0.535", "0.1486", "0.535" ], [ "0.513", "0.527", "0.882", "0.528", "0.882", "0.542", "0.513", "0.5415" ], [ "0.147", "0.538", "0.4534", "0.538", "0.4534", "0.549", "0.147", "0.549" ], [ "0.513", "0.5444", "0.7188", "0.5444", "0.7188", "0.558", "0.513", "0.558" ], [ "0.147", "0.552", "0.4575", "0.552", "0.4575", "0.5654", "0.147", "0.5654" ], [ "0.147", "0.5664", "0.4575", "0.5664", "0.4575", "0.58", "0.147", "0.58" ], [ "0.532", "0.5625", "0.881", "0.5625", "0.881", "0.576", "0.532", "0.576" ], [ "0.1453", "0.579", "0.461", "0.58", "0.461", "0.594", "0.1453", "0.5933" ], [ "0.513", "0.5776", "0.879", "0.5776", "0.879", "0.5913", "0.513", "0.5913" ], [ "0.1486", "0.5967", "0.4475", "0.5967", "0.4475", "0.607", "0.1486", "0.607" ], [ "0.511", "0.5933", "0.879", "0.5933", "0.879", "0.607", "0.511", "0.607" ], [ "0.513", "0.61", "0.881", "0.61", "0.881", "0.6235", "0.513", "0.6235" ], [ "0.11584", "0.6226", "0.1511", "0.6226", "0.1511", "0.6343", "0.11584", "0.6343" ], [ "0.1486", "0.6206", "0.2644", "0.6216", "0.2644", "0.6357", "0.1486", "0.6353" ], [ "0.511", "0.626", "0.879", "0.626", "0.879", "0.6396", "0.511", "0.6396" ], [ "0.1184", "0.6465", "0.4844", "0.6465", "0.4844", "0.661", "0.1184", "0.661" ], [ "0.511", "0.641", "0.882", "0.6426", "0.882", "0.6567", "0.511", "0.6562" ], [ "0.1184", "0.662", "0.4836", "0.662", "0.4836", "0.676", "0.1184", "0.676" ], [ "0.51", "0.6577", "0.6323", "0.6567", "0.6323", "0.6714", "0.51", "0.673" ], [ "0.11755", "0.6787", "0.486", "0.6787", "0.486", "0.6924", "0.11755", "0.6924" ], [ "0.531", "0.677", "0.881", "0.677", "0.881", "0.6904", "0.531", "0.6904" ], [ "0.1184", "0.6953", "0.4844", "0.6953", "0.4844", "0.709", "0.1184", "0.709" ], [ "0.511", "0.6914", "0.879", "0.6914", "0.879", "0.705", "0.511", "0.705" ], [ "0.12006", "0.712", "0.4844", "0.712", "0.4844", "0.7256", "0.12006", "0.7256" ], [ "0.511", "0.707", "0.879", "0.707", "0.879", "0.72", "0.511", "0.72" ], [ "0.1184", "0.728", "0.486", "0.728", "0.486", "0.741", "0.1184", "0.741" ], [ "0.511", "0.7246", "0.881", "0.7246", "0.881", "0.7383", "0.511", "0.7383" ], [ "0.11755", "0.742", "0.482", "0.742", "0.482", "0.756", "0.11755", "0.756" ], [ "0.5146", "0.742", "0.594", "0.742", "0.594", "0.7544", "0.5146", "0.7544" ], [ "0.12006", "0.7603", "0.486", "0.7603", "0.486", "0.774", "0.12006", "0.774" ], [ "0.532", "0.7583", "0.882", "0.7583", "0.882", "0.7715", "0.532", "0.7715" ], [ "0.11584", "0.775", "0.486", "0.775", "0.486", "0.788", "0.11584", "0.788" ], [ "0.5146", "0.775", "0.882", "0.775", "0.882", "0.788", "0.5146", "0.788" ], [ "0.1184", "0.792", "0.4878", "0.792", "0.4878", "0.8057", "0.1184", "0.8057" ], [ "0.513", "0.7915", "0.882", "0.7915", "0.882", "0.8047", "0.513", "0.8047" ], [ "0.1184", "0.808", "0.4878", "0.808", "0.4878", "0.822", "0.1184", "0.822" ], [ "0.513", "0.807", "0.882", "0.807", "0.882", "0.821", "0.513", "0.821" ], [ "0.11755", "0.8237", "0.3267", "0.8223", "0.3267", "0.8374", "0.11755", "0.8384" ], [ "0.513", "0.8247", "0.8774", "0.8247", "0.8774", "0.835", "0.513", "0.835" ], [ "0.1368", "0.8413", "0.4844", "0.8413", "0.4844", "0.855", "0.1368", "0.855" ], [ "0.513", "0.839", "0.881", "0.839", "0.881", "0.8525", "0.513", "0.8525" ], [ "0.1184", "0.857", "0.4844", "0.857", "0.4844", "0.8706", "0.1184", "0.8706" ], [ "0.513", "0.855", "0.882", "0.854", "0.882", "0.868", "0.513", "0.869" ], [ "0.12006", "0.8735", "0.4844", "0.8735", "0.4844", "0.887", "0.12006", "0.887" ], [ "0.513", "0.8716", "0.7153", "0.8716", "0.7153", "0.8853", "0.513", "0.8853" ], [ "0.11755", "0.8896", "0.4844", "0.8896", "0.4844", "0.903", "0.11755", "0.903" ], [ "0.531", "0.8896", "0.882", "0.8896", "0.882", "0.903", "0.531", "0.903" ], [ "0.11755", "0.905", "0.4878", "0.905", "0.4878", "0.9185", "0.11755", "0.9185" ], [ "0.511", "0.904", "0.882", "0.905", "0.882", "0.9194", "0.511", "0.9185" ] ]

[ "PRBOOST：用于交互式弱监督学习的基于提示的规则发现和增强", "Yue Yu", "Pranav Shetty", "Georgia Tech", "Rongzhi Zhang", "Georgia Tech", "Georgia Tech", "[email protected]", "[email protected]", "[email protected]", "Le Song", "Chao Zhang", "MBZUAI", "Georgia Tech", "[email protected]", "[email protected]", "Abstract", "2018），但是即使对于专家来说，编写标签规", "则的过程也很繁且耗时。一些工作尝试通过", "Weakly-supervised learning (WSL) has shown", "挖掘标记数据（VarmaandRé，2018）或枚举预", "promising results in addressing label scarcity", "定义的类型来自动发现标记规则。但是，预先", "on many NLP tasks, but manually designing", "提取的规则仅限于频繁模式或预定义类型，这", "acomprehensive,high-qualitylabelingrule", "set is tedious and difficult. We study interac-", "不足以训练准确的模型。其次，大多数现有的", "tive weakly-supervised learning—the problem", "WSL方法都是静态的，并且会受到初始弱监", "of iteratively and automatically discovering", "督 (Ratner et al., 2017; Zhou et al., 2020; Yu et al.,", "novel labeling rules from data to improve the", "2021b; Meng et al., 2020; Zhang et al., 2022) 中", "WSL model. Our proposed model, named PR-", "的噪声的影响。由于标记规则集在模型训练期", "BoosT, achieves this goal via iterative prompt-", "间保持固定，因此初始误差可能会被放大，从", "based rule discovery and model boosting. It", "而导致过度拟合的最终模型。在最近的两部", "uses boosting to identify large-error instances", "作品 (Boecking et al., 2021;Galhotra et al., 2021)", "andthendiscoverscandidaterulesfromthem", "by promptingpre-trained LMs with rule tem", "中探索了交互式规则发现，该作品征求了人类", "plates. The candidate rules are judged by hu", "对候选规则的反馈，以完善规则集。不幸的是", "man experts, and the accepted rules are used", "它们的规则形式仅限于简单的重复结构，例如", "togeneratecomplementaryweaklabelsand", "n-gram（Boeckingetal.,2021），并且巨大的规", "strengthenthecurrentmodel.Experimentsor", "则搜索空间使枚举修剪管道无法扩展到大型数", "fourtasksshow PRBoosToutperformsstate", "据集 (Galhotra et al.,2021)", "of-the-art WSL baselines up to 7.1% , and", "bridges the gaps with fully supervised mod", "由于上述原因，最先进的WSL方法在许多", "els.Our Implementation is available at ht tps :", "NLP任务上仍存在明显差距，因此性能仍低于", "//github.com/rz-zhang/PRBoost:", "完全监督的方法。如最近的一项研究E(Zhang", "etal.,2021)所示，在18个NLP基准测试中的", "1", "Introduction", "15个中，最佳的WSL方法落后于最佳的全监", "弱监督学习（WSL）最近引起了越来越多的", "督方法；就准确性或F1得分而言，平均性能", "关注，以减轻许多NLP任务中的标签稀缺问", "差距为18.84%", "题。在WSL中，训练数据是由从知识库、频", "为了弥合弱监督和完全监督方法之间的差", "繁模式或人类专家等来源获得的弱标记规则生", "距，我们提出了一个迭代规则发现和提升框", "成的。弱标记规则可以与未标记的数据匹配", "架，即用于交互式WSL的PRBOOST。与现有", "以创建大规模的弱标记，从而允许以更低的注", "的WSL和主动学习相比，PRBOOST具有三个", "释成本训练NLP模型。WSL最近在许多任务", "关键设计：", "中取得了可喜的成果，包括文本分类（Awasthi", "首先，我们设计一个规则发现模块，该模块使", "et al., 2020: Mekala and Shang, 2020: Meng et al..", "用规则模板来提示预训练的语言模型（PLM）。", "2020;Yuet al.,2021b），关系提取(Zhou et al.,", "通过将困难的实例和规则模板馈送到PLM中，", "2020）和序列标记（Lison etal.,2020；Safranchik", "该模块通过提示从PLM中提取知识，并生成", "et al., 2020; Li et al., 2021b)", "捕获输入实例的关键语义的候选规则。与基于", "尽管WSL取得了成功，但它受到两个主要", "n-gramE(Boeckingetal.,2021）的先前作品相", "因素的限制：1）标记规则，以及2）静态学", "比，我们基于提示的规则发现更具表现力，适", "习过程。首先，先验地提供一套全面、高质量", "用于任何支持提示的任务", "的标签规则是具有挑战性的。标签规则通常是", "其次，我们设计了一种boosting风格的集成", "人为编写的(Ratner et al.,2017;Hancock et al.,", "策略，以迭代地针对困难的实例并自适应地提" ]

{ "arxiv_id": "2203.09735", "language": "zh", "page_idx": 1 }

[ [ "0.12177", "0.0885", "0.486", "0.0885", "0.486", "0.0986", "0.12177", "0.0986" ], [ "0.511", "0.0873", "0.881", "0.0873", "0.881", "0.10095", "0.511", "0.10095" ], [ "0.12006", "0.1027", "0.4844", "0.1027", "0.4844", "0.1164", "0.12006", "0.1164" ], [ "0.511", "0.1027", "0.879", "0.1027", "0.879", "0.1164", "0.511", "0.1164" ], [ "0.461", "0.10925", "0.4744", "0.10925", "0.4744", "0.1134", "0.461", "0.1134" ], [ "0.1184", "0.11816", "0.4844", "0.11816", "0.4844", "0.1318", "0.1184", "0.1318" ], [ "0.511", "0.11816", "0.881", "0.11816", "0.881", "0.133", "0.511", "0.133" ], [ "0.1184", "0.1354", "0.4844", "0.1354", "0.4844", "0.1484", "0.1184", "0.1484" ], [ "0.513", "0.1354", "0.881", "0.1354", "0.881", "0.1484", "0.513", "0.1484" ], [ "0.12006", "0.1526", "0.4634", "0.1526", "0.4634", "0.1633", "0.12006", "0.1633" ], [ "0.513", "0.1509", "0.881", "0.1509", "0.881", "0.1644", "0.513", "0.1644" ], [ "0.1351", "0.1675", "0.486", "0.1675", "0.486", "0.1812", "0.1351", "0.1812" ], [ "0.513", "0.1675", "0.881", "0.1675", "0.881", "0.1812", "0.513", "0.1812" ], [ "0.12006", "0.1859", "0.4844", "0.1859", "0.4844", "0.1965", "0.12006", "0.1965" ], [ "0.513", "0.1818", "0.879", "0.1818", "0.879", "0.1953", "0.513", "0.1953" ], [ "0.12006", "0.2007", "0.4844", "0.2007", "0.4844", "0.2144", "0.12006", "0.2144" ], [ "0.513", "0.2007", "0.8774", "0.2007", "0.8774", "0.2114", "0.513", "0.2114" ], [ "0.12006", "0.2162", "0.4844", "0.2162", "0.4844", "0.2299", "0.12006", "0.2299" ], [ "0.513", "0.2156", "0.881", "0.2156", "0.881", "0.2286", "0.513", "0.2286" ], [ "0.1184", "0.2328", "0.4844", "0.2328", "0.4844", "0.2465", "0.1184", "0.2465" ], [ "0.511", "0.2322", "0.879", "0.2322", "0.879", "0.2452", "0.511", "0.2452" ], [ "0.1184", "0.2488", "0.4895", "0.2488", "0.4895", "0.2625", "0.1184", "0.2625" ], [ "0.511", "0.2465", "0.882", "0.2477", "0.882", "0.2625", "0.511", "0.2612" ], [ "0.12177", "0.2632", "0.486", "0.2632", "0.486", "0.2766", "0.12177", "0.2766" ], [ "0.511", "0.2625", "0.881", "0.2632", "0.881", "0.2778", "0.511", "0.2766" ], [ "0.12006", "0.282", "0.486", "0.282", "0.486", "0.2922", "0.12006", "0.2922" ], [ "0.513", "0.2798", "0.881", "0.2798", "0.881", "0.2935", "0.513", "0.2935" ], [ "0.1184", "0.2964", "0.4895", "0.2964", "0.4895", "0.31", "0.1184", "0.31" ], [ "0.513", "0.2957", "0.881", "0.2957", "0.881", "0.3088", "0.513", "0.3088" ], [ "0.12006", "0.3142", "0.2275", "0.3142", "0.2275", "0.326", "0.12006", "0.326" ], [ "0.5146", "0.3113", "0.881", "0.3113", "0.881", "0.3247", "0.5146", "0.3247" ], [ "0.1368", "0.33", "0.4878", "0.33", "0.4878", "0.3433", "0.1368", "0.3433" ], [ "0.511", "0.3267", "0.876", "0.3267", "0.876", "0.3403", "0.511", "0.3403" ], [ "0.1184", "0.3457", "0.486", "0.3457", "0.486", "0.3594", "0.1184", "0.3594" ], [ "0.511", "0.3433", "0.881", "0.3433", "0.881", "0.357", "0.511", "0.357" ], [ "0.1184", "0.361", "0.4878", "0.361", "0.4878", "0.3748", "0.1184", "0.3748" ], [ "0.511", "0.3599", "0.6978", "0.3599", "0.6978", "0.3735", "0.511", "0.3735" ], [ "0.1184", "0.379", "0.4844", "0.379", "0.4844", "0.3896", "0.1184", "0.3896" ], [ "0.51", "0.3772", "0.881", "0.376", "0.881", "0.3901", "0.51", "0.3914" ], [ "0.1184", "0.3943", "0.4895", "0.3943", "0.4895", "0.408", "0.1184", "0.408" ], [ "0.511", "0.3926", "0.879", "0.3926", "0.879", "0.4062", "0.511", "0.4062" ], [ "0.1184", "0.4104", "0.4878", "0.4104", "0.4878", "0.4233", "0.1184", "0.4233" ], [ "0.513", "0.4104", "0.881", "0.4104", "0.881", "0.4233", "0.513", "0.4233" ], [ "0.1184", "0.427", "0.4878", "0.427", "0.4878", "0.4407", "0.1184", "0.4407" ], [ "0.51", "0.4258", "0.879", "0.4258", "0.879", "0.4395", "0.51", "0.4395" ], [ "0.12006", "0.4448", "0.486", "0.4448", "0.486", "0.4548", "0.12006", "0.4548" ], [ "0.513", "0.4424", "0.881", "0.4424", "0.881", "0.456", "0.513", "0.456" ], [ "0.12006", "0.4602", "0.2284", "0.4602", "0.2284", "0.4714", "0.12006", "0.4714" ], [ "0.5146", "0.4578", "0.879", "0.4578", "0.879", "0.4714", "0.5146", "0.4714" ], [ "0.1351", "0.4756", "0.4878", "0.4744", "0.4878", "0.4893", "0.1351", "0.4905" ], [ "0.513", "0.4739", "0.879", "0.4739", "0.879", "0.4868", "0.513", "0.4868" ], [ "0.12006", "0.4934", "0.4844", "0.4934", "0.4844", "0.5044", "0.12006", "0.5044" ], [ "0.51", "0.4905", "0.879", "0.4905", "0.879", "0.5044", "0.51", "0.5044" ], [ "0.12006", "0.5083", "0.486", "0.5083", "0.486", "0.5215", "0.12006", "0.5215" ], [ "0.511", "0.505", "0.881", "0.506", "0.881", "0.5195", "0.511", "0.5186" ], [ "0.12006", "0.5254", "0.486", "0.5254", "0.486", "0.536", "0.12006", "0.536" ], [ "0.513", "0.5225", "0.881", "0.5225", "0.881", "0.536", "0.513", "0.536" ], [ "0.12006", "0.5405", "0.4844", "0.5405", "0.4844", "0.5503", "0.12006", "0.5503" ], [ "0.513", "0.538", "0.881", "0.538", "0.881", "0.552", "0.513", "0.552" ], [ "0.1184", "0.5557", "0.4895", "0.5557", "0.4895", "0.5693", "0.1184", "0.5693" ], [ "0.513", "0.554", "0.881", "0.554", "0.881", "0.568", "0.513", "0.568" ], [ "0.1184", "0.5723", "0.4492", "0.5723", "0.4492", "0.586", "0.1184", "0.586" ], [ "0.5146", "0.571", "0.879", "0.571", "0.879", "0.5845", "0.5146", "0.5845" ], [ "0.5146", "0.5874", "0.881", "0.5874", "0.881", "0.601", "0.5146", "0.601" ], [ "0.1142", "0.6025", "0.2747", "0.6025", "0.2747", "0.616", "0.1142", "0.616" ], [ "0.513", "0.6025", "0.559", "0.6025", "0.559", "0.6177", "0.513", "0.6177" ], [ "0.11755", "0.63", "0.4836", "0.63", "0.4836", "0.6436", "0.11755", "0.6436" ], [ "0.5073", "0.6343", "0.6665", "0.6343", "0.6665", "0.648", "0.5073", "0.648" ], [ "0.1184", "0.6465", "0.486", "0.6465", "0.486", "0.661", "0.1184", "0.661" ], [ "0.12006", "0.662", "0.4902", "0.662", "0.4902", "0.676", "0.12006", "0.676" ], [ "0.511", "0.661", "0.879", "0.662", "0.879", "0.677", "0.511", "0.676" ], [ "0.1184", "0.6787", "0.486", "0.6787", "0.486", "0.6924", "0.1184", "0.6924" ], [ "0.513", "0.6787", "0.879", "0.6787", "0.879", "0.6924", "0.513", "0.6924" ], [ "0.11755", "0.6934", "0.4836", "0.6934", "0.4836", "0.707", "0.11755", "0.707" ], [ "0.5146", "0.697", "0.879", "0.697", "0.879", "0.707", "0.5146", "0.707" ], [ "0.1184", "0.7114", "0.4878", "0.7114", "0.4878", "0.7246", "0.1184", "0.7246" ], [ "0.513", "0.7114", "0.882", "0.7114", "0.882", "0.7246", "0.513", "0.7246" ], [ "0.1184", "0.727", "0.4878", "0.727", "0.4878", "0.7407", "0.1184", "0.7407" ], [ "0.513", "0.728", "0.881", "0.728", "0.881", "0.741", "0.513", "0.741" ], [ "0.11755", "0.7437", "0.4878", "0.7437", "0.4878", "0.7573", "0.11755", "0.7573" ], [ "0.511", "0.742", "0.881", "0.7437", "0.881", "0.7573", "0.511", "0.756" ], [ "0.1184", "0.7603", "0.4878", "0.7603", "0.4878", "0.774", "0.1184", "0.774" ], [ "0.513", "0.759", "0.881", "0.759", "0.881", "0.7725", "0.513", "0.7725" ], [ "0.1184", "0.775", "0.486", "0.775", "0.486", "0.788", "0.1184", "0.788" ], [ "0.5146", "0.7754", "0.879", "0.7754", "0.879", "0.786", "0.5146", "0.786" ], [ "0.12006", "0.7935", "0.4844", "0.7935", "0.4844", "0.804", "0.12006", "0.804" ], [ "0.513", "0.7915", "0.723", "0.7915", "0.723", "0.8047", "0.513", "0.8047" ], [ "0.1184", "0.808", "0.4844", "0.808", "0.4844", "0.822", "0.1184", "0.822" ], [ "0.529", "0.8086", "0.881", "0.8086", "0.881", "0.8223", "0.529", "0.8223" ], [ "0.1184", "0.826", "0.4844", "0.826", "0.4844", "0.836", "0.1184", "0.836" ], [ "0.513", "0.8237", "0.881", "0.8247", "0.881", "0.839", "0.513", "0.8384" ], [ "0.1184", "0.839", "0.4895", "0.839", "0.4895", "0.8525", "0.1184", "0.8525" ], [ "0.513", "0.8413", "0.881", "0.8413", "0.881", "0.855", "0.513", "0.855" ], [ "0.1184", "0.8555", "0.4878", "0.8555", "0.4878", "0.869", "0.1184", "0.869" ], [ "0.511", "0.857", "0.882", "0.857", "0.882", "0.8706", "0.511", "0.8706" ], [ "0.12006", "0.8735", "0.266", "0.8735", "0.266", "0.8843", "0.12006", "0.8843" ], [ "0.511", "0.8735", "0.879", "0.8735", "0.879", "0.887", "0.511", "0.887" ], [ "0.1184", "0.8896", "0.4895", "0.8896", "0.4895", "0.903", "0.1184", "0.903" ], [ "0.51", "0.8896", "0.881", "0.8896", "0.881", "0.903", "0.51", "0.903" ], [ "0.12006", "0.9067", "0.4878", "0.9067", "0.4878", "0.9175", "0.12006", "0.9175" ], [ "0.4902", "0.9077", "0.5054", "0.909", "0.5054", "0.913", "0.4902", "0.912" ], [ "0.511", "0.9067", "0.876", "0.9067", "0.876", "0.9175", "0.511", "0.9175" ] ]

[ "出新规则。在每次迭代中，我们通过提升错误", "是，关键的区别在于AL基于各种查询策略来", "重新权衡数据，以强制规则发现模块专注于较", "标记实例E(Holub et al.,2008;Shen etal.,2017", "", "大错误的实例。这避免了枚举所有可能的规则", "Zhang et al., 2020; Ein-Dor et al., 2020; Margatina", "和对新规则实施后过滤，而是直接针对大错误", "etal.，2021；Yuetal.,2021a），而我们的方法不", "实例的规则发现，为当前模型提供补充信息", "注释单个实例，而是使用注释的规则来匹配", "第三，我们从战略上征求人类反馈来评估候", "未标记的数据。这使得我们的方法在利用人类", "选规则。人类被要求判断一个候选规则是否应", "反馈创建大规模标记数据时更具标签效率。据", "该被接受或弃权。然后将接受的高质量规则用", "我们所知，只有少数作品在我们的问题中研究", "于生成新的弱标签，这些弱标签被馈送到提升", "了交互式WSL(Boecking et al.,202l;Galhotra", "的模型训练中。由于提示生成的规则是高度可", "et al., 2021; Choi et al., 2021; Hsieh et al., 2022) ", "解释的，规则评估对于人类专家来说只是一个", "但是，他们要么使用简单的基于n-gram的规", "二元选择任务，因此毫不费力。与对单个实例", "则(Boecking et al.,2021;Hsieh et al.,2022)无法", "进行注释的传统主动学习方法不同，这种规则", "捕获句子级语义，要么遭受与上下文无关的语", "级注释的标签效率更高，因为注释的规则可以", "法规则（Galhotraetal.，2021）的巨大搜索空间。", "匹配大量实例", "与这些工作不同，我们的方法使用基于提示的", "我们在四个任务上将我们的方法与有监督，", "灵活规则表示，并且还使用boosting进行有针", "弱监督和交互式学习基线进行了比较：关系提", "对性的规则发现，以避免枚举所有可能的规则", "取，本体分类，主题分类和化学-蛋白质相互作", "并对新规则执行后过滤", "用预测。结果表明：1）我们的方法优干最先进", "语言模型提示我们的工作还与PLM的基于提", "的弱监督基线多达7.1%；2）在相同预算下，", "示的学习有关，该学习将原始任务转换为完形", "与实例级注释相比，规则级注释有助于模型实", "填空式任务，并利用PLM来填充缺少的信息E", "现更高的模型性能；3）机器发现和人工评估", "(Brownetal.2020Liuetal.,2021a)。提示已在", "的规则具有高质量，在每次选代中一致地细化", "各种任务中进行了探索，包括文本分类E(Hu", "弱标签和模型", "et al., 2021; Han et al., 2021; Schick and Schutze", "我们的主要贡献是：（1）用于交互式WSL", "2021a,b），信息提取ELesteretal.2021;Chen", "的基于提示的规则发现框架，该框架在捕获规", "etal.,2021)和文本生成E(Dou etal.,2021Li", "则生成中的微妙语义的同时提供了灵活的规则", "andLiang，2021）。最近的工作集中在生成更", "表示；（2）一种选代的增强策略，用于从硬实", "好的提示模板或学习隐式提示嵌人（Gaoetal.,", "例中发现新颖的规则，并通过一组互补的弱模", "2021；Liuetal.，2021b,c）。然而，这些工作都", "型来增强模型；（3）用于规则注释的可解释且", "没有研究提示生成弱标签。我们的工作与它们", "易于注释的交互式过程；（4）全面的实验", "正交，因为我们的目标不是优化原始任务的提", "示，而是使用提示和PLM作为规则发现的知", "2Related Work", "识源。", "弱监督学习WSL最近在各种NLP任务中引", "3Preliminaries", "起了广泛关注。尽管它们在各种任务上表现", "良好，但手动设计规则可能会很耗时。此外，", "问题公式弱监督学习（WSL）通过在未标记的", "初始规则的噪声和不完整性可能会在模型训", "实例Du上应用标记规则，为模型训练创建弱", "练（Zhangetal,2021）中传播。一些作品试图", "标记。给定一个未标记的实例&EDu：一个", "通过从数据中发现规则来减少人工设计标签", "标记规则r(）将α映射到一个扩展的标签空", "规则的工作量。例如，SnubaE(Varma and Ré,", "间：r(α)→yEo}。这里是任务的", "2018）基于具有预定义规则类型的小型标记", "原始标签集，而0是一个特殊的标签，指示&", "数据集生成启发式算法；TALLORE(Lietal.,", "是r不可匹配的。给定一组R标记规则，我们", "2021a）和GLaRAE(Zhaoetal.,2021)基于词汇", "可以在未标记的实例上应用R中的每个规则，", "信息研究NER问题的规则扩展，然后基于手", "以创建弱标记的数据集D", "动调整的阈值选择规则。但是，这些方法以静", "但是，初始的弱标签D可能具有很高的噪", "态方式发现规则，并被限制为特定于任务的规", "点且不完整，从而阻碍了WSL的性能。我们", "则类型。相反，我们的框架从整个未标记的数", "由此研究了交互式WSL的问题：如何自动发", "据集中迭代地发现规则，这可以细化规则集并", "现更多高质量的标签规则来增强WSL的性能？", "实时扩大其多样性", "除了Du和D之外，我们还假设可以访问一小", "交互式学习我们的工作与主动学习（AL）有关，", "组干净的标签Di（/Dil《DuD），任务是迭代", "因为两者都在学习过程中涉及人类注释者。但", "", "地找到一组新的规则来进行模型改进。在每个" ]

{ "arxiv_id": "2203.09735", "language": "zh", "page_idx": 2 }

[ [ "0.1184", "0.0873", "0.486", "0.0873", "0.486", "0.0974", "0.1184", "0.0974" ], [ "0.509", "0.0873", "0.881", "0.0873", "0.881", "0.10095", "0.509", "0.10095" ], [ "0.11755", "0.1027", "0.486", "0.1027", "0.486", "0.1164", "0.11755", "0.1164" ], [ "0.532", "0.1027", "0.881", "0.1027", "0.881", "0.1164", "0.532", "0.1164" ], [ "0.1184", "0.11816", "0.4878", "0.11816", "0.4878", "0.133", "0.1184", "0.133" ], [ "0.5356", "0.1194", "0.881", "0.1194", "0.881", "0.133", "0.5356", "0.133" ], [ "0.1184", "0.136", "0.4844", "0.136", "0.4844", "0.1497", "0.1184", "0.1497" ], [ "0.534", "0.1354", "0.882", "0.1354", "0.882", "0.1497", "0.534", "0.1497" ], [ "0.1184", "0.1526", "0.4844", "0.1526", "0.4844", "0.1663", "0.1184", "0.1663" ], [ "0.534", "0.1509", "0.721", "0.1509", "0.721", "0.1644", "0.534", "0.1644" ], [ "0.1184", "0.1687", "0.3535", "0.1687", "0.3535", "0.1818", "0.1184", "0.1818" ], [ "0.513", "0.1818", "0.5557", "0.1818", "0.5557", "0.1936", "0.513", "0.1936" ], [ "0.555", "0.1812", "0.777", "0.1812", "0.777", "0.1948", "0.555", "0.1948" ], [ "0.553", "0.1853", "0.559", "0.1853", "0.559", "0.1907", "0.553", "0.1907" ], [ "0.1184", "0.1853", "0.486", "0.1853", "0.486", "0.199", "0.1184", "0.199" ], [ "0.1184", "0.2007", "0.4844", "0.2007", "0.4844", "0.2144", "0.1184", "0.2144" ], [ "0.513", "0.2037", "0.881", "0.2037", "0.881", "0.2173", "0.513", "0.2173" ], [ "0.12177", "0.2185", "0.4844", "0.2185", "0.4844", "0.2286", "0.12177", "0.2286" ], [ "0.5146", "0.2214", "0.879", "0.2214", "0.879", "0.2322", "0.5146", "0.2322" ], [ "0.11755", "0.2328", "0.4844", "0.2328", "0.4844", "0.2465", "0.11755", "0.2465" ], [ "0.513", "0.2363", "0.881", "0.2363", "0.881", "0.2494", "0.513", "0.2494" ], [ "0.1184", "0.2494", "0.4844", "0.2494", "0.4844", "0.2632", "0.1184", "0.2632" ], [ "0.513", "0.253", "0.881", "0.253", "0.881", "0.2666", "0.513", "0.2666" ], [ "0.1184", "0.2654", "0.4878", "0.2654", "0.4878", "0.279", "0.1184", "0.279" ], [ "0.513", "0.2683", "0.881", "0.2683", "0.881", "0.282", "0.513", "0.282" ], [ "0.1184", "0.282", "0.4844", "0.282", "0.4844", "0.2922", "0.1184", "0.2922" ], [ "0.513", "0.2852", "0.881", "0.2852", "0.881", "0.2986", "0.513", "0.2986" ], [ "0.11755", "0.2976", "0.4895", "0.2976", "0.4895", "0.3113", "0.11755", "0.3113" ], [ "0.513", "0.3005", "0.881", "0.3005", "0.881", "0.3142", "0.513", "0.3142" ], [ "0.1184", "0.3135", "0.4878", "0.3135", "0.4878", "0.3267", "0.1184", "0.3267" ], [ "0.511", "0.3154", "0.879", "0.3154", "0.879", "0.3289", "0.511", "0.3289" ], [ "0.11584", "0.328", "0.4878", "0.3289", "0.4878", "0.3433", "0.11584", "0.3425" ], [ "0.511", "0.3308", "0.881", "0.332", "0.881", "0.3467", "0.511", "0.3457" ], [ "0.1184", "0.3467", "0.2091", "0.3467", "0.2091", "0.3582", "0.1184", "0.3582" ], [ "0.5146", "0.3499", "0.621", "0.3499", "0.621", "0.361", "0.5146", "0.361" ], [ "0.1368", "0.361", "0.4878", "0.361", "0.4878", "0.3748", "0.1368", "0.3748" ], [ "0.5283", "0.3652", "0.8774", "0.3665", "0.8774", "0.3801", "0.5283", "0.379" ], [ "0.1184", "0.379", "0.4844", "0.379", "0.4844", "0.3896", "0.1184", "0.3896" ], [ "0.509", "0.3813", "0.879", "0.3813", "0.879", "0.3943", "0.509", "0.3943" ], [ "0.1184", "0.3943", "0.4878", "0.3943", "0.4878", "0.408", "0.1184", "0.408" ], [ "0.513", "0.398", "0.881", "0.398", "0.881", "0.411", "0.513", "0.411" ], [ "0.11755", "0.4104", "0.4844", "0.4104", "0.4844", "0.4233", "0.11755", "0.4233" ], [ "0.513", "0.4146", "0.881", "0.4146", "0.881", "0.4275", "0.513", "0.4275" ], [ "0.11755", "0.427", "0.4844", "0.427", "0.4844", "0.4407", "0.11755", "0.4407" ], [ "0.513", "0.4312", "0.623", "0.4312", "0.623", "0.4424", "0.513", "0.4424" ], [ "0.11755", "0.4424", "0.4844", "0.4424", "0.4844", "0.456", "0.11755", "0.456" ], [ "0.517", "0.4478", "0.8833", "0.4465", "0.8833", "0.4685", "0.517", "0.4697" ], [ "0.12006", "0.4602", "0.482", "0.4602", "0.482", "0.4702", "0.12006", "0.4702" ], [ "0.12006", "0.4744", "0.4844", "0.4744", "0.4844", "0.488", "0.12006", "0.488" ], [ "0.511", "0.4744", "0.642", "0.4744", "0.642", "0.488", "0.511", "0.488" ], [ "0.11584", "0.4905", "0.3123", "0.4905", "0.3123", "0.5044", "0.11584", "0.5044" ], [ "0.529", "0.4893", "0.881", "0.4905", "0.881", "0.505", "0.529", "0.5044" ], [ "0.5146", "0.5073", "0.882", "0.5073", "0.882", "0.521", "0.5146", "0.521" ], [ "0.1142", "0.517", "0.2644", "0.5195", "0.2644", "0.5337", "0.1142", "0.532" ], [ "0.513", "0.524", "0.7983", "0.524", "0.7983", "0.5376", "0.513", "0.5376" ], [ "0.642", "0.5337", "0.827", "0.5444", "0.8237", "0.564", "0.6396", "0.554" ], [ "0.11755", "0.542", "0.4836", "0.542", "0.4836", "0.557", "0.11755", "0.557" ], [ "0.571", "0.5444", "0.642", "0.542", "0.644", "0.558", "0.5728", "0.56" ], [ "0.858", "0.5474", "0.881", "0.5474", "0.881", "0.558", "0.858", "0.558" ], [ "0.659", "0.5557", "0.692", "0.5557", "0.692", "0.568", "0.659", "0.568" ], [ "0.12006", "0.5625", "0.4844", "0.5625", "0.4844", "0.5723", "0.12006", "0.5723" ], [ "0.5146", "0.5684", "0.6245", "0.5684", "0.6245", "0.583", "0.5146", "0.583" ], [ "0.1184", "0.5767", "0.486", "0.5767", "0.486", "0.5903", "0.1184", "0.5903" ], [ "0.594", "0.5845", "0.6206", "0.5845", "0.6206", "0.597", "0.594", "0.597" ], [ "0.781", "0.579", "0.8096", "0.5845", "0.8037", "0.601", "0.775", "0.5957" ], [ "0.11755", "0.5913", "0.4844", "0.593", "0.4844", "0.607", "0.11755", "0.6055" ], [ "0.54", "0.5986", "0.596", "0.601", "0.594", "0.615", "0.538", "0.6133" ], [ "0.5884", "0.5967", "0.885", "0.5967", "0.885", "0.618", "0.5884", "0.618" ], [ "0.12177", "0.6094", "0.4844", "0.6094", "0.4844", "0.6226", "0.12177", "0.6226" ], [ "0.592", "0.6123", "0.621", "0.6123", "0.621", "0.629", "0.592", "0.629" ], [ "0.7793", "0.6133", "0.804", "0.6133", "0.804", "0.6274", "0.7793", "0.6274" ], [ "0.1184", "0.626", "0.2747", "0.626", "0.2747", "0.6396", "0.1184", "0.6396" ], [ "0.511", "0.6274", "0.55", "0.6274", "0.55", "0.6436", "0.511", "0.6436" ], [ "0.532", "0.6455", "0.881", "0.6455", "0.881", "0.659", "0.532", "0.659" ], [ "0.11584", "0.653", "0.486", "0.653", "0.486", "0.6665", "0.11584", "0.6665" ], [ "0.511", "0.66", "0.879", "0.66", "0.879", "0.6733", "0.511", "0.6733" ], [ "0.1394", "0.6685", "0.4844", "0.6675", "0.4844", "0.6816", "0.1394", "0.683" ], [ "0.513", "0.6777", "0.879", "0.6777", "0.879", "0.6914", "0.513", "0.6914" ], [ "0.1394", "0.685", "0.486", "0.685", "0.486", "0.6987", "0.1394", "0.6987" ], [ "0.1394", "0.6997", "0.4878", "0.6997", "0.4878", "0.7134", "0.1394", "0.7134" ], [ "0.5146", "0.6953", "0.8774", "0.6953", "0.8774", "0.706", "0.5146", "0.706" ], [ "0.513", "0.709", "0.879", "0.709", "0.879", "0.7227", "0.513", "0.7227" ], [ "0.1394", "0.7173", "0.486", "0.7173", "0.486", "0.731", "0.1394", "0.731" ], [ "0.513", "0.727", "0.8774", "0.727", "0.8774", "0.7407", "0.513", "0.7407" ], [ "0.1377", "0.733", "0.4878", "0.7324", "0.4878", "0.7466", "0.1377", "0.7476" ], [ "0.1377", "0.7466", "0.3645", "0.749", "0.3645", "0.764", "0.1377", "0.7627" ], [ "0.5146", "0.742", "0.881", "0.742", "0.881", "0.7573", "0.5146", "0.7573" ], [ "0.513", "0.7573", "0.879", "0.7573", "0.879", "0.771", "0.513", "0.771" ], [ "0.11755", "0.7754", "0.4895", "0.7754", "0.4895", "0.789", "0.11755", "0.789" ], [ "0.513", "0.775", "0.879", "0.775", "0.879", "0.788", "0.513", "0.788" ], [ "0.1394", "0.792", "0.4895", "0.792", "0.4895", "0.8057", "0.1394", "0.8057" ], [ "0.513", "0.7915", "0.881", "0.7915", "0.881", "0.8047", "0.513", "0.8047" ], [ "0.1394", "0.8086", "0.4878", "0.8086", "0.4878", "0.8223", "0.1394", "0.8223" ], [ "0.511", "0.808", "0.8774", "0.808", "0.8774", "0.822", "0.511", "0.822" ], [ "0.1377", "0.8237", "0.4895", "0.826", "0.4895", "0.839", "0.1377", "0.8374" ], [ "0.513", "0.8237", "0.881", "0.8237", "0.881", "0.8374", "0.513", "0.8374" ], [ "0.1394", "0.8413", "0.486", "0.8413", "0.486", "0.855", "0.1394", "0.855" ], [ "0.511", "0.839", "0.727", "0.839", "0.727", "0.8525", "0.511", "0.8525" ], [ "0.1394", "0.857", "0.486", "0.857", "0.486", "0.8706", "0.1394", "0.8706" ], [ "0.529", "0.857", "0.879", "0.857", "0.879", "0.8706", "0.529", "0.8706" ], [ "0.1394", "0.8735", "0.4878", "0.8735", "0.4878", "0.887", "0.1394", "0.887" ], [ "0.511", "0.8726", "0.879", "0.8726", "0.879", "0.8857", "0.511", "0.8857" ], [ "0.1368", "0.888", "0.486", "0.8896", "0.486", "0.904", "0.1368", "0.903" ], [ "0.511", "0.8896", "0.881", "0.8896", "0.881", "0.903", "0.511", "0.903" ], [ "0.1394", "0.9067", "0.1738", "0.9067", "0.1738", "0.9194", "0.1394", "0.9194" ], [ "0.513", "0.9067", "0.879", "0.9067", "0.879", "0.9175", "0.513", "0.9175" ] ]

[ "迭代中，我们假设一个固定的规则注释预算", "3. Weakly supervised model training and ensem-", "B，i。e。，最多可以向人类专家提出B候选规", "ble . We train a new weak model mt+i on the", "则Rt=r-1，以决定是否接受每条规则。", "updated weakly labeled dataset Dr . Then we", "然后使用接受的规则R+创建新的弱标记实例", "self-train the weak model mt+i and integrate it", "Dt。从DtUD,中，可以训练模型mt:X→y", "into the ensemble model.", "以提高当前WSL模型的性能，", "4.1", "Candidate Rule Generation", "", "规则表示已在WSL中为NLP任务提出了多", "个规则表示。例如，基于关键字的规则被广泛", "大错误实例上的目标规则建议我们设计了一种", "用于将某些关键字映射到其高度相关的标签", "升压式（Hastieetal..2009）策略：用于生成基", "(Boecking et al., 2021; Meng et al., 2020; Mekala", "于提示的候选规则。此策略迭代检查当前模型", "andShang，2020;Liangetal.,2020)。正则表达式", "mt较弱的特征机制，并从此类机制中提出候", "是另一种常见的规则格式，它匹配具有预定义", "选规则。我们使用小标签数据集Di来识别硬", "表面模式（Awasthietal..2020:Yuetal..2021b:", "实例，即。e。，其中模型在迭代学习过程中往", "Zhouetal.,2020）的实例。逻辑规则（Huetal.，", "往会出现累积错误。发现的规则可以补充当前", "2016;Lietal.,2021a)对原子规则执行逻辑运算", "规则集R并细化弱标签，因此在细化的弱标", "（例如连词^和否定一），因此可以捕获更高阶", "签数据上训练的下一个模型mt+1可以在弱状", "的合成模式", "态下表现更好", "我们采用基于提示的规则表示（第E4.1节），", "我们将D,中实例的权重初始化为Wi=", "它可以灵活地包含任何现有的规则表示。我们", "1/Dl,i=1,2，··，Dl。在迭代模型学习过程", "基于提示的规则依赖于目标任务的规则模板", "中，每个w；都会更新为模型在实例上的加权损", "T），其中包含一个【MASK】令牌，由PLM", "失aiEDu。具体来说，在选代tE[l，，n}", "M和一个未标记的实例&填充。从规则模板", "中，我们通过", "T中，每个候选规则都可以由r=g(M,T,α)", "Wi wi eal(yitm(ci),i= 1, 2,..,|Dil. (1)", "自动导出。这种基于提示的规则表示形式非常", "灵活，可以应用于任何支持提示的NLP任务", "对样本进行加权", "（请参见表E1中的示例）", "在公式E1中，αt是模型mt的权重，它将", "用于检测硬实例和模型集成（第E4.3节）。我", "4Methodology", "们根据Di上的模型错误率计算αt：", "i-errt +log(K -1),", "概述PRBOOST是交互式WSL的选代方法。", "Qt = log", "(2)", "errt", "在每次送代中，它从大错误实例中提出候选", "，其中errt由", "规则，征求人类对候选规则的反馈，生成弱标", "[D;]", "IDil", "签，并训练新的弱模型进行集成。图E1显示", "errt =", "Zw;I(yi+mt (ai) /Zwi.(3)", "了PRBOOST的一次选代中的过程，该过程依", "1=1", "i=1", "赖于三个关键组件：", "给出", "从直觉上讲，如果模型集合在ci上始终出", "1. Candidate rule generation.This component", "错，则样本i会收到更大的权重W；（公式F", "proposes candidate rules to be evaluated by", "1）。较大的错误通常是由覆盖率差（未标记的", "human annotators. Using the small labeled", "dataset Di , it measures the weakness of the cur-", "实例与很少或没有规则匹配）或局部特征机制", "中的主导噪声（规则匹配的标签是错误的）引", "rent model by identifying large-error instances", "起的。因此，权重可以引导规则生成器将top", "on Di , and proposes rules based on these in-", "stances using PLM prompting.", "n大错误实例Xe=【ae;]\"-i作为目标。通过", "从这些实例中提出规则，我们旨在发现可以最", "2. Rule annotation and weak label creation. This", "有效地补充当前规则集和模型集成的新颖规则", "component collects human feedback to im-", "基于提示的规则建议对于诸如关系提取和文", "prove the weak supervision quality. It takes as", "本分类之类的各种NLP任务，我们可以利用", "input the candidate rules proposed by the pre-", "提示来构造信息丰富的规则模板，这自然会导", "vious component, and asks humans to select", "致WSL的表达性标记规则", "the high-quality ones. Then the human-selected", "受此启发，我们设计了一个基于PLM提示", "rules Rt are used to generate weak labels for", "的规则建议模块。我们在表1中提供了基于提", "the unlabeled instances Du in a soft-matching", "示的规则的具体示例。输入实例来自在干净数", "way.", "据集Di上标识的大错误实例。对于每个任务" ]

{ "arxiv_id": "2203.09735", "language": "zh", "page_idx": 3 }

[ [ "0.1855", "0.0873", "0.2627", "0.0873", "0.2627", "0.0974", "0.1855", "0.0974" ], [ "0.2737", "0.0861", "0.367", "0.0873", "0.367", "0.0986", "0.2737", "0.0974" ], [ "0.629", "0.0873", "0.81", "0.0873", "0.81", "0.10095", "0.629", "0.10095" ], [ "0.3955", "0.0974", "0.5474", "0.0998", "0.5474", "0.114", "0.3955", "0.11224" ], [ "0.2167", "0.1057", "0.2334", "0.1057", "0.2334", "0.1134", "0.2167", "0.1134" ], [ "0.409", "0.11755", "0.8037", "0.11755", "0.8037", "0.1277", "0.409", "0.1277" ], [ "0.3997", "0.1395", "0.5146", "0.1395", "0.5146", "0.152", "0.3997", "0.152" ], [ "0.4072", "0.155", "0.7886", "0.155", "0.7886", "0.1687", "0.4072", "0.1687" ], [ "0.2988", "0.1644", "0.367", "0.1675", "0.3652", "0.1787", "0.297", "0.1758" ], [ "0.6323", "0.2031", "0.8096", "0.2031", "0.8096", "0.2162", "0.6323", "0.2162" ], [ "0.3955", "0.2227", "0.4492", "0.2227", "0.4492", "0.234", "0.3955", "0.234" ], [ "0.5356", "0.2214", "0.7715", "0.2214", "0.7715", "0.2351", "0.5356", "0.2351" ], [ "0.4626", "0.228", "0.471", "0.228", "0.471", "0.2328", "0.4626", "0.2328" ], [ "0.3972", "0.2351", "0.4417", "0.2351", "0.4417", "0.2465", "0.3972", "0.2465" ], [ "0.4626", "0.2369", "0.4702", "0.2369", "0.4702", "0.2446", "0.4626", "0.2446" ], [ "0.3972", "0.2465", "0.4358", "0.2465", "0.4358", "0.2583", "0.3972", "0.2583" ], [ "0.461", "0.2477", "0.4744", "0.2477", "0.4744", "0.2583", "0.461", "0.2583" ], [ "0.5356", "0.2452", "0.763", "0.2452", "0.763", "0.2588", "0.5356", "0.2588" ], [ "0.3376", "0.2612", "0.4534", "0.2632", "0.4534", "0.279", "0.3376", "0.2766" ], [ "0.586", "0.2632", "0.735", "0.2632", "0.735", "0.2766", "0.586", "0.2766" ], [ "0.532", "0.3005", "0.817", "0.3005", "0.817", "0.3142", "0.532", "0.3142" ], [ "0.1814", "0.2654", "0.2896", "0.2654", "0.2896", "0.2756", "0.1814", "0.2756" ], [ "0.69", "0.3184", "0.7583", "0.3218", "0.7573", "0.3333", "0.6885", "0.33" ], [ "0.4995", "0.3247", "0.567", "0.328", "0.565", "0.3425", "0.4978", "0.339" ], [ "0.617", "0.3374", "0.6245", "0.3374", "0.6245", "0.3403", "0.617", "0.3403" ], [ "0.635", "0.332", "0.6484", "0.332", "0.6484", "0.3362", "0.635", "0.3362" ], [ "0.4148", "0.3467", "0.4224", "0.3467", "0.4224", "0.3499", "0.4148", "0.3499" ], [ "0.6885", "0.3445", "0.784", "0.3445", "0.784", "0.3557", "0.6885", "0.3557" ], [ "0.1914", "0.3706", "0.3105", "0.3718", "0.3105", "0.386", "0.1914", "0.3855" ], [ "0.3687", "0.3723", "0.4978", "0.3723", "0.4978", "0.386", "0.3687", "0.386" ], [ "0.594", "0.3694", "0.6533", "0.3718", "0.6514", "0.386", "0.5938", "0.3843" ], [ "0.11584", "0.3943", "0.882", "0.3955", "0.882", "0.4104", "0.11584", "0.4092" ], [ "0.11755", "0.4104", "0.881", "0.4104", "0.881", "0.4233", "0.11755", "0.4233" ], [ "0.11755", "0.4246", "0.881", "0.4246", "0.881", "0.4382", "0.11755", "0.4382" ], [ "0.11584", "0.4395", "0.8076", "0.4395", "0.8076", "0.4531", "0.11584", "0.4531" ], [ "0.1486", "0.456", "0.2048", "0.456", "0.2048", "0.4673", "0.1486", "0.4673" ], [ "0.2083", "0.4548", "0.706", "0.4548", "0.706", "0.4685", "0.2083", "0.4685" ], [ "0.1335", "0.4673", "0.2007", "0.4702", "0.199", "0.485", "0.1318", "0.4822" ], [ "0.215", "0.4697", "0.6963", "0.4697", "0.6963", "0.4827", "0.215", "0.4827" ], [ "0.1528", "0.485", "0.199", "0.485", "0.199", "0.4963", "0.1528", "0.4963" ], [ "0.2124", "0.484", "0.8774", "0.484", "0.8774", "0.4988", "0.2124", "0.4988" ], [ "0.1503", "0.5005", "0.199", "0.5005", "0.199", "0.5127", "0.1503", "0.5127" ], [ "0.2091", "0.4988", "0.8555", "0.4995", "0.8555", "0.514", "0.2091", "0.513" ], [ "0.1301", "0.513", "0.2007", "0.5156", "0.2007", "0.53", "0.1301", "0.528" ], [ "0.215", "0.5166", "0.6465", "0.5166", "0.6465", "0.527", "0.215", "0.527" ], [ "0.1511", "0.5293", "0.2024", "0.5293", "0.2024", "0.5435", "0.1511", "0.5435" ], [ "0.2124", "0.53", "0.8286", "0.53", "0.8286", "0.5444", "0.2124", "0.5444" ], [ "0.1486", "0.5464", "0.2007", "0.5464", "0.2007", "0.558", "0.1486", "0.558" ], [ "0.2107", "0.546", "0.8774", "0.546", "0.8774", "0.5586", "0.2107", "0.5586" ], [ "0.2124", "0.56", "0.7925", "0.56", "0.7925", "0.5737", "0.2124", "0.5737" ], [ "0.1335", "0.5723", "0.2024", "0.576", "0.2007", "0.5903", "0.1318", "0.5874" ], [ "0.215", "0.5747", "0.398", "0.5747", "0.398", "0.5884", "0.215", "0.5884" ], [ "0.1528", "0.5903", "0.2048", "0.5903", "0.2048", "0.6016", "0.1528", "0.6016" ], [ "0.2091", "0.589", "0.8213", "0.589", "0.8213", "0.604", "0.2091", "0.604" ], [ "0.11755", "0.61", "0.8833", "0.61", "0.8833", "0.6235", "0.11755", "0.6235" ], [ "0.11755", "0.6245", "0.882", "0.6245", "0.882", "0.638", "0.11755", "0.638" ], [ "0.11584", "0.6396", "0.669", "0.6396", "0.669", "0.653", "0.11584", "0.653" ], [ "0.1184", "0.662", "0.486", "0.662", "0.486", "0.677", "0.1184", "0.677" ], [ "0.511", "0.662", "0.881", "0.662", "0.881", "0.677", "0.511", "0.677" ], [ "0.1184", "0.6777", "0.4844", "0.6777", "0.4844", "0.6914", "0.1184", "0.6914" ], [ "0.509", "0.6777", "0.882", "0.676", "0.882", "0.6934", "0.509", "0.6953" ], [ "0.12177", "0.695", "0.4895", "0.695", "0.4895", "0.708", "0.12177", "0.708" ], [ "0.513", "0.695", "0.881", "0.695", "0.881", "0.708", "0.513", "0.708" ], [ "0.11755", "0.709", "0.4878", "0.7104", "0.4878", "0.7246", "0.11755", "0.7236" ], [ "0.513", "0.7104", "0.881", "0.7104", "0.881", "0.7236", "0.513", "0.7236" ], [ "0.1184", "0.727", "0.486", "0.727", "0.486", "0.7407", "0.1184", "0.7407" ], [ "0.5146", "0.727", "0.879", "0.727", "0.879", "0.7407", "0.5146", "0.7407" ], [ "0.11755", "0.742", "0.486", "0.742", "0.486", "0.7573", "0.11755", "0.7573" ], [ "0.513", "0.742", "0.882", "0.742", "0.882", "0.7573", "0.513", "0.7573" ], [ "0.11755", "0.759", "0.1948", "0.759", "0.1948", "0.774", "0.11755", "0.774" ], [ "0.511", "0.7583", "0.854", "0.7583", "0.854", "0.7715", "0.511", "0.7715" ], [ "0.1351", "0.7754", "0.486", "0.7754", "0.486", "0.789", "0.1351", "0.789" ], [ "0.6943", "0.778", "0.8135", "0.778", "0.8135", "0.792", "0.6943", "0.792" ], [ "0.538", "0.7876", "0.6753", "0.7876", "0.6753", "0.8003", "0.538", "0.8003" ], [ "0.858", "0.786", "0.882", "0.786", "0.882", "0.799", "0.858", "0.799" ], [ "0.687", "0.792", "0.8286", "0.7935", "0.8286", "0.808", "0.687", "0.807" ], [ "0.687", "0.804", "0.7095", "0.8057", "0.708", "0.8164", "0.686", "0.814" ], [ "0.11755", "0.7905", "0.486", "0.7915", "0.486", "0.8057", "0.11755", "0.8047" ], [ "0.1184", "0.808", "0.4844", "0.808", "0.4844", "0.822", "0.1184", "0.822" ], [ "0.1184", "0.8247", "0.486", "0.8247", "0.486", "0.8384", "0.1184", "0.8384" ], [ "0.511", "0.822", "0.881", "0.8223", "0.881", "0.8374", "0.511", "0.836" ], [ "0.1184", "0.8403", "0.486", "0.8403", "0.486", "0.854", "0.1184", "0.854" ], [ "0.51", "0.8384", "0.881", "0.839", "0.881", "0.854", "0.51", "0.8525" ], [ "0.11755", "0.8555", "0.4844", "0.8555", "0.4844", "0.869", "0.11755", "0.869" ], [ "0.511", "0.8555", "0.882", "0.8555", "0.882", "0.8706", "0.511", "0.8706" ], [ "0.1184", "0.8726", "0.486", "0.8726", "0.486", "0.887", "0.1184", "0.887" ], [ "0.511", "0.8706", "0.882", "0.8716", "0.882", "0.8857", "0.511", "0.8853" ], [ "0.11755", "0.888", "0.486", "0.887", "0.486", "0.902", "0.11755", "0.903" ], [ "0.513", "0.887", "0.881", "0.887", "0.881", "0.902", "0.513", "0.902" ], [ "0.11755", "0.905", "0.4844", "0.905", "0.4844", "0.9185", "0.11755", "0.9185" ], [ "0.511", "0.8994", "0.65", "0.904", "0.6484", "0.9214", "0.51", "0.9175" ] ]

[ "Clean Data D,", "Data Weights wi", "1.Candidate Rules Generation", "Large-error Instance xe", "++", "Microsoft is an American technology corporation founded by Bill Gates.", "Prompt Template xp", "[Input] The PERSON Bill Gates [MASK] the ORGANIZATION Microsoft.", "Model mt-1", "2.InteractiveRule Evaluation", "founded", "Rule1:PERSON[founded]ORGANIZATION", "V", "started", "V", "called", "区", "Rule2:PERSON[started]ORGANIZATION", "p(MASK=DIxp)", "Human-selected Rules R+", "3.WeaklySupervised Model Training&Ensemble", "Human Annotators", "Self-training", "D, UDt-1", "", "", "", "Model Ensemble", "Unmatched Data Du", "Rule-matched Data Dr", "Model mt", "Figure l: Overall framework for PRBOOST. In each iteration, PRBOOST (1) identifies large-error instances", "from the limited clean data and converts each large-error instance to a prompt template for prompting-based rule", "discovery; (2) presents candidate rules to human experts for annotation and uses accepted rules to generate new", "weak labels; (3) trains a new weak model with self-training and ensembles it with the previous models.", "Input :", "Microsoft is an American technology corporation founded by Bill Gates.", "Prompt :", "[ Input ] The Person Bill Gates [Mask] the Organization Microsoft.", "Rule :", "[Entity Pair ==(Person, Org) } ^( [Mask] == founded ]^[ st,j ≥ threshold ] → per:found", "Input :", "Marvell Software Solutions Israel is a wholly owned subsidiary of Marvell Technology Group", "Prompt :", "[Input]TheMarvell SoftwareSolutionsIsraelisa [Mask].", "Rule :", "( [Mask] == subsidiary V corporation V company ] ^( st,j ≥ threshold -→Company", "Input :", "Liverpool short of firepower for crucial encounter. Rafael Benitez must gamble with Liverpools", "Champions League prospects tonight but lacks the ammunition to make it a fair fight.", "Prompt :", "[Mask] News: [Input]", "Rule:", "{ [Mask] == Liverpool V Team V Football V Sports ] ^ [ st, ≥ threshold ] → Sports", "Table I: The examples of prompt-based rules for relation extraction, ontology classification, and news topic clas-", "sification. Here [Input ] denotes the original input, [Mask] denotes the mask token, and A, V are the logical", "operators. We use bold words to show the ground-truth label of the original input.", "我们都有一个特定于任务的模板来重塑原始输", "定一个大错误实例ae,EXe，我们首先通过", "入以提示PLM。生成的提示通常包括作为上", "ap=T(ace）将其转换为提示。这样的提示由", "下文的原始输入和要由PLM填充的掩码令牌。", "原始输人的关键组件和【MASK】令牌组成。通", "最终规则包含多个原子部分，以捕获不同的信", "过继承原始输入，我们为【MASK】令牌构造上", "息视图。每个规则都伴随着原始输入实例的真", "下文，以便由预先训练的LMM预测。为了", "实标签，这样的标签将分配给此规则匹配的未", "完成规则，我们将每个αp馈送到M，以获", "标记实例", "得「MASK】今牌在词汇表V上的概率分布", "例如，如表日1所示，关系提取任务的提示", "exp(o- M(ap))", "p(MASK = | p) =", "(4)", " exp(u-M(ap:)", "VEV", "可以是“实体【MASK】实体”，它在保持关键", "语义的同时使用关系短语对原始输入进行重", "新短语。以新闻主题分类为例，通过填充提示", "其中M）表示M的输出向量，是预测", "中的屏蔽槽，PLM提出了基于候选关键字的", "的XMATHX嵌入1，。我们收集具有最高", "主题分类规则。与从语料库的表面模式中提取", "p(MASK=|p）的top-k预测以形成候", "的规则不同（e。g。，n-gram规则），这样的", "选规则。通过使用提示预测填充基于&e的规", "基于提示的规则建议可以生成原始输入中未", "则，我们在送代t中获得了候选规则集，表示", "出现的单词---此功能对于模型泛化很重要。给", "为 Rt=(rg-1 " ]

{ "arxiv_id": "2203.09735", "language": "zh", "page_idx": 4 }

[ [ "0.1814", "0.1256", "0.8145", "0.1256", "0.8145", "0.1401", "0.1814", "0.1401" ], [ "0.2566", "0.1509", "0.7485", "0.1509", "0.7485", "0.1718", "0.2566", "0.1718" ], [ "0.246", "0.1724", "0.749", "0.1724", "0.749", "0.1875", "0.246", "0.1875" ], [ "0.2754", "0.1901", "0.7236", "0.1901", "0.7236", "0.2052", "0.2754", "0.2052" ], [ "0.2754", "0.2026", "0.7236", "0.2039", "0.7236", "0.2184", "0.2754", "0.2172" ], [ "0.3406", "0.2184", "0.6567", "0.2184", "0.6567", "0.2317", "0.3406", "0.2317" ], [ "0.4167", "0.2311", "0.581", "0.2317", "0.581", "0.2468", "0.4167", "0.2456" ], [ "0.08905", "0.2468", "0.911", "0.2468", "0.911", "0.26", "0.08905", "0.26" ], [ "0.25", "0.2734", "0.3145", "0.2734", "0.3145", "0.2886", "0.25", "0.2886" ], [ "0.546", "0.2747", "0.649", "0.2734", "0.649", "0.2854", "0.546", "0.286" ], [ "0.6724", "0.2715", "0.834", "0.2715", "0.834", "0.2822", "0.6724", "0.2822" ], [ "0.6724", "0.2827", "0.755", "0.2827", "0.755", "0.2935", "0.6724", "0.2935" ], [ "0.1013", "0.2954", "0.46", "0.2954", "0.46", "0.3093", "0.1013", "0.3093" ], [ "0.603", "0.2954", "0.6436", "0.2954", "0.6436", "0.3062", "0.603", "0.3062" ], [ "0.603", "0.3037", "0.6562", "0.3037", "0.6562", "0.3145", "0.603", "0.3145" ], [ "0.751", "0.308", "0.9087", "0.308", "0.9087", "0.318", "0.751", "0.318" ], [ "0.744", "0.3176", "0.9067", "0.3176", "0.9067", "0.3276", "0.744", "0.3276" ], [ "0.10297", "0.3093", "0.4631", "0.3093", "0.4631", "0.3225", "0.10297", "0.3225" ], [ "0.10297", "0.3213", "0.4631", "0.3213", "0.4631", "0.336", "0.10297", "0.336" ], [ "0.10297", "0.334", "0.4631", "0.334", "0.4631", "0.348", "0.10297", "0.348" ], [ "0.568", "0.3447", "0.6274", "0.3447", "0.6274", "0.3516", "0.568", "0.3516" ], [ "0.7573", "0.3416", "0.88", "0.3416", "0.88", "0.3484", "0.7573", "0.3484" ], [ "0.10297", "0.348", "0.4617", "0.348", "0.4617", "0.361", "0.10297", "0.361" ], [ "0.744", "0.3484", "0.884", "0.3484", "0.884", "0.3591", "0.744", "0.3591" ], [ "0.10297", "0.3591", "0.4631", "0.3591", "0.4631", "0.3726", "0.10297", "0.3726" ], [ "0.607", "0.3623", "0.6587", "0.3623", "0.6587", "0.3726", "0.607", "0.3726" ], [ "0.7686", "0.3591", "0.841", "0.3591", "0.841", "0.3694", "0.7686", "0.3694" ], [ "0.10297", "0.3718", "0.46", "0.3718", "0.46", "0.3865", "0.10297", "0.3865" ], [ "0.6006", "0.375", "0.6567", "0.375", "0.6567", "0.3865", "0.6006", "0.3865" ], [ "0.749", "0.3757", "0.8784", "0.377", "0.8784", "0.387", "0.749", "0.3865" ], [ "0.10297", "0.3838", "0.4592", "0.3838", "0.4592", "0.3977", "0.10297", "0.3977" ], [ "0.683", "0.3914", "0.6953", "0.3914", "0.6953", "0.3945", "0.683", "0.3945" ], [ "0.7427", "0.3865", "0.898", "0.3865", "0.898", "0.3965", "0.7427", "0.3965" ], [ "0.10297", "0.3965", "0.4617", "0.3965", "0.4617", "0.4104", "0.10297", "0.4104" ], [ "0.751", "0.3958", "0.8423", "0.3958", "0.8423", "0.406", "0.751", "0.406" ], [ "0.1013", "0.4077", "0.4617", "0.4092", "0.4617", "0.4236", "0.1013", "0.4224" ], [ "0.10297", "0.4216", "0.4617", "0.4216", "0.4617", "0.4363", "0.10297", "0.4363" ], [ "0.5215", "0.4248", "0.9126", "0.4248", "0.9126", "0.4382", "0.5215", "0.4382" ], [ "0.10297", "0.4343", "0.46", "0.4343", "0.46", "0.4475", "0.10297", "0.4475" ], [ "0.5195", "0.4395", "0.9126", "0.4395", "0.9126", "0.4526", "0.5195", "0.4526" ], [ "0.10297", "0.4475", "0.4617", "0.4475", "0.4617", "0.461", "0.10297", "0.461" ], [ "0.5195", "0.4526", "0.91", "0.4526", "0.91", "0.4666", "0.5195", "0.4666" ], [ "0.1054", "0.4634", "0.1732", "0.4634", "0.1732", "0.4702", "0.1054", "0.4702" ], [ "0.179", "0.4602", "0.46", "0.4602", "0.46", "0.4734", "0.179", "0.4734" ], [ "0.10297", "0.473", "0.4617", "0.473", "0.4617", "0.486", "0.10297", "0.486" ], [ "0.5195", "0.4685", "0.5483", "0.4685", "0.5483", "0.481", "0.5195", "0.481" ], [ "0.10297", "0.4849", "0.4617", "0.4849", "0.4617", "0.4988", "0.10297", "0.4988" ], [ "0.10297", "0.4976", "0.4631", "0.4976", "0.4631", "0.511", "0.10297", "0.511" ], [ "0.10297", "0.5103", "0.4617", "0.5103", "0.4617", "0.5234", "0.10297", "0.5234" ], [ "0.539", "0.513", "0.91", "0.513", "0.91", "0.5234", "0.539", "0.5234" ], [ "0.10297", "0.523", "0.4617", "0.523", "0.4617", "0.536", "0.10297", "0.536" ], [ "0.5215", "0.526", "0.7183", "0.526", "0.7183", "0.539", "0.5215", "0.539" ], [ "0.7114", "0.526", "0.911", "0.526", "0.911", "0.539", "0.7114", "0.539" ], [ "0.10455", "0.537", "0.4592", "0.537", "0.4592", "0.5474", "0.10455", "0.5474" ], [ "0.5215", "0.539", "0.911", "0.539", "0.911", "0.553", "0.5215", "0.553" ], [ "0.10297", "0.549", "0.46", "0.549", "0.46", "0.5625", "0.10297", "0.5625" ], [ "0.5215", "0.5537", "0.911", "0.5537", "0.911", "0.5674", "0.5215", "0.5674" ], [ "0.10297", "0.56", "0.4617", "0.56", "0.4617", "0.5737", "0.10297", "0.5737" ], [ "0.1013", "0.5713", "0.4592", "0.5737", "0.4592", "0.587", "0.1013", "0.5854" ], [ "0.5215", "0.5674", "0.911", "0.5674", "0.911", "0.5806", "0.5215", "0.5806" ], [ "0.10455", "0.586", "0.46", "0.586", "0.46", "0.5967", "0.10455", "0.5967" ], [ "0.523", "0.5835", "0.732", "0.5835", "0.732", "0.5933", "0.523", "0.5933" ], [ "0.10455", "0.5996", "0.4617", "0.5996", "0.4617", "0.61", "0.10455", "0.61" ], [ "0.5376", "0.5967", "0.9126", "0.5967", "0.9126", "0.61", "0.5376", "0.61" ], [ "0.10297", "0.6113", "0.4631", "0.6113", "0.4631", "0.625", "0.10297", "0.625" ], [ "0.5215", "0.6113", "0.9126", "0.6113", "0.9126", "0.625", "0.5215", "0.625" ], [ "0.1013", "0.6226", "0.392", "0.6226", "0.392", "0.636", "0.1013", "0.636" ], [ "0.5215", "0.625", "0.9126", "0.625", "0.9126", "0.638", "0.5215", "0.638" ], [ "0.5215", "0.638", "0.9126", "0.638", "0.9126", "0.6514", "0.5215", "0.6514" ], [ "0.2427", "0.657", "0.353", "0.657", "0.353", "0.6704", "0.2427", "0.6704" ], [ "0.5215", "0.6514", "0.911", "0.6514", "0.911", "0.6655", "0.5215", "0.6655" ], [ "0.2133", "0.6577", "0.2303", "0.6577", "0.2303", "0.6675", "0.2133", "0.6675" ], [ "0.232", "0.661", "0.2385", "0.661", "0.2385", "0.6655", "0.232", "0.6655" ], [ "0.523", "0.6665", "0.911", "0.6665", "0.911", "0.6797", "0.523", "0.6797" ], [ "0.08826", "0.675", "0.4763", "0.675", "0.4763", "0.688", "0.08826", "0.688" ], [ "0.5215", "0.6797", "0.911", "0.6797", "0.911", "0.6934", "0.5215", "0.6934" ], [ "0.08826", "0.6895", "0.478", "0.6895", "0.478", "0.7026", "0.08826", "0.7026" ], [ "0.523", "0.6934", "0.911", "0.6934", "0.911", "0.707", "0.523", "0.707" ], [ "0.08905", "0.7026", "0.478", "0.7026", "0.478", "0.7163", "0.08905", "0.7163" ], [ "0.523", "0.7085", "0.911", "0.7085", "0.911", "0.7217", "0.523", "0.7217" ], [ "0.08826", "0.7173", "0.478", "0.7173", "0.478", "0.731", "0.08826", "0.731" ], [ "0.5215", "0.7217", "0.9126", "0.7217", "0.9126", "0.735", "0.5215", "0.735" ], [ "0.08826", "0.731", "0.4697", "0.731", "0.4697", "0.744", "0.08826", "0.744" ], [ "0.5215", "0.735", "0.7837", "0.735", "0.7837", "0.7485", "0.5215", "0.7485" ], [ "0.08905", "0.744", "0.4763", "0.744", "0.4763", "0.7583", "0.08905", "0.7583" ], [ "0.5376", "0.7505", "0.9126", "0.7505", "0.9126", "0.7646", "0.5376", "0.7646" ], [ "0.0907", "0.7583", "0.4756", "0.7583", "0.4756", "0.7715", "0.0907", "0.7715" ], [ "0.5195", "0.763", "0.911", "0.7646", "0.911", "0.779", "0.5195", "0.778" ], [ "0.08826", "0.7715", "0.4795", "0.7715", "0.4795", "0.785", "0.08826", "0.785" ], [ "0.5215", "0.778", "0.9126", "0.778", "0.9126", "0.7915", "0.5215", "0.7915" ], [ "0.08826", "0.786", "0.478", "0.786", "0.478", "0.8", "0.08826", "0.8" ], [ "0.523", "0.7925", "0.911", "0.7925", "0.911", "0.806", "0.523", "0.806" ], [ "0.0866", "0.8", "0.2345", "0.8", "0.2345", "0.813", "0.0866", "0.813" ], [ "0.523", "0.806", "0.911", "0.806", "0.911", "0.8193", "0.523", "0.8193" ], [ "0.10455", "0.813", "0.478", "0.813", "0.478", "0.827", "0.10455", "0.827" ], [ "0.5215", "0.8193", "0.911", "0.8193", "0.911", "0.8335", "0.5215", "0.8335" ], [ "0.0866", "0.8276", "0.478", "0.8276", "0.478", "0.8413", "0.0866", "0.8413" ], [ "0.0866", "0.839", "0.1724", "0.8413", "0.1724", "0.856", "0.0866", "0.854" ], [ "0.5215", "0.834", "0.911", "0.834", "0.911", "0.8477", "0.5215", "0.8477" ], [ "0.5215", "0.8477", "0.9126", "0.8477", "0.9126", "0.8613", "0.5215", "0.8613" ], [ "0.08496", "0.8613", "0.4756", "0.8623", "0.4756", "0.877", "0.08496", "0.8755" ], [ "0.5195", "0.8613", "0.911", "0.86", "0.911", "0.875", "0.5195", "0.8755" ], [ "0.0866", "0.8755", "0.4495", "0.8755", "0.4495", "0.8896", "0.0866", "0.8896" ], [ "0.5215", "0.8755", "0.911", "0.8755", "0.911", "0.8896", "0.5215", "0.8896" ] ]

[ "WhatdoyouMEME？为模因中的视觉语义角色标记生成解释", "Shivam Sharma, 1,4 Siddhant Agarwal, 1 Tharun Suresh, 1", "Preslav Nakov, 2 Md. Shad Akhtar, 1 Tanmoy Chakraborty", "'Indraprastha Institute of Information Technology Delhi, India", "2Mohamed bin Zayed University of Artifcial Intelligence, UAE", "3Indian Institute of Technology Delhi, India", "4Wipro AI Labs, India", "[ shivams, siddhant20247, tharun20119, shad.akhtar J @itd.ac.in, [email protected], [email protected]", "Abstract", "TRUMP2WEEKSAG0", "Entity: democratic party", "Role: victim", "Memes are powerful means for effective communication", "The Coronavi", "a Democrat Hoax!", "ocraticparty is depictec", "being unfairly criticizec", "on social media. Their effortless amalgamation of viral", "visuals and compelling messages can have far-reaching", "implications with proper marketing. Previous research", "TRUMPTODAY", "democratic party is", "on memes has primarily focused on characterizing their", "rtrayed as a victim of", "affective spectrum and detecting whether the meme's", "T'm declaring the", "allegations", "message insinuates any intended harm, such as hate", "National Emergency", "democratic partyis", "offense, racism , etc.However, memes often use ab", "", "ortrayedasbeingfalsely", "straction, which can be elusive. Here, we introduce a", "used of ahoax", "novel task- ExCLAIM, generating explanations for", "visual semantic role labeling in memes. To this end, we", "Figure l: A novel task of EXCLAIM. Given a meme, a", "curate ExHVV, a novel dataset that offers natural lan-", "referred entity and its semantic role, generate a natural", "guage explanations of connotative roles for three types", "language explanation that emulates human-like reason-", "ofentities", "-heroes ,villains ,and victims,encom-", "passing 4,680 entities present in 3K memes. We also", "ing.", "benchmark ExHVV with several strong unimodal and", "multimodal baselines. Moreover, we posit LUMEN, a", "novel multimodal, multi-task learning framework that", "我们尝试在检测模因的语义角色标签时利用所学的", "endeavors to address EXCLAIM optimally by jointly", "功能：英雄，恶棍和victim", "，以生成合理的自然语言解", "learningtopredictthecorrectsemanticrolesandcor-", "释。在这样做的过程中，我们的目标是将模因与基于社", "respondingly to generate suitable natural language ex-", "交媒体的关键用例（如内容审核和数字营销）相关联。。", "planations. LUMEN distinctly outperforms the best", "baseline across 18 standard natural language genera-", "生成的解释可以帮助对角色标签进行情境化，并有助于", "tion evaluation metrics.Our systematic evaluation and", "对战略洞察力进行回顾性推断", "analyses demonstrate that characteristic multimodal", "模因在低毁、伤害或美化目标实体时，经常含蓄地讽", "cues required for adjudicating semantic roles are also", "刺、讽刺、幽默和讽刺。。由于需要进行抽象推理和多模", "helpful for generating suitable explanations", "式语境化，因此为这些方面生成解释具有挑战性，因为", "角色标签和相应的解释是相辅相成的。例如，图E1描", "Introduction", "绘了一个模因，其中唐纳德·特朗普被描绘成错误地暗", "1", "", "示民主党炮制了“冠状病毒”的诡计，但没有明确说明。", "近年来，模因已成为交流复杂思想和讨论社会问题的必", "因此，根据模因的符号学，唐纳德·特朗普需要被判定", "要条件。除了是互联网上的一种发展现象外，模因视觉", "为恶棍，涉及他在所描绘的场景中的讽刺立场逆转。因", "食谱还通过适应不断发展的不同社会观点而不断变异，", "此，民主党作为受害者的角色需要根据隐含的讽刺来确", "使它们变得深奥。由于它们的设计可访问性，模因越来", "定。。EXCLAIM要求一个解释代理来考虑这种复杂性，", "越多地传播以造成各种伤害（Pramanicketal.202la）", "同时为上述语义角色生成合适的解释", "包括讨厌（Kielaetal.2020a），冒犯（Shangetal.2021）", "众所周知，文本模态有助于检测仇恨模因（Kielaetal.", "，拖钓（SuryawanshiandChakravarthi2021）等。最近", "2020a），有害模因及其目标类型（Pramanicketal.", "已经朝着自动检测此类模因（Sharmaetal.2022b）的", "2021a,b），模因情感（Sharmaetal.2020），模态内不", "方向发展。。然而，在考虑多模式和上下文依赖等方面", "协调(Panetal.2020）以及模因（Sharmaetal.2022c)", "时，限制开始具体化", "中多种情感内涵的语义角色标签。另一方面，一个罕见", "例如，在模因（Sharmaetal.2022a,c）中检测社会实", "的任务表现出视觉影响是检测有害实体Sharma etal.", "体的多模态叙事框架并对固有的上下文推理进行建模", "具有挑战性", "2022a），它包含一个单一的情感内涵。此外，已观察", "到图像模态在检测模因情绪（Singh，Bauwelinck，and", "Copyright 2023, Association for the Advancement of Ar-", "Lefever2020；Ruizetal.2020）方面表现不佳。因此，模", "tificial Intelligence (www.aaai.org). All rights reserved.", "态影响通常因任务而异。但是，可以为任务复杂性或特" ]

{ "arxiv_id": "2212.00715", "language": "zh", "page_idx": 1 }

[ [ "0.08826", "0.06946", "0.4795", "0.0682", "0.4795", "0.0821", "0.08826", "0.0833" ], [ "0.63", "0.06946", "0.694", "0.06946", "0.694", "0.0802", "0.63", "0.0802" ], [ "0.7075", "0.06946", "0.7563", "0.06946", "0.7563", "0.0802", "0.7075", "0.0802" ], [ "0.773", "0.07135", "0.8", "0.0745", "0.7974", "0.0852", "0.7695", "0.0821" ], [ "0.814", "0.0745", "0.8423", "0.0745", "0.8423", "0.0833", "0.814", "0.0833" ], [ "0.705", "0.07764", "0.7417", "0.07764", "0.7417", "0.0884", "0.705", "0.0884" ], [ "0.5933", "0.0745", "0.619", "0.0745", "0.619", "0.0833", "0.5933", "0.0833" ], [ "0.6274", "0.0789", "0.665", "0.0789", "0.665", "0.089", "0.6274", "0.089" ], [ "0.6685", "0.0802", "0.6914", "0.0802", "0.6914", "0.0871", "0.6685", "0.0871" ], [ "0.738", "0.0789", "0.759", "0.0789", "0.759", "0.0871", "0.738", "0.0871" ], [ "0.08905", "0.0852", "0.1929", "0.0852", "0.1929", "0.09534", "0.08905", "0.09534" ], [ "0.5864", "0.089", "0.625", "0.09155", "0.6235", "0.1016", "0.585", "0.09973" ], [ "0.671", "0.0903", "0.693", "0.0903", "0.693", "0.1029", "0.671", "0.1029" ], [ "0.706", "0.0903", "0.728", "0.0903", "0.728", "0.1029", "0.706", "0.1029" ], [ "0.751", "0.0884", "0.761", "0.0966", "0.751", "0.1048", "0.74", "0.0966" ], [ "0.7754", "0.09155", "0.796", "0.09155", "0.796", "0.1016", "0.7754", "0.1016" ], [ "0.814", "0.0903", "0.841", "0.0903", "0.841", "0.1016", "0.814", "0.1016" ], [ "0.10297", "0.09784", "0.4795", "0.09784", "0.4795", "0.1111", "0.10297", "0.1111" ], [ "0.6357", "0.09216", "0.66", "0.09216", "0.66", "0.101", "0.6357", "0.101" ], [ "0.7485", "0.09784", "0.7607", "0.1029", "0.7534", "0.11237", "0.7417", "0.1067" ], [ "0.814", "0.09784", "0.841", "0.101", "0.8384", "0.1111", "0.8105", "0.108" ], [ "0.777", "0.09973", "0.794", "0.09973", "0.794", "0.1092", "0.777", "0.1092" ], [ "0.589", "0.0985", "0.6235", "0.101", "0.6216", "0.1111", "0.5884", "0.1092" ], [ "0.638", "0.1016", "0.6562", "0.1016", "0.6562", "0.1092", "0.638", "0.1092" ], [ "0.6763", "0.1029", "0.689", "0.1029", "0.689", "0.1092", "0.6763", "0.1092" ], [ "0.709", "0.101", "0.726", "0.101", "0.726", "0.1092", "0.709", "0.1092" ], [ "0.709", "0.108", "0.728", "0.1111", "0.7256", "0.1187", "0.706", "0.11554" ], [ "0.749", "0.1067", "0.759", "0.11426", "0.751", "0.1193", "0.7427", "0.113" ], [ "0.8164", "0.1092", "0.839", "0.1092", "0.839", "0.1206", "0.8164", "0.1206" ], [ "0.638", "0.10986", "0.6543", "0.10986", "0.6543", "0.1187", "0.638", "0.1187" ], [ "0.591", "0.1092", "0.6216", "0.1092", "0.6216", "0.1206", "0.591", "0.1206" ], [ "0.675", "0.10986", "0.6895", "0.10986", "0.6895", "0.1193", "0.675", "0.1193" ], [ "0.778", "0.10986", "0.7915", "0.10986", "0.7915", "0.1193", "0.778", "0.1193" ], [ "0.8145", "0.11554", "0.847", "0.1193", "0.8447", "0.132", "0.812", "0.1276" ], [ "0.6357", "0.11743", "0.6587", "0.1193", "0.6567", "0.1276", "0.634", "0.1256" ], [ "0.5884", "0.1187", "0.625", "0.1187", "0.625", "0.1288", "0.5884", "0.1288" ], [ "0.0866", "0.1111", "0.478", "0.1111", "0.478", "0.125", "0.0866", "0.125" ], [ "0.7417", "0.1187", "0.759", "0.1187", "0.759", "0.1276", "0.7417", "0.1276" ], [ "0.777", "0.1187", "0.794", "0.1187", "0.794", "0.1276", "0.777", "0.1276" ], [ "0.7075", "0.1193", "0.7305", "0.1193", "0.7305", "0.1276", "0.7075", "0.1276" ], [ "0.6763", "0.1193", "0.6914", "0.1193", "0.6914", "0.1276", "0.6763", "0.1276" ], [ "0.0866", "0.1256", "0.478", "0.1256", "0.478", "0.1395", "0.0866", "0.1395" ], [ "0.0989", "0.1395", "0.478", "0.1395", "0.478", "0.1528", "0.0989", "0.1528" ], [ "0.5186", "0.1414", "0.7363", "0.1433", "0.7363", "0.1572", "0.5186", "0.1547" ], [ "0.738", "0.1477", "0.744", "0.1477", "0.744", "0.1528", "0.738", "0.1528" ], [ "0.74", "0.1433", "0.9087", "0.1433", "0.9087", "0.1572", "0.74", "0.1572" ], [ "0.0866", "0.1528", "0.478", "0.1528", "0.478", "0.1666", "0.0866", "0.1666" ], [ "0.5195", "0.1572", "0.9126", "0.1572", "0.9126", "0.1704", "0.5195", "0.1704" ], [ "0.08496", "0.1654", "0.4795", "0.1666", "0.4795", "0.1812", "0.08496", "0.1799" ], [ "0.5215", "0.1724", "0.5728", "0.1724", "0.5728", "0.1843", "0.5215", "0.1843" ], [ "0.08826", "0.1812", "0.478", "0.1812", "0.478", "0.1945", "0.08826", "0.1945" ], [ "0.08826", "0.1957", "0.478", "0.1957", "0.478", "0.209", "0.08826", "0.209" ], [ "0.0866", "0.2084", "0.2688", "0.209", "0.2688", "0.2203", "0.0866", "0.2197" ], [ "0.5195", "0.2108", "0.9126", "0.2122", "0.9126", "0.2267", "0.5195", "0.226" ], [ "0.1054", "0.2229", "0.478", "0.2229", "0.478", "0.2361", "0.1054", "0.2361" ], [ "0.5195", "0.2267", "0.9087", "0.2267", "0.9087", "0.2405", "0.5195", "0.2405" ], [ "0.08826", "0.2361", "0.478", "0.2361", "0.478", "0.25", "0.08826", "0.25" ], [ "0.5215", "0.2405", "0.911", "0.2405", "0.911", "0.2551", "0.5215", "0.2551" ], [ "0.0866", "0.2488", "0.478", "0.25", "0.478", "0.2644", "0.0866", "0.2632" ], [ "0.5215", "0.2551", "0.774", "0.2551", "0.774", "0.2683", "0.5215", "0.2683" ], [ "0.08826", "0.2644", "0.478", "0.2644", "0.478", "0.2778", "0.08826", "0.2778" ], [ "0.08826", "0.2778", "0.4763", "0.2778", "0.4763", "0.2917", "0.08826", "0.2917" ], [ "0.5312", "0.2766", "0.911", "0.2766", "0.911", "0.2903", "0.5312", "0.2903" ], [ "0.0866", "0.2917", "0.4756", "0.2922", "0.4756", "0.3062", "0.0866", "0.305" ], [ "0.5215", "0.2903", "0.911", "0.2903", "0.911", "0.3037", "0.5215", "0.3037" ], [ "0.08826", "0.305", "0.478", "0.305", "0.478", "0.3193", "0.08826", "0.3193" ], [ "0.5215", "0.3037", "0.911", "0.3037", "0.911", "0.3176", "0.5215", "0.3176" ], [ "0.08826", "0.3193", "0.478", "0.3193", "0.478", "0.3333", "0.08826", "0.3333" ], [ "0.5195", "0.3164", "0.911", "0.3176", "0.911", "0.332", "0.5195", "0.3308" ], [ "0.0866", "0.3333", "0.2222", "0.3333", "0.2222", "0.3467", "0.0866", "0.3467" ], [ "0.523", "0.3333", "0.9087", "0.3333", "0.9087", "0.3435", "0.523", "0.3435" ], [ "0.5215", "0.3452", "0.9126", "0.3452", "0.9126", "0.3591", "0.5215", "0.3591" ], [ "0.08496", "0.3528", "0.4756", "0.3542", "0.4756", "0.3687", "0.08496", "0.3674" ], [ "0.5215", "0.3591", "0.9126", "0.3591", "0.9126", "0.3726", "0.5215", "0.3726" ], [ "0.10706", "0.3662", "0.4763", "0.3662", "0.4763", "0.3806", "0.10706", "0.3806" ], [ "0.5215", "0.3738", "0.9126", "0.3738", "0.9126", "0.387", "0.5215", "0.387" ], [ "0.10864", "0.3833", "0.1609", "0.3833", "0.1609", "0.3945", "0.10864", "0.3945" ], [ "0.523", "0.387", "0.911", "0.387", "0.911", "0.401", "0.523", "0.401" ], [ "0.0866", "0.3984", "0.4756", "0.3997", "0.4756", "0.4136", "0.0866", "0.4124" ], [ "0.523", "0.401", "0.9126", "0.401", "0.9126", "0.414", "0.523", "0.414" ], [ "0.10706", "0.4136", "0.4795", "0.4136", "0.4795", "0.428", "0.10706", "0.428" ], [ "0.5215", "0.4153", "0.911", "0.4167", "0.911", "0.428", "0.5215", "0.4268" ], [ "0.10864", "0.428", "0.3635", "0.428", "0.3635", "0.4414", "0.10864", "0.4414" ], [ "0.523", "0.4312", "0.911", "0.4312", "0.911", "0.4414", "0.523", "0.4414" ], [ "0.08417", "0.443", "0.4795", "0.4443", "0.4795", "0.4583", "0.08417", "0.457" ], [ "0.5215", "0.4426", "0.911", "0.4426", "0.911", "0.4558", "0.5215", "0.4558" ], [ "0.1054", "0.4583", "0.4763", "0.459", "0.4763", "0.473", "0.1054", "0.4717" ], [ "0.5215", "0.4558", "0.9126", "0.4558", "0.9126", "0.4697", "0.5215", "0.4697" ], [ "0.10706", "0.473", "0.335", "0.473", "0.335", "0.486", "0.10706", "0.486" ], [ "0.5195", "0.4697", "0.9126", "0.4702", "0.9126", "0.4841", "0.5195", "0.483" ], [ "0.5215", "0.4841", "0.911", "0.4841", "0.911", "0.4976", "0.5215", "0.4976" ], [ "0.0866", "0.4905", "0.4763", "0.4905", "0.4763", "0.504", "0.0866", "0.504" ], [ "0.524", "0.4976", "0.911", "0.4976", "0.911", "0.511", "0.524", "0.511" ], [ "0.1054", "0.504", "0.478", "0.5024", "0.478", "0.5176", "0.1054", "0.518" ], [ "0.523", "0.512", "0.556", "0.512", "0.556", "0.5244", "0.523", "0.5244" ], [ "0.10706", "0.5195", "0.1487", "0.5195", "0.1487", "0.531", "0.10706", "0.531" ], [ "0.528", "0.535", "0.91", "0.535", "0.91", "0.5454", "0.528", "0.5454" ], [ "0.2035", "0.5454", "0.3594", "0.5454", "0.3594", "0.56", "0.2035", "0.56" ], [ "0.523", "0.5474", "0.911", "0.5474", "0.911", "0.561", "0.523", "0.561" ], [ "0.08826", "0.565", "0.4756", "0.565", "0.4756", "0.579", "0.08826", "0.579" ], [ "0.5215", "0.5625", "0.911", "0.5625", "0.911", "0.5757", "0.5215", "0.5757" ], [ "0.08826", "0.58", "0.4763", "0.58", "0.4763", "0.5933", "0.08826", "0.5933" ], [ "0.5215", "0.5757", "0.91", "0.5757", "0.91", "0.589", "0.5215", "0.589" ], [ "0.08905", "0.5933", "0.3015", "0.5933", "0.3015", "0.607", "0.08905", "0.607" ], [ "0.5215", "0.5884", "0.911", "0.5884", "0.911", "0.603", "0.5215", "0.603" ], [ "0.5195", "0.6016", "0.911", "0.601", "0.911", "0.6157", "0.5195", "0.616" ], [ "0.09314", "0.617", "0.4631", "0.617", "0.4631", "0.6274", "0.09314", "0.6274" ], [ "0.5215", "0.616", "0.911", "0.616", "0.911", "0.63", "0.5215", "0.63" ], [ "0.08905", "0.629", "0.478", "0.629", "0.478", "0.643", "0.08905", "0.643" ], [ "0.523", "0.63", "0.911", "0.63", "0.911", "0.643", "0.523", "0.643" ], [ "0.08826", "0.6426", "0.478", "0.6426", "0.478", "0.656", "0.08826", "0.656" ], [ "0.523", "0.6445", "0.9126", "0.6445", "0.9126", "0.6577", "0.523", "0.6577" ], [ "0.08826", "0.659", "0.4763", "0.659", "0.4763", "0.6694", "0.08826", "0.6694" ], [ "0.5215", "0.657", "0.911", "0.6577", "0.911", "0.6724", "0.5215", "0.672" ], [ "0.0866", "0.672", "0.478", "0.672", "0.478", "0.685", "0.0866", "0.685" ], [ "0.5215", "0.672", "0.9126", "0.672", "0.9126", "0.685", "0.5215", "0.685" ], [ "0.08826", "0.685", "0.4763", "0.685", "0.4763", "0.6987", "0.08826", "0.6987" ], [ "0.5195", "0.685", "0.9126", "0.686", "0.9126", "0.6997", "0.5195", "0.6987" ], [ "0.0866", "0.6987", "0.478", "0.6987", "0.478", "0.712", "0.0866", "0.712" ], [ "0.5215", "0.6997", "0.9126", "0.6997", "0.9126", "0.7134", "0.5215", "0.7134" ], [ "0.08826", "0.711", "0.478", "0.712", "0.478", "0.7266", "0.08826", "0.726" ], [ "0.5215", "0.7134", "0.71", "0.7134", "0.71", "0.7266", "0.5215", "0.7266" ], [ "0.08826", "0.726", "0.478", "0.726", "0.478", "0.7393", "0.08826", "0.7393" ], [ "0.5376", "0.726", "0.9126", "0.7266", "0.9126", "0.741", "0.5376", "0.7407" ], [ "0.08496", "0.7407", "0.4763", "0.7407", "0.4763", "0.754", "0.08496", "0.754" ], [ "0.523", "0.741", "0.911", "0.741", "0.911", "0.755", "0.523", "0.755" ], [ "0.0866", "0.755", "0.478", "0.755", "0.478", "0.768", "0.0866", "0.768" ], [ "0.5215", "0.755", "0.7812", "0.755", "0.7812", "0.768", "0.5215", "0.768" ], [ "0.08826", "0.768", "0.4756", "0.768", "0.4756", "0.782", "0.08826", "0.782" ], [ "0.0866", "0.782", "0.4756", "0.782", "0.4756", "0.7954", "0.0866", "0.7954" ], [ "0.5664", "0.784", "0.865", "0.784", "0.865", "0.7974", "0.5664", "0.7974" ], [ "0.08496", "0.7944", "0.4763", "0.7954", "0.4763", "0.8096", "0.08496", "0.808" ], [ "0.6855", "0.802", "0.7793", "0.802", "0.7793", "0.813", "0.6855", "0.813" ], [ "0.08826", "0.811", "0.4763", "0.811", "0.4763", "0.8213", "0.08826", "0.8213" ], [ "0.0866", "0.8237", "0.4795", "0.8237", "0.4795", "0.837", "0.0866", "0.837" ], [ "0.527", "0.8193", "0.911", "0.8193", "0.911", "0.8335", "0.527", "0.8335" ], [ "0.0866", "0.836", "0.482", "0.8384", "0.482", "0.853", "0.0866", "0.851" ], [ "0.5215", "0.834", "0.911", "0.834", "0.911", "0.8477", "0.5215", "0.8477" ], [ "0.523", "0.8477", "0.9126", "0.8477", "0.9126", "0.8613", "0.523", "0.8613" ], [ "0.1054", "0.86", "0.478", "0.863", "0.478", "0.8774", "0.1054", "0.875" ], [ "0.5195", "0.86", "0.911", "0.8613", "0.911", "0.8755", "0.5195", "0.875" ], [ "0.08826", "0.8755", "0.4656", "0.8755", "0.4656", "0.8896", "0.08826", "0.8896" ], [ "0.524", "0.8755", "0.911", "0.8755", "0.911", "0.8896", "0.524", "0.8896" ] ]

[ "定于模态的输出配置推广类似的东西，就像EXCLAIM", "U.S. Politics", "Covid-19", "Test", "Total", "Train", "Split", "Train", "val", "Val", "所要求的那样？", "Villain", "217", "654", "To", "347", "3004", "Bateman（2014）将文本-图像关系（Barthesand", "1708", "AT", "1110", "104", "Victim", "531", "71", "357", "185", "00", "566", "276", "Hero", "35", "50", "4680", "2515", "Total", "Heath1978）以图形方式研究为系统网络，并强调文", "145", "501", "1196", "323", "本会放大图像，而图像会抑制它。Osterroth（2018）进", "步将模因解构为三个部分：（i）最初的言语成分会", "Table 1:Dataset summary,", "", "tabulating the entity", "影响普通用户，（i）有经验的用户的视觉部分，以及", "specific sample counts w.r.t. different domains and", "(i）曲折的结论性第二言语部分。这恢复了视觉模糊，", "splits :", "强调了批判性模因分析的抽象推理。我们旨在通过从多", "个相关目标（LeeandShen2022）进行共享学习来研究", "EXCLAIM的上下文增强", "宣传技术（Dimitrovetal.2021），有害性（Pramanick", "总而言之，我们使用几个单峰和多峰基线对我们建议", "etal.2021b）以及（Sharmaet al.2022a）中的有害目", "的数据集ExHVV，通常涉及单/多模式编码器-解码器", "标以及模因(Sharma.and Pulabaigari.2020;Sharma,", "架构。我们进一步提出了LUMEN，这是一个多模式编", "Pulabaigari,andDas2020)的检测", "码器-解码器框架，它通过多任务学习结合了实体特定", "的角色标签信息和解释生成功能。我们使用多个标准自", "Visual Question Answering (VQA).Antol et al.", "然语言生成（NLG）评估措施来比较这些系统的性能", "（2015）完成了有关VQA的早期杰出工作，该框架具", "以评估其生成质量。我们终于揭示了LUMEN的局限", "有鼓励开放式问题和候选人答案的框架。从那时起，已", "性，同时强调了解决EXCLAIM所面临的挑战。我们", "经观察到多种变化。Antol etal.（2015）通过共同表示", "的贡献总结如下：", "图像和问题对答案进行分类。其他方法则通对过不限于", "co/soft/hard-attention 机制 (Lu et al.2016;Anderson", "1. EXCLAIM: A novel task formulation, soliciting ex-", "etal.2018;Malinowskietal.2018）的注意力类型检查", "planation generation for semantic role labelling in", "跨模式交互，从而有效地学习问题令牌与局部图像区域", "memes.", "之间的显式相关性。值得注意的是，有一系列尝试将常", "2. ExHVV: A multimodal dataset comprising natu-", "识推理纳人 (Zellers et al. 2019;Wu et al. 2016,2017;", "ral language explanations accompanying the sets of", "Marinoetal.2019）。其中许多研究还利用来自外部知", "memes, entities, and semantic roles.", "识库的信息来解决VQA任务。诸如UpDn（Anderson", "3.LUMEN: A novel multimodal, multi-task learning", "et al. 2018) 和 LXMERT (Tan and Bansal 2019) 之类", "framework that facilitates shared feature learning", "的通用模型明确地利用了VQA任务的非线性变换和变", "from semantically related tasks.", "压器，而诸如LMH（Clark，Yatskar，andZettlemoyer", "2019）和SSL（Zhuetal.2021）之类的其他模型则解决", "4. An extensive study of the explanation generation", "了限制VQA性能的关键语言先验问题，尽管增强幅度", "quality, across 18 standard NLG evaluation mea-", "很小", "sures.", "Cross-ModalAssociation：由干多模态数据的通入增", "2Related Work", "加，跨模态关联最近受到了极大的关注。对于跨模态检", "本节简要讨论了有关模因分析的相关研究，这些研究主", "索和视觉语言预训练，跨模态相似性的精确测量势在必", "要试图捕捉模因的情感方面，例如敌意和情绪，以及有", "行。传统技术主要使用模态的串联，然后自我关注学习", "关模因的其他值得注意的研究", "跨模态比对（Wangetal.2016）。在以对象为中心的方", "法之后，Zeng，Zhang，andLi（2021）和Lietal.（2020)", "MemeAnalysis：最近组织了儿个共享任务，最近", "提出了一种多粒度对齐方法，该方法捕获了多个对象的", "个任务是检测模因（Sharmaetal.2022c）中的英雄，恶", "视觉概念之间的关系，同时将它们与文本和其他元数据", "棍和受害者实体。其他包括巨魔模因分类（Suryawanshi", "对齐。另一方面，几种方法还学习了图像和文本的粗粒", "andChakravarthi2021）和通过其情感，类型和强度预", "度特征之间的对齐方式，同时在其方法（Huang etal.", "测（Sharmaetal.2020）进行的模因情感分析。值得注", "2020；Kim，Son，andKim2021）中忽略了对象检测。后", "意的是，由Kiela etal.（2020b）引人并随后由Zhou.", "来的方法尝试了多种方法，包括从视觉效果中进行跨模", "Chen，andYang（2021）跟进的仇恨模因检测引起了极", "式语义学习和对比损失公式（Yuanetal.2021;Jiaetal.", "大的兴趣，并提出了各种解决方案。其中一些工作包括", "2021; Radford et al. 2021)", "微调VisualBERT(Liet al.2019)和UNITER（Chen", "尽管通常涵盖了跨模态和与模因相关的应用程序，但", "etal.2020），以及使用基于Detectron的表示（Veli-", "模因的几个细粒度方面仍有待研究。在这里，我们试图", "ogluandRose2020;Lippeetal.2020）进行仇恨模因检", "解决一个这样的新任务：EXCLAIM", "测。另一方面，有系统的努力涉及使用变压器（Muen", "nighoff 2020; Vaswani et al. 2017),ViLBERT, VLP", "3ExHVV:Dataset Curation and", "UNITER (Sandulescu 2020; Lu et al. 2019; Zhou et al.", "Annotation", "2020:Chenetal.2020)和LXMERT(TanandBansal", "2019）进行双流集成的统一和双流编码器。除此之外，", "Dataset：由于EXCLAIM需要具有模因的多模式数", "其他任务还涉及反犹太主义（Chandra etal.2021），", "据集，所引用实体的语义角色标签以及相关的自然语", "言解释，因此我们利用HVVMemes。此数据集最近在", "1The source code for this work can be found at: https:", "CONSTRAINT-2022（Sharmaetal.2022c）上作为共", "//github.com/LCS2-IITD/LUMEN-Explaining-Memes.", "享任务的一部分发布。原始数据集HVVMemes通过将" ]

{ "arxiv_id": "2212.00715", "language": "zh", "page_idx": 2 }

[ [ "0.0972", "0.0682", "0.1365", "0.0707", "0.1356", "0.0833", "0.09644", "0.0808" ], [ "0.2418", "0.07135", "0.3774", "0.07135", "0.3774", "0.0821", "0.2418", "0.0821" ], [ "0.6396", "0.06757", "0.7876", "0.06946", "0.7876", "0.084", "0.6396", "0.0821" ], [ "0.1136", "0.089", "0.1259", "0.089", "0.1259", "0.0985", "0.1136", "0.0985" ], [ "0.1544", "0.0871", "0.4631", "0.0871", "0.4631", "0.09784", "0.1544", "0.09784" ], [ "0.5215", "0.09216", "0.91", "0.09216", "0.91", "0.1067", "0.5215", "0.1067" ], [ "0.1544", "0.09973", "0.2035", "0.09973", "0.2035", "0.10986", "0.1544", "0.10986" ], [ "0.152", "0.1092", "0.4631", "0.1092", "0.4631", "0.1225", "0.152", "0.1225" ], [ "0.5215", "0.1067", "0.91", "0.1067", "0.91", "0.1206", "0.5215", "0.1206" ], [ "0.539", "0.1206", "0.911", "0.1206", "0.911", "0.1351", "0.539", "0.1351" ], [ "0.1095", "0.1111", "0.125", "0.108", "0.1283", "0.1193", "0.1136", "0.1225" ], [ "0.1136", "0.1225", "0.1259", "0.1225", "0.1259", "0.1332", "0.1136", "0.1332" ], [ "0.1528", "0.1218", "0.3987", "0.1218", "0.3987", "0.1351", "0.1528", "0.1351" ], [ "0.524", "0.1225", "0.544", "0.1225", "0.544", "0.1307", "0.524", "0.1307" ], [ "0.5215", "0.1338", "0.911", "0.1338", "0.911", "0.1477", "0.5215", "0.1477" ], [ "0.1569", "0.1401", "0.3823", "0.1401", "0.3823", "0.1509", "0.1569", "0.1509" ], [ "0.1544", "0.1515", "0.4453", "0.1528", "0.4453", "0.1666", "0.1544", "0.1654" ], [ "0.523", "0.1483", "0.91", "0.1483", "0.91", "0.1622", "0.523", "0.1622" ], [ "0.1544", "0.1692", "0.3938", "0.1692", "0.3938", "0.1799", "0.1544", "0.1799" ], [ "0.523", "0.1635", "0.9126", "0.1635", "0.9126", "0.1768", "0.523", "0.1768" ], [ "0.11926", "0.1704", "0.1259", "0.1755", "0.1209", "0.1812", "0.11273", "0.1749" ], [ "0.152", "0.1787", "0.407", "0.1787", "0.407", "0.1925", "0.152", "0.1925" ], [ "0.523", "0.1768", "0.911", "0.1768", "0.911", "0.1901", "0.523", "0.1901" ], [ "0.152", "0.1901", "0.464", "0.1925", "0.464", "0.2058", "0.152", "0.2039" ], [ "0.1095", "0.1818", "0.125", "0.1787", "0.1283", "0.1901", "0.1136", "0.1932" ], [ "0.1136", "0.1932", "0.1259", "0.1932", "0.1259", "0.2026", "0.1136", "0.2026" ], [ "0.523", "0.1901", "0.911", "0.1901", "0.911", "0.2039", "0.523", "0.2039" ], [ "0.1544", "0.2058", "0.2035", "0.2058", "0.2035", "0.2166", "0.1544", "0.2166" ], [ "0.5195", "0.2026", "0.9126", "0.2039", "0.9126", "0.2184", "0.5195", "0.2172" ], [ "0.1177", "0.2153", "0.1259", "0.2216", "0.11926", "0.2267", "0.1111", "0.2203" ], [ "0.1528", "0.2166", "0.4192", "0.2166", "0.4192", "0.2267", "0.1528", "0.2267" ], [ "0.523", "0.2184", "0.6045", "0.2184", "0.6045", "0.2299", "0.523", "0.2299" ], [ "0.1095", "0.2299", "0.1209", "0.2247", "0.1283", "0.2343", "0.1168", "0.2393" ], [ "0.1528", "0.226", "0.4631", "0.226", "0.4631", "0.2393", "0.1528", "0.2393" ], [ "0.1544", "0.2393", "0.4395", "0.2393", "0.4395", "0.25", "0.1544", "0.25" ], [ "0.5376", "0.2343", "0.914", "0.2343", "0.914", "0.2474", "0.5376", "0.2474" ], [ "0.1136", "0.2507", "0.125", "0.2507", "0.125", "0.26", "0.1136", "0.26" ], [ "0.1544", "0.2507", "0.437", "0.2507", "0.437", "0.2615", "0.1544", "0.2615" ], [ "0.5215", "0.2474", "0.911", "0.2474", "0.911", "0.2615", "0.5215", "0.2615" ], [ "0.152", "0.2615", "0.4617", "0.2615", "0.4617", "0.2747", "0.152", "0.2747" ], [ "0.523", "0.2615", "0.911", "0.2615", "0.911", "0.2747", "0.523", "0.2747" ], [ "0.5215", "0.2734", "0.911", "0.2747", "0.911", "0.289", "0.5215", "0.2886" ], [ "0.11273", "0.262", "0.1274", "0.262", "0.1274", "0.2715", "0.11273", "0.2715" ], [ "0.08905", "0.2886", "0.4739", "0.2886", "0.4739", "0.3018", "0.08905", "0.3018" ], [ "0.5215", "0.2874", "0.9126", "0.2886", "0.9126", "0.303", "0.5215", "0.3018" ], [ "0.523", "0.3018", "0.911", "0.3018", "0.911", "0.315", "0.523", "0.315" ], [ "0.5195", "0.315", "0.9126", "0.3164", "0.9126", "0.3308", "0.5195", "0.33" ], [ "0.5215", "0.33", "0.911", "0.33", "0.911", "0.3435", "0.5215", "0.3435" ], [ "0.08905", "0.3384", "0.4722", "0.3384", "0.4722", "0.3484", "0.08905", "0.3484" ], [ "0.0866", "0.3484", "0.4763", "0.3499", "0.4763", "0.3643", "0.0866", "0.363" ], [ "0.524", "0.3447", "0.91", "0.3447", "0.91", "0.358", "0.524", "0.358" ], [ "0.0907", "0.363", "0.4756", "0.363", "0.4756", "0.377", "0.0907", "0.377" ], [ "0.523", "0.358", "0.7817", "0.358", "0.7817", "0.3718", "0.523", "0.3718" ], [ "0.08905", "0.379", "0.4763", "0.379", "0.4763", "0.3926", "0.08905", "0.3926" ], [ "0.5215", "0.3914", "0.568", "0.3914", "0.568", "0.4016", "0.5215", "0.4016" ], [ "0.564", "0.3901", "0.716", "0.3901", "0.716", "0.404", "0.564", "0.404" ], [ "0.0866", "0.404", "0.478", "0.4048", "0.478", "0.4204", "0.0866", "0.4192" ], [ "0.08905", "0.3926", "0.4763", "0.3926", "0.4763", "0.406", "0.08905", "0.406" ], [ "0.08905", "0.4192", "0.4763", "0.4192", "0.4763", "0.433", "0.08905", "0.433" ], [ "0.523", "0.414", "0.911", "0.414", "0.911", "0.428", "0.523", "0.428" ], [ "0.523", "0.428", "0.9004", "0.428", "0.9004", "0.4414", "0.523", "0.4414" ], [ "0.892", "0.4312", "0.9062", "0.428", "0.91", "0.4375", "0.8965", "0.44" ], [ "0.5215", "0.4414", "0.9126", "0.4414", "0.9126", "0.455", "0.5215", "0.455" ], [ "0.08826", "0.433", "0.4763", "0.433", "0.4763", "0.4463", "0.08826", "0.4463" ], [ "0.08826", "0.4458", "0.478", "0.4463", "0.478", "0.461", "0.08826", "0.4602" ], [ "0.08905", "0.461", "0.4763", "0.461", "0.4763", "0.4749", "0.08905", "0.4749" ], [ "0.5215", "0.455", "0.9126", "0.455", "0.9126", "0.4697", "0.5215", "0.4697" ], [ "0.08826", "0.4749", "0.4763", "0.4749", "0.4763", "0.488", "0.08826", "0.488" ], [ "0.523", "0.4697", "0.911", "0.4697", "0.911", "0.483", "0.523", "0.483" ], [ "0.5215", "0.483", "0.911", "0.483", "0.911", "0.4968", "0.5215", "0.4968" ], [ "0.08905", "0.4893", "0.478", "0.4893", "0.478", "0.5024", "0.08905", "0.5024" ], [ "0.08826", "0.5024", "0.478", "0.5024", "0.478", "0.5166", "0.08826", "0.5166" ], [ "0.523", "0.4976", "0.9126", "0.4976", "0.9126", "0.511", "0.523", "0.511" ], [ "0.08826", "0.5166", "0.478", "0.5166", "0.478", "0.53", "0.08826", "0.53" ], [ "0.5254", "0.511", "0.9126", "0.511", "0.9126", "0.5244", "0.5254", "0.5244" ], [ "0.523", "0.5244", "0.911", "0.5244", "0.911", "0.5386", "0.523", "0.5386" ], [ "0.08826", "0.531", "0.3823", "0.531", "0.3823", "0.5444", "0.08826", "0.5444" ], [ "0.524", "0.5405", "0.91", "0.5405", "0.91", "0.5503", "0.524", "0.5503" ], [ "0.523", "0.552", "0.91", "0.552", "0.91", "0.565", "0.523", "0.565" ], [ "0.09644", "0.5625", "0.4763", "0.5625", "0.4763", "0.5757", "0.09644", "0.5757" ], [ "0.5215", "0.5664", "0.911", "0.5664", "0.911", "0.58", "0.5215", "0.58" ], [ "0.08826", "0.5757", "0.4763", "0.5757", "0.4763", "0.5903", "0.08826", "0.5903" ], [ "0.5215", "0.579", "0.9126", "0.58", "0.9126", "0.5947", "0.5215", "0.5933" ], [ "0.08826", "0.589", "0.478", "0.589", "0.478", "0.603", "0.08826", "0.603" ], [ "0.5215", "0.5933", "0.911", "0.5933", "0.911", "0.607", "0.5215", "0.607" ], [ "0.08826", "0.603", "0.4763", "0.603", "0.4763", "0.616", "0.08826", "0.616" ], [ "0.5215", "0.608", "0.91", "0.608", "0.91", "0.622", "0.5215", "0.622" ], [ "0.08826", "0.617", "0.478", "0.617", "0.478", "0.631", "0.08826", "0.631" ], [ "0.5186", "0.619", "0.777", "0.616", "0.777", "0.6353", "0.5186", "0.638" ], [ "0.08826", "0.631", "0.4763", "0.631", "0.4763", "0.6445", "0.08826", "0.6445" ], [ "0.08905", "0.6465", "0.4756", "0.6465", "0.4756", "0.657", "0.08905", "0.657" ], [ "0.5195", "0.653", "0.7036", "0.653", "0.7036", "0.6665", "0.5195", "0.6665" ], [ "0.08826", "0.659", "0.4763", "0.659", "0.4763", "0.6724", "0.08826", "0.6724" ], [ "0.08826", "0.6724", "0.4756", "0.6724", "0.4756", "0.686", "0.08826", "0.686" ], [ "0.523", "0.678", "0.911", "0.678", "0.911", "0.6914", "0.523", "0.6914" ], [ "0.08905", "0.688", "0.4756", "0.688", "0.4756", "0.6987", "0.08905", "0.6987" ], [ "0.523", "0.6914", "0.589", "0.6914", "0.589", "0.7026", "0.523", "0.7026" ], [ "0.584", "0.6895", "0.9067", "0.6895", "0.9067", "0.704", "0.584", "0.704" ], [ "0.08826", "0.7007", "0.478", "0.7007", "0.478", "0.714", "0.08826", "0.714" ], [ "0.5195", "0.702", "0.911", "0.7007", "0.911", "0.7153", "0.5195", "0.7163" ], [ "0.08826", "0.714", "0.3506", "0.714", "0.3506", "0.728", "0.08826", "0.728" ], [ "0.523", "0.7183", "0.91", "0.7183", "0.91", "0.729", "0.523", "0.729" ], [ "0.1054", "0.733", "0.4763", "0.733", "0.4763", "0.7437", "0.1054", "0.7437" ], [ "0.523", "0.729", "0.911", "0.729", "0.911", "0.742", "0.523", "0.742" ], [ "0.08826", "0.7456", "0.4763", "0.7456", "0.4763", "0.759", "0.08826", "0.759" ], [ "0.5195", "0.741", "0.911", "0.7407", "0.911", "0.755", "0.5195", "0.756" ], [ "0.08826", "0.759", "0.406", "0.759", "0.406", "0.773", "0.08826", "0.773" ], [ "0.523", "0.755", "0.911", "0.755", "0.911", "0.768", "0.523", "0.768" ], [ "0.5215", "0.768", "0.9126", "0.768", "0.9126", "0.782", "0.5215", "0.782" ], [ "0.08905", "0.785", "0.4763", "0.785", "0.4763", "0.799", "0.08905", "0.799" ], [ "0.524", "0.782", "0.8823", "0.782", "0.8823", "0.7954", "0.524", "0.7954" ], [ "0.8774", "0.785", "0.911", "0.785", "0.911", "0.7944", "0.8774", "0.7944" ], [ "0.10455", "0.799", "0.3962", "0.799", "0.3962", "0.8125", "0.10455", "0.8125" ], [ "0.524", "0.7935", "0.9126", "0.7935", "0.9126", "0.8066", "0.524", "0.8066" ], [ "0.523", "0.806", "0.911", "0.8066", "0.911", "0.8213", "0.523", "0.821" ], [ "0.08905", "0.8228", "0.478", "0.8228", "0.478", "0.836", "0.08905", "0.836" ], [ "0.523", "0.821", "0.911", "0.821", "0.911", "0.834", "0.523", "0.834" ], [ "0.1054", "0.836", "0.4192", "0.836", "0.4192", "0.8496", "0.1054", "0.8496" ], [ "0.523", "0.8335", "0.9004", "0.8354", "0.9004", "0.8486", "0.523", "0.8467" ], [ "0.8945", "0.837", "0.911", "0.837", "0.911", "0.845", "0.8945", "0.845" ], [ "0.5215", "0.8477", "0.9126", "0.8486", "0.9126", "0.863", "0.5215", "0.8623" ], [ "0.08905", "0.8613", "0.4763", "0.8613", "0.4763", "0.875", "0.08905", "0.875" ], [ "0.5376", "0.86", "0.9126", "0.8613", "0.9126", "0.8755", "0.5376", "0.875" ], [ "0.5254", "0.866", "0.5376", "0.866", "0.5376", "0.8726", "0.5254", "0.8726" ], [ "0.536", "0.866", "0.5435", "0.866", "0.5435", "0.8706", "0.536", "0.8706" ], [ "0.08826", "0.8755", "0.3064", "0.8755", "0.3064", "0.8896", "0.08826", "0.8896" ], [ "0.5195", "0.8735", "0.778", "0.8706", "0.778", "0.888", "0.5195", "0.8916" ], [ "0.7754", "0.8755", "0.911", "0.8755", "0.911", "0.8896", "0.7754", "0.8896" ] ]

[ "S.No.", "Annotation Guidelines", "4Methodology", "1", "Explanations should consider meme author's per-", "本节介绍LUMEN的详细信息，LUMEN是我们提出", "spective", "Explanations should emphasize meme's content only", "的多模式，多任务学习框架，用于解决EXCLAIM（c。", "Fig。2）。正如在1节中激励EXCLAIM时举例说", "2", "3", "Annotations should be narrated/written-", "F", "明的那样，角色标签和相应的解释可能是互补的；我们", ".Using reported speech for opinions.", ". In the simple present form when stating facts", "假设预测一个应该有助于推断另一个。LUMEN包含四", "Explanations should belexicallydiverse", "个关键组件：（a）视觉识别，（b）实体语义，（c）解", "N", "Cases labeled as others should be skipped.", "释生成和（d）角色标签预测。后三者还有助于对子任", "Entity should constitute as explanation's primary", "5", "6", "务（ST）进行建模-序列分类（ST-1)，解释生成（ST-2）", "subject..", "和角色标签预测（ST-3），作为多任务框架的一部分在", "", "Obvious erroneous labels should be rectified", "LUMEN中", "8", "Explanations may follow a standardized format: [En-", "tity] [Action/Abstract Word/Phrases] [Description]", "视觉识别模块通过基于OCR的模因的文本提取，字", "9", "Each explanation should refer to a single entity", "幕生成和视觉特征提取做出贡献。实体语义模块联合", "Explanations may borrow facts from the OCR text.", "编码模因的嵌入文本和候选实体，同时优化序列分类", "任务。另一方面，解释生成模块首先对特殊配置的提示", "10", "Table 2: Prescribed guidelines for ExHVV annotation", "以及模因文本及其标题进行编码，然后对编码的隐藏", "表示的自回归解码进行优化。最后，前三个模块的输出", "通过其多模式融合被用于角色标签预测任务。有效地，", "LUMEN通过从三个相应目标优化联合损失公式来联", "meme实体与角色相关联来注释它们的内涵标签：英雄", "恶棍，受害者或其他跨COVID-19和美国政治领域的", "合训练序列分类，角色预测和解释生成。我们将在以下", "人。我们通过对模因中存在的实体提供的内涵标签进行", "小节中详细说明这些组件中的每一个", "自然语言解释来增强HVVMemes，方法是重新注释它", "4.1", "Visual Recognition", "ExHVV的摘要。我们只考虑英雄，恶棍和受害者，因", "们，并偶尔在明显的地方纠正角色标签。表1显示了", "为其他情况要么本质上是模棱两可的，要么会寻求可", "由于模因主要以视觉形式包含所有内容，因此需要首", "先利用某种视觉处理机制来挖掘每个输入特征提示。", "头", "此，我们首先提取模因的嵌人文本，并使用Google的", "能超出范围的解释。r。t。这项工作。这导致ExHVV", "中类别英雄的实体特定样本计数增加了78，小人增", "加了616，受害者增加了159。Annotatorsarerequested-", "OCRVisionAPI从给定的模因图像中执行基于OCR的", "toexplainwhyanentitycouldbeahero,villain,orvictims.这", "嵌人文本提取。我们完全考虑模因的文本，并在预处理", "时丢弃任何固有的换行符。模因的图像形式的视觉背景", "导致了总共4680标记的样品（c。f。表1）。由于注释", "者的解释风格不同，因此在词汇和length方面，注释", "可能无法全面捕捉模因的预期信息。然而，它确实提供", "者的解释合理地有所不同。注释中还强调了任务的主观", "了一种协调的理解，可以获得对模因信息的完整观点。", "Blaier，Malkiel，andWolf（2021）的证据表明，利用模", "性，因为模因通常会导致儿种有效的解释", "因标题可以改善仇恨模因的识别结果。此外：模因中存", "在的人员，位置和实体等其他提示也有助于下游任务。", "AnnotationProcess．专门为获得解释而训练了两个", "因此，我们将这些辅助信息与OCR文本一起使用。对", "注释器W。r。t。EXCLAIM。他们被规定为标准注释", "于图像字幕，我们使用最近发布的OFA模型（Wang", "指南(c。f。表2），这是为了帮助他们而起草的。r。t。", "etal.2022）。最后，我们使用基于ViT的模型对meme", "任务要求和歧义解决。注释指南明确强调将模因作者的", "视觉对象进行编码，并从其最后一个隐藏层中提取合并", "观点视为标准化参考。还向注释者发布了具有某些用例", "输出，以获得视觉嵌人VhER768", "的动词列表，以保持注释之间的一致性。两位注释者都", "听取了简报。r。t。注释指南由整合者提供，其任务是", "4.2EntitySemantics", "用125个实体注释20个随机模因的初始公共集，以评", "估和简化注释模式。然后，考虑到来自整合程序的反馈", "如前所述，模因中的言语内容突出地构成了其语义底色", "对测试集进行注释。然后，其余的模因在注释者之间平", "的特征。", "为了从嵌人在模因中的文本内容中捕获语义", "均分配，合并者裁定模棱两可的情况。这些方面被容纳", "我们根据经验将DeBERTa（Heetal.202i）的最后一个", "在最终数据集中，以确保一致的注释", "隐藏层的输出指定为我们的首选。除了展示其在各种任", "通过明确捕获所提出解释的语义有效性和多样性，通", "务中的优势，即MNLISQuAD，RACE等。，DeBERTa", "过人工评估来评估注释质量。此评估中提出的问题以及", "在检测模因（Kun，Bankoti，andKiskovski2022）中实", "可能的答案选择（以及分配的相应分数）为：", "体的语义角色标签方面表现出了卓越的性能。我们将其", "deberta-v2-xlarge预训练变体（#层：24，H：1536和", ": Does the annotator provide a valid explanation?", "#参数：900M）用于与后者类似的目标，除了将", "“其", "No/Partially/Yes (0/1/2), no label (-1)?", "他”类别视为EXCLAIM尽式中的自标语义角色。在", "使用最后一个隐藏层输出时，我们还将此设置微调到序", ": Are the given pair of annotations diverse or not?", "列分类目标。此模型的输人配置为一对序列：[CLS]A", "No/Yes (0/1), invalid (no) annotation (-1)?", "[SEP]B[SEP]，其中A对应于OCR提取的文本，", "B", "是候选实体。我们在LUMENw中联合微调此模型。r。", "从该评估中获得的平均归一化效度和多样性分数分别", "三个自标标签，英雄，恶棍和受害者，并获得嵌人实", "10", "", "为两个注释者的0.81和0.84。", "体语义的均值池表示：ThER1536", "，以及给定序列的" ]

{ "arxiv_id": "2212.00715", "language": "zh", "page_idx": 3 }

[ [ "0.3005", "0.0873", "0.7", "0.0873", "0.7", "0.10156", "0.3005", "0.10156" ], [ "0.2527", "0.111", "0.7524", "0.11224", "0.7524", "0.133", "0.2527", "0.1318" ], [ "0.272", "0.1318", "0.7314", "0.133", "0.7314", "0.1478", "0.272", "0.1467" ], [ "0.2317", "0.1484", "0.769", "0.1497", "0.769", "0.1644", "0.2317", "0.1633" ], [ "0.2477", "0.165", "0.7515", "0.1663", "0.7515", "0.1812", "0.2477", "0.1799" ], [ "0.2418", "0.1818", "0.7583", "0.1829", "0.7583", "0.1978", "0.2418", "0.1965" ], [ "0.3762", "0.1996", "0.629", "0.1996", "0.629", "0.2131", "0.3762", "0.2131" ], [ "0.2351", "0.2162", "0.482", "0.2162", "0.482", "0.231", "0.2351", "0.231" ], [ "0.525", "0.2162", "0.769", "0.2162", "0.769", "0.231", "0.525", "0.231" ], [ "0.2627", "0.2542", "0.3418", "0.2542", "0.3418", "0.2683", "0.2627", "0.2683" ], [ "0.532", "0.2546", "0.881", "0.2546", "0.881", "0.2683", "0.532", "0.2683" ], [ "0.513", "0.2708", "0.881", "0.2708", "0.881", "0.284", "0.513", "0.284" ], [ "0.147", "0.279", "0.4575", "0.279", "0.4575", "0.2935", "0.147", "0.2935" ], [ "0.513", "0.2852", "0.879", "0.2852", "0.879", "0.2986", "0.513", "0.2986" ], [ "0.147", "0.2935", "0.4575", "0.2935", "0.4575", "0.308", "0.147", "0.308" ], [ "0.147", "0.308", "0.4575", "0.308", "0.4575", "0.3218", "0.147", "0.3218" ], [ "0.513", "0.3027", "0.881", "0.3027", "0.881", "0.3164", "0.513", "0.3164" ], [ "0.1453", "0.3218", "0.4568", "0.3218", "0.4568", "0.3362", "0.1453", "0.3362" ], [ "0.513", "0.3196", "0.8564", "0.3196", "0.8564", "0.3333", "0.513", "0.3333" ], [ "0.1453", "0.3362", "0.4575", "0.335", "0.4575", "0.3486", "0.1453", "0.3499" ], [ "0.511", "0.3342", "0.879", "0.3342", "0.879", "0.3474", "0.511", "0.3474" ], [ "0.1436", "0.3499", "0.4568", "0.3508", "0.4568", "0.3652", "0.1436", "0.364" ], [ "0.5146", "0.3516", "0.881", "0.3516", "0.881", "0.3652", "0.5146", "0.3652" ], [ "0.147", "0.3665", "0.4568", "0.3652", "0.4568", "0.3772", "0.147", "0.3777" ], [ "0.5146", "0.3677", "0.882", "0.3677", "0.882", "0.3813", "0.5146", "0.3813" ], [ "0.1486", "0.3813", "0.4568", "0.3813", "0.4568", "0.3914", "0.1486", "0.3914" ], [ "0.513", "0.3843", "0.881", "0.3843", "0.881", "0.398", "0.513", "0.398" ], [ "0.147", "0.3938", "0.4568", "0.3938", "0.4568", "0.4067", "0.147", "0.4067" ], [ "0.5146", "0.3997", "0.8774", "0.3997", "0.8774", "0.4133", "0.5146", "0.4133" ], [ "0.1436", "0.4062", "0.4575", "0.4067", "0.4575", "0.4204", "0.1436", "0.4192" ], [ "0.511", "0.4146", "0.881", "0.415", "0.881", "0.43", "0.511", "0.4287" ], [ "0.1453", "0.4216", "0.4575", "0.4216", "0.4575", "0.4353", "0.1453", "0.4353" ], [ "0.147", "0.4353", "0.4575", "0.4353", "0.4575", "0.449", "0.147", "0.449" ], [ "0.513", "0.4316", "0.879", "0.4316", "0.879", "0.4453", "0.513", "0.4453" ], [ "0.1486", "0.452", "0.455", "0.452", "0.455", "0.462", "0.1486", "0.462" ], [ "0.5146", "0.4478", "0.879", "0.4478", "0.879", "0.4614", "0.5146", "0.4614" ], [ "0.1453", "0.4644", "0.4568", "0.4644", "0.4568", "0.478", "0.1453", "0.478" ], [ "0.513", "0.4631", "0.879", "0.4631", "0.879", "0.4768", "0.513", "0.4768" ], [ "0.1453", "0.4785", "0.455", "0.4797", "0.455", "0.4912", "0.1453", "0.4905" ], [ "0.5146", "0.4797", "0.882", "0.4797", "0.882", "0.4934", "0.5146", "0.4934" ], [ "0.1486", "0.4946", "0.4534", "0.4946", "0.4534", "0.505", "0.1486", "0.505" ], [ "0.5146", "0.4963", "0.882", "0.4963", "0.882", "0.5103", "0.5146", "0.5103" ], [ "0.147", "0.509", "0.455", "0.509", "0.455", "0.5195", "0.147", "0.5195" ], [ "0.5146", "0.5127", "0.881", "0.5127", "0.881", "0.5254", "0.5146", "0.5254" ], [ "0.1453", "0.521", "0.4575", "0.5215", "0.4575", "0.535", "0.1453", "0.5337" ], [ "0.513", "0.528", "0.881", "0.528", "0.881", "0.5415", "0.513", "0.5415" ], [ "0.1453", "0.5337", "0.4568", "0.5337", "0.4568", "0.5474", "0.1453", "0.5474" ], [ "0.511", "0.5435", "0.879", "0.5435", "0.879", "0.557", "0.511", "0.557" ], [ "0.1486", "0.552", "0.4568", "0.552", "0.4568", "0.5625", "0.1486", "0.5625" ], [ "0.511", "0.5586", "0.882", "0.56", "0.882", "0.5747", "0.511", "0.5737" ], [ "0.1486", "0.5654", "0.4568", "0.5654", "0.4568", "0.576", "0.1486", "0.576" ], [ "0.1453", "0.5776", "0.4575", "0.5776", "0.4575", "0.5913", "0.1453", "0.5913" ], [ "0.513", "0.5767", "0.7197", "0.5767", "0.7197", "0.5903", "0.513", "0.5903" ], [ "0.147", "0.5933", "0.4534", "0.5933", "0.4534", "0.604", "0.147", "0.604" ], [ "0.529", "0.593", "0.882", "0.5933", "0.882", "0.608", "0.529", "0.607" ], [ "0.147", "0.6055", "0.4575", "0.6055", "0.4575", "0.619", "0.147", "0.619" ], [ "0.513", "0.61", "0.882", "0.61", "0.882", "0.6235", "0.513", "0.6235" ], [ "0.147", "0.6216", "0.4568", "0.6206", "0.4568", "0.632", "0.147", "0.633" ], [ "0.5283", "0.6245", "0.881", "0.626", "0.881", "0.64", "0.5283", "0.6396" ], [ "0.1486", "0.6357", "0.4568", "0.6357", "0.4568", "0.6465", "0.1486", "0.6465" ], [ "0.5146", "0.6426", "0.881", "0.6426", "0.881", "0.6562", "0.5146", "0.6562" ], [ "0.147", "0.651", "0.4534", "0.651", "0.4534", "0.661", "0.147", "0.661" ], [ "0.147", "0.662", "0.4575", "0.662", "0.4575", "0.676", "0.147", "0.676" ], [ "0.5146", "0.659", "0.881", "0.659", "0.881", "0.6733", "0.5146", "0.6733" ], [ "0.1453", "0.6777", "0.44", "0.6777", "0.44", "0.6914", "0.1453", "0.6914" ], [ "0.513", "0.675", "0.881", "0.675", "0.881", "0.6885", "0.513", "0.6885" ], [ "0.511", "0.6904", "0.8774", "0.6904", "0.8774", "0.7036", "0.511", "0.7036" ], [ "0.11755", "0.706", "0.1503", "0.706", "0.1503", "0.7144", "0.11755", "0.7144" ], [ "0.1436", "0.7036", "0.2627", "0.7036", "0.2627", "0.7173", "0.1436", "0.7173" ], [ "0.513", "0.707", "0.881", "0.707", "0.881", "0.72", "0.513", "0.72" ], [ "0.11755", "0.727", "0.4844", "0.7256", "0.4844", "0.7407", "0.11755", "0.741" ], [ "0.511", "0.7236", "0.881", "0.7236", "0.881", "0.737", "0.511", "0.737" ], [ "0.513", "0.7383", "0.881", "0.7383", "0.881", "0.753", "0.513", "0.753" ], [ "0.11755", "0.7446", "0.4836", "0.7446", "0.4836", "0.755", "0.11755", "0.755" ], [ "0.1184", "0.7603", "0.486", "0.7603", "0.486", "0.774", "0.1184", "0.774" ], [ "0.513", "0.755", "0.881", "0.755", "0.881", "0.7686", "0.513", "0.7686" ], [ "0.511", "0.7695", "0.879", "0.7695", "0.879", "0.783", "0.511", "0.783" ], [ "0.1184", "0.7754", "0.486", "0.7754", "0.486", "0.789", "0.1184", "0.789" ], [ "0.11755", "0.792", "0.486", "0.7915", "0.486", "0.8057", "0.11755", "0.807" ], [ "0.5146", "0.788", "0.879", "0.788", "0.879", "0.799", "0.5146", "0.799" ], [ "0.513", "0.8027", "0.881", "0.8027", "0.881", "0.8164", "0.513", "0.8164" ], [ "0.1184", "0.8086", "0.4844", "0.8086", "0.4844", "0.8223", "0.1184", "0.8223" ], [ "0.12006", "0.8247", "0.4844", "0.8247", "0.4844", "0.8384", "0.12006", "0.8384" ], [ "0.5146", "0.821", "0.879", "0.821", "0.879", "0.8306", "0.5146", "0.8306" ], [ "0.511", "0.835", "0.869", "0.835", "0.869", "0.8486", "0.511", "0.8486" ], [ "0.1184", "0.8413", "0.4844", "0.8413", "0.4844", "0.855", "0.1184", "0.855" ], [ "0.11755", "0.8555", "0.4844", "0.857", "0.4844", "0.8716", "0.11755", "0.8706" ], [ "0.517", "0.857", "0.881", "0.857", "0.881", "0.8706", "0.517", "0.8706" ], [ "0.11584", "0.8706", "0.3594", "0.8726", "0.3594", "0.887", "0.11584", "0.8853" ], [ "0.513", "0.8726", "0.634", "0.8726", "0.634", "0.8857", "0.513", "0.8857" ], [ "0.147", "0.8936", "0.4878", "0.8936", "0.4878", "0.9067", "0.147", "0.9067" ], [ "0.51", "0.8896", "0.881", "0.8896", "0.881", "0.903", "0.51", "0.903" ], [ "0.11755", "0.905", "0.315", "0.9062", "0.315", "0.9194", "0.11755", "0.9185" ], [ "0.511", "0.904", "0.881", "0.905", "0.881", "0.9194", "0.511", "0.9185" ] ]

[ "TADA：变压器的高效任务无关域适配", "Chia-Chien Hungl, 2, 3*, Lukas Lange', Jannik Strotgen3, 4", "'NEC Laboratories Europe GmbH, Heidelberg, Germany", "2Data and Web Science Group, University of Mannheim, Germany", "3Bosch Center for Artificial Intelligence, Renningen, Germany", "4Karlsruhe University of Applied Sciences, Karlsruhe, Germany", "[email protected]", "[email protected]", "[email protected]", "Abstract", "为了传授有用的领域知识，出现了两种主要", "的利用变压器的领域自适应方法：（1)E从头开", "Intermediate training of pre-trained trans-", "始进行大规模预训练E(Beltagyetal.,2019:Wu", "former-based language models on domain-", "specific data leads to substantial gains for", "etal.,2020)依赖于在预训练期间结合了各种自", "downstream tasks.To increase efficiency", "我监督目标的大型特定领域语料库。但是，", "and prevent catastrophic forgetting alleviated", "泛的培训过程在时间和资源上都是低效的，因", "from full domain-adaptive pre-training, ap", "为它需要大量（未）标记的领域专用语料库和", "proaches such as adapters have been devel", "大量计算能力。（2)E域自适应中间预训练E", "oped.However,theserequire additionalpa", "(Gururanganetal.，2020）被认为是更轻的重量，", "rameters for each layer, and are criticized for", "因为它只需要少量的域内数据和更少的纪元从", "their limited expressiveness.In this work,", "以前的检查点在PTLM上持续训练。然而，完", "we introduce TADA, a novel task-agnostic do-", "main adaptation method which is modular,", "全预训练模型（i。e。，更新所有PTLM参数）", "parameter-efficient,andthus,data-efficient.", "可能会导致灾难性的遗忘和干扰（McCloskey", "Within TADA, we retrain the embeddings", "and Cohen,1989;Houlsbyet al.,2019），特别是", "to learn domain-aware input representations", "对于更长的自适应选代。为了克服这些限制，", "and tokenizers for the transformerencoder", "已经引人了诸如适配器E(Rebuffietal.，2017;", "while freezing all other parameters of the", "Houlsby et al.,2019)和稀疏微调E(Guo etal.,", "model. Then, task-specific fine-tuning is per.", "2021;BenZakenetal.，2022）之类的替代方案。", "formed. We further conduct experiments with", "然而，这些方法仍然是参数和时间效率低下", "meta-embeddings and newlyintroduced meta-", "的，因为它们要么添加额外的参数，要么需要", "tokenizers,resultingin onemodelpertaskin", "multi-domain use cases. Our broad evaluation", "复杂的训练步骤和/或模型", "in 4 downstream tasks for 14 domains across", "在这项工作中，我们提出了Task-一个诺斯", "single- and multi-domain setups and high- and", "替Domain一个变压器的daptation（TADA），", "low-resource scenarios reveals that TADA is", "个新颖的领域专业化框架。如图E1所示，它", "an effective and efficient alternative to full", "由两个步骤组成：（1）我们对预训练的基于变", "domain-adaptivepre-training and adaptersfoi", "domain adaptation, while not introducing addi", "压器的语言模型进行中间训练（e。g。，BERT)", "tional parameters or complex training steps", "在未标记的特定于领域的文本语料库上，以便", "将领域知识注入到转换器中。在这里，我们固", "1", "Introduction", "定编码器的参数权重，同时只更新嵌人的权重", "预训练语言模型(Radfordetal.,2018;Devlin", "(i。e。，基于嵌入的域自适应预训练)。因此,", "我们从原始PTLM获得具有共享编码器的每", "etal.2019）利l用转换器（Vaswanietal2017）", "已成为在各种自然语言处理（NLP）任务中取", "个域的域专用嵌人，而无需添加用于域自适应", "的其他参数。（2）然后可以针对单域或多域场", "得令人瞩目的成就的关键技术。然而，这些预", "先训练的基于变压器的语言模型（PTLM）是", "景（Langeetal.2021b）中的下游任务对获得的", "域专用嵌入以及编码器进行微调，其中后者是", "在大规模和异构的语料库上训练的，重点是泛", "化性，而没有解决特定领域的问题。在实践中", "使用元款入Fl(CoatesandBollegala.2018:Kiela", "etal.,2018）和一种新颖的元标记方法进行的。", "缺乏此类域相关信息会严重损害下游应用程序", "的性能，如大量研究（i。a。，ZhuandGoldberg，", "Contributions我们通过以下贡献推进了领", "2009; Ruder and Plank, 2018; ?)", "域专业化领域：", "*Research work conducted during internship at Bosch", "（i）我们提出了一种模块化，参数高效且与任", "Center for Artificial Intelligence", "务无关的域自适应方法（TADA），而无需为" ]

{ "arxiv_id": "2305.12717", "language": "zh", "page_idx": 1 }