File size: 137,811 Bytes
837b615 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 |
{
"id": "2204.13346",
"annotator": "moosa",
"input": [
"\\pdfoutput=1\n",
"\\documentclass[11pt]{article}\n",
"\\usepackage{acl}\n",
"\\usepackage{times}\n",
"\\usepackage{latexsym}\n",
"\\usepackage[T1]{fontenc}\n",
"\\usepackage[utf8]{inputenc}\n",
"\\usepackage{microtype}\n",
"\\usepackage{amsmath}\n",
"\\usepackage{amssymb}\n",
"\\usepackage{xcolor}\n",
"\\usepackage{arydshln}\n",
"\\usepackage{tikz}\n",
"\\usetikzlibrary{arrows, decorations.text, shapes.geometric, positioning, decorations.pathreplacing, calligraphy}\n",
"\\usepackage{pgfplots}\n",
"\\usepackage{pgfmath}\n",
"\\usepackage{pgffor}\n",
"\\pgfplotsset{compat=1.17}\n",
"\\usepackage{subcaption}\n",
"\\usepackage{mathtools}\n",
"\\usepackage{multirow}\n",
"\\usepackage{booktabs}\n",
"\\usepackage{pifont}\n",
"\\usepackage{ifthen}\n",
"\\newcommand{\\cref}{\\colorbox{blue!10}{\\textsc{Ref}}}\n",
"\\newcommand{\\csrc}{\\colorbox{red!10}{\\textsc{Src}}}\n",
"\\newcommand{\\csrcref}{\\colorbox{green!10}{\\textsc{Src+Ref}}}\n",
"\\newcommand{\\ccref}{\\fboxsep0.5ex\\colorbox{blue!10}{\\textsc{Ref}}}\n",
"\\newcommand{\\ccsrc}{\\fboxsep0.5ex\\colorbox{red!10}{\\textsc{Src}}}\n",
"\\newcommand{\\ccsrcref}{\\fboxsep0.5ex\\colorbox{green!10}{\\textsc{Src+Ref}}}\n",
"\\newcommand{\\cmodel}{UniTE}\n",
"\\title{\\cmodel: Unified Translation Evaluation}\n",
"\\author{Yu Wan$^{a,b}$\\thanks{~~Work was done when Yu Wan was interning at DAMO Academy, Alibaba Group.}~~~Dayiheng Liu$^b$\\thanks{~~Dayiheng Liu and Derek F. Wong are co-corresponding authors.}~~~Baosong Yang$^b$~~~Haibo Zhang$^b$~~~Boxing Chen$^b$\\\\\\textbf{Derek F. Wong}$^{b\\dagger}$~~~\\textbf{Lidia S. Chao}$^a$ \\\\\n",
" $^a$NLP$^2$CT Lab,\n",
" University of Macau\\\\\n",
" {\\tt [email protected], \\{derekfw,lidiasc\\}@umac.mo} \\\\\n",
" $^b$Alibaba Group\\\\\n",
" {\\tt \\{liudayiheng.ldyh,yangbaosong.ybs,zhanhui.zhb,}\\\\\n",
" {\\tt boxing.cbx\\}@alibaba-inc.com}}\n",
"\\begin{document}\n",
"\\maketitle\n",
"\\begin{abstract}\n",
"Recent methods, despite their promising results, are specifically designed and optimized on one of them. \n",
"This limits the convenience of these methods, and overlooks the commonalities among tasks.\n",
"In this paper, we propose \\cmodel, which is the first unified framework engaged with abilities to handle all three evaluation tasks.\n",
"Concretely, we propose monotonic regional attention to control the interaction among input segments, and unified pretraining to better adapt multi-task learning. \n",
"We testify our framework on WMT 2019 Metrics and WMT 2020 Quality Estimation benchmarks. Extensive analyses show that our \\textit{single model} can universally surpass various state-of-the-art or winner methods across tasks.\n",
"Both source code and associated models are available at \\href{https://github.com/NLP2CT/UniTE}{https://github.com/NLP2CT/UniTE}.\n",
"\\end{abstract}\n",
"\\section{Introduction}\n",
"Automatically evaluating the translation quality with the given reference segment(s), is of vital importance to identify the performance of Machine Translation (MT) models~\\cite{freitag2020bleu,mathur2020tangled,zhao2020limitations,kocmi2021ship}.\n",
"Based on the input contexts, translation evaluation can be mainly categorized into three classes:\n",
"1) reference-only evaluation (\\cref) approaches like BLEU~\\cite{papineni2002bleu} and BLEURT~\\cite{sellam2020bleurt}, which evaluate the hypothesis by referring the golden reference at target side;\n",
"2) source-only evaluation (\\csrc) methods like YiSi-2~\\cite{lo2019yisi} and TransQuest~\\cite{ranasinghe2020transquest}, which are also referred as quality estimation (QE).\n",
"These methods estimate the quality of the hypothesis based on the source sentence without using references;\n",
"With the help of powerful pretrained language models~\\citep[PLMs,][]{devlin2019bert,conneau2020unsupervised}, model-based approaches (\\textit{e.g.}, BLEURT, TransQuest, and COMET) have shown promising results in recent WMT competitions~\\cite{ma2019results,mathur2020results,freitag2021results,fonseca2019findings,specia2020findings,specia2021findings}.\n",
"Nevertheless, each existing MT evaluation work is usually designed for one specific task, \\textit{e.g.}, BLEURT is only used for \\cref~task and can not support \\csrc~and \\csrcref~tasks.\n",
"Moreover, those approaches preserve the same core -- evaluating the quality of translation by referring to the given segments.\n",
"We believe that it is valuable, as well as feasible, to unify the capabilities of all MT evaluation tasks (\\cref, \\csrc~and \\csrcref) into one model.\n",
"Among the promising advantages are ease of use and improved robustness through knowledge transfer across evaluation tasks.\n",
"To achieve this, two important challenges need to be addressed:\n",
"1) How to design a model framework that can unify all translation evaluation tasks?\n",
"2) How to make the powerful PLMs better adapt to the unified evaluation model?\n",
"In this paper, we propose \\textbf{\\cmodel} - \\textbf{Uni}fied \\textbf{T}ranslation \\textbf{E}valuation, a novel approach which unifies the functionalities of \\cref, \\csrc~and \\csrcref~tasks into one model.\n",
"To solve the first challenge as mentioned above, based on the multilingual PLM, we utilize layerwise coordination which concatenates all input segments into one sequence as the unified input form.\n",
"To further unify the modeling of three evaluation tasks, we propose a novel Monotonic Regional Attention (MRA) strategy, which allows partial semantic flows for a specific evaluation task. For the second challenge, a multi-task learning-based unified pretraining is proposed.\n",
"To be concrete, we collect the high-quality translations and degrade low-quality translations of NMT models as synthetic data.\n",
"Then we propose a novel ranking-based data labeling strategy to provide the training signal.\n",
"Finally, the multilingual PLM is continuously pretrained on synthetic dataset with multi-task learning manner.\n",
"Besides, our proposed models, named \\cmodel-MRA and \\cmodel-UP respectively, can benefit from finetuning with human-annotated data over three tasks at once, not requiring extra task-specific training.\n",
"Experimental results demonstrate the superiority of \\cmodel.\n",
"Compared to various strong baseline systems on each task, \\cmodel, which unifies \\cref, \\csrc~and \\csrcref~tasks into one \\textit{single model}, achieves consistently absolute improvements of Kendall's $\\tau$ correlations at 1.1, 2.3 and 1.1 scores on English-targeted translation directions of WMT 2019 Metric Shared task~\\cite{fonseca2019findings}, respectively.\n",
"Meanwhile, after introducing multilingual-targeted support for our unified pretraining strategy, a single model named \\cmodel-MUP also gives dominant results against existing methods on non-English-targeted translation evaluation tasks. \n",
"Furthermore, our method can also achieve competitive results over WMT 2020 QE task compared with the winner submission~\\cite{ranasinghe2020transquest}.\n",
"Ablation studies reveal that, the proposed MRA and unified pretraining strategies are both important for model performance, making the model preserve the outstanding performance and multi-task transferability concurrently.\n",
"\\section{Related Work}\n",
"In this section, we briefly introduce the three directions of translation evaluation.\n",
"\\cref~assesses the translation quality via comparing the translation candidate and the given reference. \n",
"In this setting, the two inputs are written in the same language, thus being easily applied in most of the metric tasks. \n",
"In the early stages, statistical methods are dominant solutions due to their strengths in wide language support and intuitive design. \n",
"These methods measure the surface text similarity for a range of linguistic features, including n-gram~\\citep[BLEU,][]{papineni2002bleu}, token~\\citep[TER,][]{snover2006a}, and character~\\citep[ChrF \\& ChrF++,][]{popovic2015chrf,popovic2017chrf++}.\n",
"However, recent studies pointed out that these metrics have low consistency with human judgments and insufficiently evaluate high-qualified MT systems~\\cite{freitag2020bleu,rei2020comet,mathur2020tangled}.\n",
"Consequently, with the rapid development of PLMs, researchers have been paying their attention to model-based approaches.\n",
"To further improve the model, \\newcite{sellam2020bleurt} pretrained a specific PLM for the translation evaluation (BLEURT), while \\newcite{lo2019yisi} combined statistical and representative features (YiSi-1). Both these methods achieve higher correlations with human judgments than statistical counterparts.\n",
"\\subsection{Source-Only Evaluation}\n",
"\\csrc , which also refers to quality estimation\\footnote{Refer to ``quality estimation'' or ``reference-free metric'' in WMT (\\href{http://www.statmt.org/wmt19/qe-task.html}{http://www.statmt.org/wmt19/qe-task.html}, \\href{http://www.statmt.org/wmt21/metrics-task.html}{http://www.statmt.org/wmt21/metrics-task.html}).}, is an important translation evaluation task especially for the scenario where the ground-truth reference is unavailable.\n",
"It takes the source-side sentence and the translation candidate as inputs for the quality estimation.\n",
"To achieve this, the methods are required to model cross-lingual semantic alignments. \n",
"In this way, methods in this context can evaluate the translation candidate via utilizing the features from both sides. \n",
"As a rising paradigm among translation evaluation tasks, \\csrcref~also profits from the development of cross-lingual PLMs. For example, \n",
"finetuning PLMs over human-annotated datasets~\\citep[COMET,][]{rei2020comet} achieves new state-of-the-art results among all evaluation approaches in WMT 2020~\\cite{mathur2020results}.\n",
"\\begin{figure*}\n",
" \\centering\n",
" \\scalebox{0.9}\n",
" {\n",
" \\begin{tikzpicture}\n",
" \\draw[rectangle, draw=orange, fill=orange!10, very thick] (-2.875, 0) rectangle node{} (2.875, 3.5);\n",
" \\draw[rectangle, color=orange, draw=orange!30, fill=orange!20] (-2.625, 0.25) rectangle node{Pretrained Language Model} (2.625, 1.0);\n",
" \\draw[rectangle, color=orange, draw=orange!30, fill=orange!20] (-2.625, 1.25) rectangle node{Pooling Layer} (2.625, 2.0);\n",
" \\draw[rectangle, color=orange, draw=orange!30, fill=orange!20] (-2.625, 2.25) rectangle node{Feedforward Network} (2.625, 3.0);\n",
" \\draw[color=orange!80!black] (-2, 3.25) node{\\textbf{\\cmodel}};\n",
" \\draw[rectangle, draw=gray, fill=gray!10, very thick] (-2.875, -1.5) rectangle node(){} (2.875, -0.25);\n",
" \\draw[rectangle, draw=gray!30, fill=gray!20] (-2.625, -1.375) rectangle node(){\\small{$\\mathbf{x}_1$}} (-2.0, -0.875);\n",
" \\draw[rectangle, draw=gray!30, fill=gray!20] (-1.75, -1.375) rectangle node(){\\small{$\\mathbf{x}_2$}} (-1.125, -0.875);\n",
" \\draw[rectangle, draw=gray!30, fill=gray!20] (-0.875, -1.375) rectangle node(){\\small{$\\mathbf{x}_3$}} (-0.25, -0.875);\n",
" \\draw[rectangle, draw=gray!30, fill=gray!20] (0.0, -1.375) rectangle node(){\\small{$\\mathbf{x}_4$}} (0.625, -0.875);\n",
" \\draw (1.3125, -1.125) node{$\\cdots$};\n",
" \\draw[rectangle, draw=gray!30, fill=gray!20] (2.0, -1.375) rectangle node(){\\small{$\\mathbf{x}_L$}} (2.625, -0.875);\n",
" \\draw[color=gray!80!black] (-1.25, -0.5) node{\\textbf{Input Sequence}};\n",
" \\draw[rectangle, color=violet!80!black, draw=violet!50, fill=violet!10, very thick] (-0.75, 3.75) rectangle node{\\textbf{Score}} (0.75, 4.25);\n",
" \\draw [->, color=black!50, very thick](0, -0.25) -- (0, 0.0);\n",
" \\draw [->, color=black!50, very thick](0, 3.5) -- (0, 3.75);\n",
" \\draw [rectangle, draw=black!50, dashed, very thick] (-3.0, -1.75) rectangle node{} (3.0, 4.5);\n",
" \\filldraw[draw=blue!30, fill=blue!10, very thick] (3.5, 4.0) -- (5.0, 4.0) -- (5.0, 4.25) -- (5.5, 3.75) -- (5.0, 3.25) -- (5.0, 3.5) -- (3.5, 3.5) -- cycle;\n",
" \\draw (4.5, 3.75) node{\\textsc{Ref}};\n",
" \\filldraw[draw=red!30, fill=red!10, very thick] (3.5, 1.625) -- (5.0, 1.625) -- (5.0, 1.875) -- (5.5, 1.375) -- (5.0, 0.875) -- (5.0, 1.125) -- (3.5, 1.125) -- cycle;\n",
" \\draw (4.5, 1.375) node{\\textsc{Src}};\n",
" \\filldraw[draw=green!30, fill=green!10, very thick] (3.5, -0.75) -- (5.0, -0.75) -- (5.0, -0.5) -- (5.5, -1.0) -- (5.0, -1.5) -- (5.0, -1.25) -- (3.5, -1.25) -- cycle;\n",
" \\draw (4.5, -1) node{\\textsc{Src+Ref}};\n",
" \\filldraw[draw=blue!30, fill=blue!10, very thick] ;\n",
" \\draw (4.5, 3.75) node{\\textsc{Ref}};\n",
" \\draw[rectangle, draw=blue!50, very thick, dashed] (6.0, 2.5) rectangle node(){} (9.0, 4.5);\n",
" \\draw[rectangle, draw=orange!10, fill=orange!5, very thick, dashed] (6.125, 3.0) rectangle node{} (8.875, 4.0);\n",
" \\draw[rectangle, draw=gray!50, fill=gray!30, very thick] (6.125, 2.625) rectangle node{\\tiny{Hyp; Ref}}(8.875, 2.875);\n",
" \\draw[rectangle, draw=violet!10, fill=violet!5, very thick, dashed] (7.125, 4.125) rectangle node{} (7.875, 4.375);\n",
" \\draw[rectangle, color=orange, draw=orange!50, fill=orange!30, very thick] (6.25, 3.125) rectangle node{\\tiny{PLM}} (8.75, 3.375);\n",
" \\foreach \\x in {0,...,3}\n",
" \\foreach \\y in {0,...,3}\n",
" {\n",
" \\draw[draw=gray!50, fill=white] (10+0.25*\\x, 3+0.25*\\y) circle (0.1);\n",
" }\n",
" \\draw [color=brown, draw=brown!30, decorate, decoration = {brace}, very thick] (9.9, 3.875) -- (10.35, 3.875) node[pos=0.5, above=-0.025]{\\tiny{Hyp}};\n",
" \\draw [color=cyan, draw=cyan!30, decorate, decoration = {brace}, very thick] (10.4, 3.875) -- (10.85, 3.875) node[pos=0.5, above=0.025]{\\tiny{Ref}};\n",
" \\draw [color=brown, draw=brown!30, decorate, decoration = {brace}, very thick] (9.875, 3.4) -- (9.875, 3.85) node[pos=0.5, left=0.025]{\\tiny{Hyp}};\n",
" \\draw [color=cyan, draw=cyan!30, decorate, decoration = {brace}, very thick] (9.875, 2.9) -- (9.875, 3.35) node[pos=0.5, left=0.025]{\\tiny{Ref}};\n",
" \\draw [color=black] (12.5, 3.6) node{\\small{Full Attention}};\n",
" \\draw[rectangle, draw=red!50, very thick, dashed] (6.0, 0.375) rectangle node(){} (9.0, 2.375);\n",
" \\draw[rectangle, draw=orange!10, fill=orange!5, very thick, dashed] (6.125, 0.875) rectangle node{} (8.875, 1.875);\n",
" \\draw[rectangle, draw=gray!50, fill=gray!30, very thick] (6.125, 0.5) rectangle node{\\tiny{Hyp; Src}}(8.875, 0.75);\n",
" \\draw[rectangle, draw=violet!10, fill=violet!5, very thick, dashed] (7.125, 2.0) rectangle node{} (7.875, 2.25);\n",
" \\draw [color=black!80] (10.625, 0.4) node{\\tiny{Key}};\n",
" \\draw [color=black!80] (9.25, -0.875) node{\\rotatebox{90}{\\tiny{Query}}};\n",
" \\draw[rectangle, color=orange, draw=orange!50, fill=orange!30, very thick] (6.25, 1.0) rectangle node{\\tiny{PLM}} (8.75, 1.25);\n",
" \\foreach \\x in {0,...,3}\n",
" \\foreach \\y in {0,...,3}\n",
" {\n",
" \\draw[draw=gray!50, fill=white] (10+0.25*\\x, 0.875+0.25*\\y) circle (0.1);\n",
" }\n",
" \\draw [color=brown, draw=brown!30, decorate, decoration = {brace}, very thick] (9.9, 1.75) -- (10.35, 1.75) node[pos=0.5, above=-0.025]{\\tiny{Hyp}};\n",
" \\draw [color=magenta, draw=magenta!30, decorate, decoration = {brace}, very thick] (10.4, 1.75) -- (10.85, 1.75) node[pos=0.5, above=0.025]{\\tiny{Src}};\n",
" \\draw [color=brown, draw=brown!30, decorate, decoration = {brace}, very thick] (9.875, 1.275) -- (9.875, 1.725) node[pos=0.5, left=0.025]{\\tiny{Hyp}};\n",
" \\draw [color=magenta, draw=magenta!30, decorate, decoration = {brace}, very thick] (9.875, 0.775) -- (9.875, 1.225) node[pos=0.5, left=0.025]{\\tiny{Src}};\n",
" \\draw [color=black] (12.5, 1.375) node{\\small{Full Attention}};\n",
" \\draw[rectangle, draw=green!50, very thick, dashed] (6.0, -1.75) rectangle node(){} (9.0, 0.25);\n",
" \\draw[rectangle, draw=orange!10, fill=orange!5, very thick, dashed] (6.125, -1.25) rectangle node{} (8.875, -0.25);\n",
" \\draw[rectangle, draw=gray!50, fill=gray!30, very thick] (6.125, -1.625) rectangle node{\\tiny{Hyp; Src; Ref}}(8.875, -1.375);\n",
" \\draw[rectangle, draw=violet!10, fill=violet!5, very thick, dashed] (7.125, -0.125) rectangle node{} (7.875, 0.125);\n",
" \\draw[rectangle, color=orange, draw=orange!50, fill=orange!30, very thick] (6.25, -1.125) rectangle node{\\tiny{PLM}} (8.75, -0.875);\n",
" \\foreach \\x in {0,...,5}\n",
" \\foreach \\y in {0,...,5}\n",
" {\n",
" \\ifthenelse{\\(\\x<2 \\AND \\y>1 \\AND \\y<4\\) \\OR \\(\\x<4 \\AND \\y<2\\)}\n",
" {\n",
" \\draw[draw=gray!50, fill=white] (10+0.25*\\x, -1.5+0.25*\\y) circle (0.1) node {\\textcolor{red}{\\tiny{\\ding{55}}}};\n",
" }\n",
" {\n",
" \\draw[draw=gray!50, fill=white] (10+0.25*\\x, -1.5+0.25*\\y) circle (0.1);\n",
" }\n",
" }\n",
" \\draw [color=brown, draw=brown!30, decorate, decoration = {brace}, very thick] (9.9, -0.125) -- (10.35, -0.125) node[pos=0.5, above=-0.025]{\\tiny{Hyp}};\n",
" \\draw [color=magenta, draw=magenta!30, decorate, decoration = {brace}, very thick] (10.4, -0.125) -- (10.85, -0.125) node[pos=0.5, above=0.025]{\\tiny{Src}};\n",
" \\draw [color=cyan, draw=cyan!30, decorate, decoration = {brace}, very thick] (10.9, -0.125) -- (11.35, -0.125) node[pos=0.5, above=0.025]{\\tiny{Ref}};\n",
" \\draw [color=brown, draw=brown!30, decorate, decoration = {brace}, very thick] (9.875, -0.6) -- (9.875, -0.15) node[pos=0.5, left=0.025]{\\tiny{Hyp}};\n",
" \\draw [color=magenta, draw=magenta!30, decorate, decoration = {brace}, very thick] (9.875, -1.1) -- (9.875, -0.65) node[pos=0.5, left=0.025]{\\tiny{Src}};\n",
" \\draw [color=cyan, draw=cyan!30, decorate, decoration = {brace}, very thick] (9.875, -1.6) -- (9.875, -1.15) node[pos=0.5, left=0.025]{\\tiny{Ref}};\n",
" \\node[draw, color=black, draw=none, fill=none, align=left] at (12.5, -0.75) () {\\small{Monotonic} \\\\ \\small{Regional} \\\\ \\small{Attention}};\n",
" \\end{tikzpicture}\n",
" }\n",
" \\label{fig.architecture}\n",
"\\end{figure*}\n",
"\\section{Methodology}\n",
"As mentioned above, massive methods are proposed for different automatic evaluation tasks.\n",
"On the one hand, it is inconvenient and expensive to develop and employ different metrics for different evaluation scenarios.\n",
"On the other hand, separate models absolutely overlook the commonalities among these evaluation tasks, of which knowledge potentially benefits all three tasks.\n",
"In order to fulfill the aim of unifying the functionalities on \\cref, \\csrc, and \\csrcref~into one model, in this section, we introduce \\textbf{\\cmodel} (Figure~\\ref{fig.architecture}).\n",
"\\subsection{Model Architecture}\n",
"\\label{sec:model}\n",
"By receiving a data example composing of hypothesis, source, and reference segment, \\cmodel~first modifies it into concatenated sequence following the given setting as \\cref, \\csrc, or \\csrcref:\n",
"\\begin{align}\n",
" \\mathbf{x}_\\textsc{Ref} & = \\texttt{Concat} (\\mathbf{h}, \\mathbf{r}) \\in \\mathbb{R}^{(l_h + l_r)}, \\notag \\\\\n",
" \\mathbf{x}_\\textsc{Src} & = \\texttt{Concat}(\\mathbf{h}, \\mathbf{s}) \\in \\mathbb{R}^{(l_h + l_s)}, \\\\\n",
" \\mathbf{x}_\\textsc{Src+Ref} & = \\texttt{Concat}(\\mathbf{h}, \\mathbf{s}, \\mathbf{r}) \\in \\mathbb{R}^{(l_h + l_s + l_r)}, \\notag\n",
"\\end{align}\n",
"where $\\mathbf{h}$, $\\mathbf{s}$ and $\\mathbf{r}$ are hypothesis, source and reference segments, with the corresponding sequence lengths being $l_h$, $l_s$ and $l_r$, respectively.\n",
"The input sequence is then fed to PLM to derive representations $\\tilde{\\mathbf{H}}$.\n",
"Take \\cref~as an example:\n",
"\\begin{align}\n",
" \\tilde{\\mathbf{H}}_\\textsc{Ref} = \\texttt{PLM}(\\mathbf{x}_\\textsc{Ref}) \\in \\mathbb{R}^{(l_h + l_r) \\times d},\n",
"\\end{align}\n",
"where $d$ is the model size of PLM.\n",
"According to~\\newcite{ranasinghe2020transquest}, we use the first output representation as the input of feedforward layer. \n",
"Compared to existing methods~\\cite{zhang2020bertscore,rei2020comet} which take sentence-level representations for evaluation, the advantages of our architecture design are as follows.\n",
"First, our \\cmodel~model can benefit from layer-coordinated semantical interactions inside every one of PLM layers, which is proven effective on capturing diverse linguistic features ~\\cite{he2018layer,lin2019open,jawahar2019what,tenney2019bert,rogers2020primer}.\n",
"Second, for the unified approach of our model, the concatenation provides the unifying format for all task inputs, turning our model into a more general architecture.\n",
"When conducting different evaluation tasks, our model requires no further modification inside.\n",
"Note here, to keep the consistency across all evaluation tasks, as well as ease the unified learning, $\\mathbf{h}$ is always located at the beginning of the input sequence.\n",
"After deriving $\\tilde{\\mathbf{H}}_{\\textsc{Ref}}$, a pooling block is arranged after PLM which gives sequence-level representations $\\mathbf{H}_{\\textsc{Ref}}$.\n",
"Finally, a feedforward network takes $\\mathbf{H}_{\\textsc{Ref}}$ as input, and gives a scalar $p$ as prediction:\n",
"\\begin{align}\n",
" \\mathbf{H}_\\textsc{Ref} & = \\texttt{Pool}(\\tilde{\\mathbf{H}}_\\textsc{Ref}) \\in \\mathbb{R}^{d}, \\\\\n",
" p_\\textsc{Ref} & = \\texttt{FeedForward}(\\mathbf{H}_\\textsc{Ref}) \\in \\mathbb{R}^{1}.\n",
"\\end{align}\n",
"For training, we encourage the model to reduce the mean squared error with respect to given score $q$:\n",
"\\begin{align}\n",
" \\mathcal{L}_\\textsc{Ref} = (p_\\textsc{Ref} - q) ^ 2.\n",
"\\end{align}\n",
"However, for the pretraining of most PLMs~\\citep[e,g., XLM-R,][]{conneau2020unsupervised}, the input patterns are designed to receive two segments at most.\n",
"Thus there exists a gap between the pretraining of PLM and the joint training of \\cmodel~where the concatenation of three fragments is used as input.\n",
"To alleviate this issue, we propose two strategies: \\textbf{Monotonic Regional Attention} as described in~\\S\\ref{sec:mra} and \\textbf{Unified Pretraining} in~\\S\\ref{sec:up}.\n",
"\\subsection{Monotonic Regional Attention}\n",
"\\label{sec:mra}\n",
"To fill the modeling gap between the pretraining of PLM and the joint training of three downstream tasks, a natural idea is to unify the number of involved segments when modeling semantics for \\csrc, \\cref~and \\csrcref~tasks.\n",
"Following this, we propose to modify the attention mask of \\csrcref~to simulate the modeling of two segments in \\csrc~and \\cref.\n",
"Specifically, when calculating the attention logits, semantics from a specific segment are only allowed to derive information from two segments at most.\n",
"Considering the conventional attention module:\n",
"\\begin{align}\n",
" \\mathbf{A} = \\mathrm{Softmax}(\\frac{\\mathbf{Q}\\mathbf{K}^\\top}{\\sqrt{d}}) \\in \\mathbb{R}^{L \\times L},\n",
"\\end{align}\n",
"where $L$ is the sequential length for input, $\\mathbf{Q}, \\mathbf{K} \\in \\mathbb{R}^{L \\times d}$ are query and key representations, respectively.\\footnote{For simplicity, we omit the multi-head mechanism.}\n",
"As to monotonic regional attention (MRA), we simply add a mask $\\mathbf{M}$ to the softmax logits to control attention flows:\n",
"\\begin{align}\n",
" \\mathbf{A} & = \\mathrm{Softmax}(\\frac{\\mathbf{Q}\\mathbf{K}^\\top}{\\sqrt{d}} + \\mathbf{M}) \\in \\mathbb{R}^{L \\times L}, \\\\\n",
" \\mathbf{M}_{ij} & = \n",
" \\begin{cases}\n",
" -\\infty & (i,j) \\in \\mathbf{U}, \\\\\n",
" 0 & \\text{otherwise},\n",
" \\end{cases}\n",
"\\end{align}\n",
"where $\\mathbf{U}$ stores the index pairs of all masked areas. \n",
"Following this idea, the key of MRA is how to design the matrix $\\mathbf{U}$.\n",
"For the cases where interactions inside each segment, we believe that these self-interactions are beneficial to the modeling.\n",
"For other cases where interactions are arranged across segments, three patterns are included: hypothesis-reference, source-reference, and hypothesis-source.\n",
"Intuitively, the former two parts are beneficial for model training, since they might contribute the monolingual signals and cross-lingual disambiguation to evaluation, respectively.\n",
"\\tikzset{\n",
"semi/.style={\n",
" semicircle,\n",
" draw,\n",
" }\n",
"}\n",
"\\begin{figure}\n",
" \\centering\n",
" \\scalebox{0.7}{\n",
" \\begin{tikzpicture}\n",
" {\n",
" \\node[fill=brown!20, draw=brown!50, circle, minimum size=30pt] (hyp) at (0, 0) {$\\mathbf{h}$};\n",
" \\node[fill=magenta!20, draw=magenta!50, circle, minimum size=30pt] (src) at (-1.25, -1.5) {$\\mathbf{s}$};\n",
" \\node[fill=cyan!20, draw=cyan!50, circle, minimum size=30pt] (ref) at (1.25, -1.5) {$\\mathbf{r}$};\n",
" \\draw[brown!80, thick, ->] (0.35, 0.45) arc (-30:210:0.4);\n",
" \\draw[magenta!80, thick, ->] (-1.7, -1.15) arc (60:300:0.4);\n",
" \\draw[cyan!80, thick, ->] (1.7, -1.15) arc (120:-120:0.4);\n",
" \\draw[cyan!80, thick, ->] (0.65, -1.5) -- (-0.65, -1.5);\n",
" \\draw[cyan!80, thick, ->] (0.95, -1.0) -- (0.45, -0.35);\n",
" \\draw[magenta!80, thick, ->] (-0.95, -1.0) -- (-0.45, -0.35);\n",
" \\node[fill=brown!20, draw=brown!50, circle, minimum size=30pt] (hyp) at (6, 0) {$\\mathbf{h}$};\n",
" \\node[fill=magenta!20, draw=magenta!50, circle, minimum size=30pt] (src) at (4.75, -1.5) {$\\mathbf{s}$};\n",
" \\node[fill=cyan!20, draw=cyan!50, circle, minimum size=30pt] (ref) at (7.25, -1.5) {$\\mathbf{r}$};\n",
" \\draw[brown!80, thick, ->] (6.35, 0.45) arc (-30:210:0.4);\n",
" \\draw[magenta!80, thick, ->] (4.3, -1.15) arc (60:300:0.4);\n",
" \\draw[cyan!80, thick, ->] (7.7, -1.15) arc (120:-120:0.4);\n",
" \\draw[cyan!80, thick, ->] (6.65, -1.45) -- (5.35, -1.45);\n",
" \\draw[magenta!80, thick, ->] (5.35, -1.55) -- (6.65, -1.55);\n",
" \\draw[cyan!80, thick, ->] (6.98, -0.95) -- (6.48, -0.3);\n",
" \\draw[brown!80, thick, ->] (6.42, -0.4) -- (6.92, -1.05);\n",
" \\draw[magenta!80, thick, ->] (5.05, -1.0) -- (5.55, -0.35);\n",
" }\n",
" \\end{tikzpicture}\n",
" }\n",
" \\caption{Attention flows in monotonic regional attention. $\\mathbf{h}$, $\\mathbf{s}$ and $\\mathbf{r}$ are hypothesis, source and reference, respectively. We prevent specified interactions in \\fboxsep0.5ex\\colorbox{green!10}{\\textsc{Src+Ref}}~training via modifying the attention mask with regional properties. We show the hard (left) and soft design (right, no $\\mathbf{h}\\rightarrow\\mathbf{s}$) in this figure.}\n",
" \\label{fig.monotonic}\n",
"\\end{figure}\n",
"To give more fine-grained designs, we propose two approaches for \\cmodel-MRA, which apply the MRA mechanism into \\cmodel~model (Figure~\\ref{fig.monotonic}):\n",
"\\begin{itemize}\n",
" \\item Hard MRA. Only monotonic attention flows are allowed. Interactions between any two segments are strictly unidirectional through the entire PLM, where\n",
" $\\mathbf{U}$ stores the index pairs of unidirectional interactions of $\\mathbf{h}\\rightarrow\\mathbf{r}$, $\\mathbf{s}\\rightarrow\\mathbf{r}$ and $\\mathbf{h}\\rightarrow\\mathbf{s}$, where ``$\\rightarrow$'' denotes the direction of attention flows.\n",
" \\item Soft MRA. Specific attention flows are forbidden inside each attention module.\n",
" The involved two segments may interact inside a higher layer.\n",
" In practice, index pairs which denoting $\\mathbf{h}\\rightarrow\\mathbf{s}$ or $\\mathbf{s}\\rightarrow\\mathbf{h}$ between source and hypothesis are stored in $\\mathbf{U}$.\n",
"\\end{itemize}\n",
"\\subsection{Unified Pretraining}\n",
"\\label{sec:up}\n",
"To further bridge the modeling gap between PLM and the joint training of \\cmodel~mentioned in~\\S\\ref{sec:model}, we propose a unified pretraining strategy including the following main stages: 1) collecting and downgrading synthetic data; 2) labeling examples with a novel ranking-based strategy; 3) multi-task learning for unified pretraining and finetuning.\n",
"\\paragraph{Synthetic Data Collection}\n",
"As our approach aims at evaluating the quality of translations, generated hypotheses with NMT models are ideal synthetic data.\n",
"To further improve the diversity of synthetic data quality,\n",
"we follow existing experiences~\\cite{sellam2020bleurt,wan2021robleurt} to apply the word and span dropping strategy to downgrade a portion of hypotheses.\n",
"The collected data totally contains $N$ triplets composing of hypothesis, source and reference segments, which is formed as $\\mathcal{D}' = \\{\\langle\\mathbf{h}^i, \\mathbf{s}^i, \\mathbf{r}^i\\rangle\\}_{i=1}^{N}$.\n",
"\\paragraph{Data Labeling}\n",
"After obtaining the synthetic data, the next step is to augment each data pair with a label which serves as the signal of unified pretraining.\n",
"To stabilize the model training, as well as normalize the distributions across all score systems and languages, we propose a novel ranking-based approach.\n",
"This method is based on the idea of Borda count~\\cite{ho1994decision,emerson2013original}, which provides more precise and well-distributed synthetic data labels than Z-score normalization.\n",
"Then, we tag each example with its rank index $\\tilde{q}^i$ referring to $\\hat{q}^i$:\n",
"\\begin{align}\n",
" \\tilde{q}^i = \\texttt{IndexOf}(\\hat{q}^i, \\mathcal{Q}),\n",
"\\end{align}\n",
"where $\\mathcal{Q}$ is the list storing all the sorted $\\hat{q}^i$ descendingly.\n",
"Then, we use the conventional Z-score strategy to normalize the scores:\n",
"\\begin{align}\n",
" q^i = \\frac{\\tilde{q}^i - \\mu}\\sigma,\n",
"\\end{align}\n",
"where $\\mu$ and $\\sigma$ are the mean and the standard deviation of values in $\\mathcal{Q}$, respectively.\n",
"The dataset thus updates its format to $\\mathcal{D} = \\{\\langle\\mathbf{h}^i, \\mathbf{s}^i, \\mathbf{r}^i, q^i\\rangle\\}_{i=1}^{N}$.\n",
"Note here that, an example with higher $\\hat{q}^i$ is assigned with higher $\\tilde{q}^i$, thus a larger value of $q^i$.\n",
"Compared to related approaches which apply Z-score normalization~\\cite{bojar2018findings}, or leave the conventional labeled scores as signals for learning~\\citep[\\textit{i.e.}, knowledge distillation,][]{kim2016sequence,phuong2019towards}, our approach can alleviate the bias of chosen model for labeling and prior distributional disagreement of scores.\n",
"For example, different methods may give scores with different distributions.\n",
"Especially for translation directions of low-resource, scores may follow skewed distribution~\\cite{sellam2020bleurt}, which has a disagreement with rich-resource scenarios.\n",
"Our method can unify the distribution of all labeling data into the same scale, which can also be easily applied by the ensembling strategy.\n",
"\\paragraph{Multi-task Pretrainig and Finetuning}\n",
"To unify all evaluation scenarios into one model, we apply multi-task learning for both pretraining and finetuning.\n",
"For each step, we arrange three substeps for all input formats, yielding $\\mathcal{L}_\\textsc{Ref}$, $\\mathcal{L}_\\textsc{Src}$, and $\\mathcal{L}_\\textsc{Src+Ref}$, respectively.\n",
"The final learning objective is to reduce the summation of all losses:\n",
"\\begin{align}\n",
" \\mathcal{L} = \\mathcal{L}_\\textsc{Ref} + \\mathcal{L}_\\textsc{Src} + \\mathcal{L}_\\textsc{Src+Ref}.\n",
"\\end{align}\n",
"\\begin{table*}[t]\n",
" \\small\n",
" \\centering\n",
" \\scalebox{1.0}\n",
" {\n",
" \\begin{tabular}{lcccccccc}\n",
" \\toprule\n",
" \\multirow{2}{*}{\\textbf{Model}} & \\multicolumn{4}{c}{\\textbf{High-resource}} & \\multicolumn{3}{c}{\\textbf{Zero-shot}} & \\multirow{2}{*}{\\textbf{Avg.}} \\\\\n",
" \\cmidrule(l{2pt}r{2pt}){2-5}\\cmidrule(l{2pt}r{2pt}){6-8}\n",
" & \\textit{\\underline{De-En}} & \\textit{\\underline{Ru-En}} & \\textit{\\underline{Zh-En}} & \\underline{Fi-En} & Gu-En & Kk-En & Lt-En & \\\\\n",
" \\midrule\n",
" \\cdashline{1-9}\\noalign{\\vskip 0.1ex}\n",
" $^\\heartsuit$BLEU~\\cite{papineni2002bleu} & ~~5.4 & 11.5 & 32.1 & 23.6 & 19.4 & 27.6 & 24.9 & \\colorbox{blue!10}{20.6} \\\\\n",
" $^\\spadesuit$ChrF~\\cite{popovic2015chrf} & 12.3 & 17.7 & 37.1 & 29.2 & 24.0 & 32.3 & 30.4 & \\colorbox{blue!10}{26.1} \\\\\n",
" $^\\heartsuit$BERTScore~\\cite{zhang2020bertscore} & 19.0 & 22.1 & 43.0 & 35.4 & 29.2 & 35.1 & 38.1 & \\colorbox{blue!10}{31.7} \\\\\n",
" $^\\heartsuit$BLEURT~\\cite{sellam2020bleurt} & 17.4 & 22.0 & 43.6 & 37.4 & 31.3 & 37.2 & 38.8 & \\colorbox{blue!10}{32.5} \\\\\n",
" $^\\spadesuit$YiSi-1~\\cite{lo2019yisi} & 16.4 & 21.7 & 42.6 & 34.7 & 31.2 & \\textbf{44.0} & 37.6 & \\colorbox{blue!10}{32.6} \\\\\n",
" $^\\heartsuit$PRISM-ref~\\cite{thompson2020automatic} & 20.4 & \\textbf{22.5} & 43.8 & 35.7 & 31.3 & 43.4 & 38.2 & \\colorbox{blue!10}{33.6} \\\\\n",
" $^\\heartsuit$BARTScore~\\cite{yuan2021bartscore} & 23.8 & 21.9 & 44.7 & 37.4 & 31.8 & 37.6 & 38.6 & \\colorbox{blue!10}{33.7} \\\\\n",
" $^\\diamondsuit$XLM-R+Concat~\\cite{takahashi2020automatic} & 24.5 & 21.8 & 45.8 & 37.0 & 31.5 & 37.4 & 39.5 & \\colorbox{blue!10}{33.9} \\\\\n",
" $^\\diamondsuit$RoBERTa+Concat~\\cite{takahashi2020automatic} & 25.1 & 22.4 & 46.4 & 36.2 & 30.8 & 38.0 & \\textbf{40.0} & \\colorbox{blue!10}{34.1} \\\\\n",
" \\cdashline{1-9}\\noalign{\\vskip 0.1ex}\n",
" \\cmodel-MRA & 25.2 & 22.4 & 46.4 & 36.5 & 31.6 & 38.4 & 39.1 & \\colorbox{blue!10}{34.2} \\\\\n",
" \\cmodel-UP & \\textbf{25.9} & 21.9 & \\textbf{46.7} & \\textbf{37.9} & \\textbf{32.2} & 38.7 & \\textbf{40.0} & \\colorbox{blue!10}{\\textbf{34.8}} \\\\\n",
" \\midrule\n",
" \\cdashline{1-9}\\noalign{\\vskip 0.1ex}\n",
" $^\\spadesuit$YiSi-2~\\cite{lo2019yisi} & ~~6.8 & ~~5.3 & 25.3 & 12.6 & ~-0.1 & ~~9.6 & ~~7.5 & \\colorbox{red!10}{~~9.5} \\\\\n",
" $^\\heartsuit$PRISM-src~\\cite{thompson2020automatic} & 10.9 & \\textbf{17.8} & 33.6 & 30.0 & 10.2 & \\textbf{39.1} & 35.6 & \\colorbox{red!10}{25.3} \\\\\n",
" $^\\heartsuit$MTransQuest~\\cite{ranasinghe2020transquest} & 11.1 & 14.0 & 32.1 & 29.7 & 27.2 & 31.6 & 30.7 & \\colorbox{red!10}{25.2} \\\\\n",
" $^\\diamondsuit$MTransQuest~\\cite{ranasinghe2020transquest} & 17.0 & 17.3 & 37.6 & 29.2 & 26.5 & 31.9 & 34.2 & \\colorbox{red!10}{27.7} \\\\\n",
" $^\\diamondsuit$XLM-R+Concat~\\cite{takahashi2020automatic} & 16.9 & 17.6 & 38.1 & 29.1 & 26.2 & 31.6 & 34.3 & \\colorbox{red!10}{27.7} \\\\\n",
" \\cdashline{1-9}\\noalign{\\vskip 0.1ex}\n",
" \\cmodel-MRA & 17.4 & 17.7 & 41.0 & \\textbf{34.3} & 29.0 & 32.7 & \\textbf{36.2} & \\colorbox{red!10}{29.7} \\\\\n",
" \\cmodel-UP & \\textbf{19.3} & 16.9 & \\textbf{41.4} & 34.0 & \\textbf{29.7} & 33.6 & 35.4 & \\colorbox{red!10}{\\textbf{30.0}} \\\\\n",
" \\midrule\n",
" \\cdashline{1-9}\\noalign{\\vskip 0.1ex}\n",
" $^\\diamondsuit$XLM-R+Concat~\\cite{takahashi2020automatic} & 24.0 & 22.0 & 44.7 & 35.7 & 30.4 & 37.2 & 38.9 & \\colorbox{green!10}{33.4} \\\\\n",
" $^\\diamondsuit$COMET~\\cite{rei2020comet} & 23.4 & 20.7 & 45.8 & 36.2 & 30.9 & 37.9 & 40.3 & \\colorbox{green!10}{33.6} \\\\\n",
" \\cdashline{1-9}\\noalign{\\vskip 0.1ex}\n",
" \\cmodel-MRA & 25.6 & \\textbf{22.9} & 46.9 & 37.6 & 31.6 & 38.5 & \\textbf{40.5} & \\colorbox{green!10}{34.8} \\\\\n",
" \\cmodel-UP & \\textbf{26.0} & 22.0 & \\textbf{47.2} & \\textbf{37.7} & \\textbf{32.3} & \\textbf{39.4} & 40.0 & \\colorbox{green!10}{\\textbf{35.0}} \\\\\n",
" \\bottomrule\n",
" \\end{tabular}\n",
" }\n",
" \\caption{Kendall's Tau correlation (\\\n",
" \\label{table.main_results_metric_x-en}\n",
"\\end{table*}\n",
"\\begin{table*}[t]\n",
" \\centering\n",
" \\scalebox{0.675}{\n",
" \\begin{tabular}{lcccccccccccc}\n",
" \\toprule\n",
" \\multirow{2}{*}{\\textbf{Model}} & \\multicolumn{5}{c}{\\textbf{High-resource}} & \\multicolumn{6}{c}{\\textbf{Zero-shot}} & \\multirow{2}{*}{\\textbf{Avg.}} \\\\\n",
" \\cmidrule(l{2pt}r{2pt}){2-6}\\cmidrule(l{2pt}r{2pt}){7-12}\n",
" & \\textit{\\underline{En-Cs}} & \\textit{\\underline{En-De}} & \\textit{\\underline{En-Ru}} & \\textit{\\underline{En-Zh}} & \\underline{En-Fi} & En-Gu & En-Kk & En-Lt & De-Cs & De-Fr & Fr-De & \\\\\n",
" \\midrule\n",
" \\cdashline{1-13}\\noalign{\\vskip 0.1ex}\n",
" $^\\heartsuit$BLEU~\\cite{papineni2002bleu} & 36.4 & 24.8 & 46.9 & 23.5 & 39.5 & 46.3 & 36.3 & 33.3 & 22.2 & 22.6 & 17.3 & \\colorbox{blue!10}{31.7} \\\\\n",
" $^\\spadesuit$ChrF~\\cite{popovic2015chrf} & 44.4 & 32.1 & 54.8 & 24.1 & 51.8 & 54.8 & 51.0 & 43.8 & 34.1 & 28.7 & 27.4 & \\colorbox{blue!10}{40.6} \\\\\n",
" $^\\heartsuit$BERTScore~\\cite{zhang2020bertscore} & 50.0 & 36.3 & \\textbf{58.5} & 35.6 & 52.7 & 56.8 & 54.0 & 46.4 & 35.8 & 32.9 & 30.0 & \\colorbox{blue!10}{44.5} \\\\\n",
" $^\\spadesuit$YiSi-1~\\cite{lo2019yisi} & 47.5 & 35.1 & \\textbf{58.5} & 35.5 & 53.7 & 55.1 & 54.6 & 47.0 & 37.6 & 34.9 & 31.0 & \\colorbox{blue!10}{44.6} \\\\\n",
" $^\\spadesuit$BLEURT~\\cite{sellam2020learning} & 60.3 & 42.2 & 49.2 & 33.7 & 61.5 & 57.7 & 55.8 & 58.4 & 46.1 & 44.9 & \\textbf{42.7} & \\colorbox{blue!10}{50.2} \\\\\n",
" $^\\diamondsuit$XLM-R+Concat~\\cite{takahashi2020automatic} & 60.2 & 43.0 & 58.1 & 41.0 & 60.2 & 60.8 & 60.1 & 58.8 & 47.0 & 45.1 & 40.9 & \\colorbox{blue!10}{52.3} \\\\\n",
" \\cdashline{1-13}\\noalign{\\vskip 0.1ex}\n",
" \\cmodel-UP & 60.1 & 44.4 & 50.7 & \\textbf{45.3} & 62.2 & 62.1 & 61.1 & \\textbf{61.5} & 48.3 & \\textbf{47.3} & 42.3 & \\colorbox{blue!10}{53.2} \\\\\n",
" \\cmodel-MUP & \\textbf{62.1} & \\textbf{45.6} & 52.2 & 44.8 & \\textbf{62.5} & \\textbf{63.0} & \\textbf{61.9} & 61.4 & \\textbf{49.1} & 46.9 & 42.3 & \\colorbox{blue!10}{\\textbf{53.8}} \\\\\n",
" \\midrule\n",
" \\cdashline{1-13}\\noalign{\\vskip 0.1ex}\n",
" $^\\heartsuit$MTransQuest~\\cite{ranasinghe2020transquest} & 35.8 & 28.4 & 31.1 & 29.0 & 50.8 & 52.7 & 56.3 & 43.9 & 35.7 & 23.7 & ~~9.4 & \\colorbox{red!10}{36.1} \\\\\n",
" $^\\diamondsuit$MTransQuest~\\cite{ranasinghe2020transquest} & 40.2 & 33.1 & \\textbf{32.9} & 32.8 & 54.2 & 57.2 & 60.2 & 49.1 & 40.4 & 29.8 & 17.9 & \\colorbox{red!10}{40.7} \\\\\n",
" $^\\diamondsuit$XLM-R+Concat~\\cite{takahashi2020automatic} & 53.5 & 38.0 & 30.2 & 34.0 & 53.9 & 55.9 & 53.5 & 53.8 & 35.7 & 32.5 & 31.5 & \\colorbox{red!10}{42.9} \\\\\n",
" \\cdashline{1-13}\\noalign{\\vskip 0.1ex}\n",
" \\cmodel-UP & 52.3 & 41.7 & 27.3 & \\textbf{40.7} & 60.7 & 59.1 & 60.4 & 56.8 & 40.7 & \\textbf{37.0} & 32.1 & \\colorbox{red!10}{46.3} \\\\\n",
" \\cmodel-MUP & \\textbf{55.9} & \\textbf{43.8} & 28.7 & 40.6 & \\textbf{61.9} & \\textbf{60.5} & \\textbf{61.1} & \\textbf{59.3} & \\textbf{41.4} & 35.6 & \\textbf{36.7} & \\colorbox{red!10}{\\textbf{47.8}} \\\\\n",
" \\midrule\n",
" \\cdashline{1-13}\\noalign{\\vskip 0.1ex}\n",
" $^\\diamondsuit$XLM-R+Concat~\\cite{takahashi2020automatic} & 60.9 & 43.3 & 53.3 & 40.8 & 60.4 & 60.1 & 59.1 & 59.3 & 46.4 & 44.9 & 40.5 & \\colorbox{green!10}{51.7} \\\\\n",
" $^\\diamondsuit$COMET~\\cite{rei2020comet} & 61.0 & 44.6 & \\textbf{58.3} & 42.3 & 62.3 & 60.7 & 59.0 & 60.6 & 45.7 & 46.8 & 38.8 & \\colorbox{green!10}{52.7} \\\\\n",
" \\cdashline{1-13}\\noalign{\\vskip 0.1ex}\n",
" \\cmodel-UP & 60.0 & 44.9 & 49.7 & \\textbf{45.6} & 62.7 & 62.6 & 62.0 & 61.0 & 48.0 & 45.5 & \\textbf{42.4} & \\colorbox{green!10}{53.1} \\\\\n",
" \\cmodel-MUP & \\textbf{62.2} & \\textbf{46.0} & 54.6 & 44.9 & \\textbf{63.2} & \\textbf{63.2} & \\textbf{63.0} & \\textbf{61.8} & \\textbf{48.7} & \\textbf{47.5} & \\textbf{42.4} & \\colorbox{green!10}{\\textbf{54.3}} \\\\\n",
" \\bottomrule\n",
" \\end{tabular}\n",
" }\n",
" \\caption{Kendall's Tau correlation (\\\n",
" }\n",
" \\label{table.main_results_metric_en-x}\n",
"\\end{table*}\n"
],
"output": {
"What experiments do you suggest doing?": [
"1. Perform statistical significance testing between reference-only and source-reference evaluation of the proposed method.",
"2. Alternate the position of source and reference to ablate the effect of positional encoding.",
"3. Conduct fine-grained analysis on the error patterns of unified model vs separate model on the reference-only, source-only and source-reference evaluation scenario.",
"4. Conduct pseudo-labeling with other approaches. Instead of using an earlier checkpoint of the proposed approach use some well-established method such as BERTScore to pseudo-label the pretraining data."
],
"Why do you suggest these experiments?": [
"1. To understand whether the proposed approach is using the source sentence while performing evaluation. Since the performance gain is minute when using both source and reference compared to just using reference, and since using just using source it is important to establish whether the performance gain of using both source and reference is significant.",
"2. To understand whether positional encoding is the culprit in the performance reduction for the source to hypothesis direction. By alternating the position of the source and reference the effect of positional encoding can be separated out.",
"3. To understand why the performance drops significantly when the model is trained on each task separately. This is surprising since separate training should make the model evaluation and training be on the same task. Specifically, since source-only evaluation does not have access to reference it is important to understand how the unified model is exploiting access to reference during training to perform well during evaluation.",
"4. To disentangle the benefits of the proposed ranking approach from the monotonic regional attention proposed by the authors. Employing already established models to pseudo-label the data will also make the method more reproducible and the impact of score distribution of evaluation metric on the different ranking approaches can be analyzed."
]
},
"paper_info": {
"title": "UniTE: Unified Translation Evaluation",
"authors": [
"Yu Wan",
"Dayiheng Liu",
"Baosong Yang",
"Haibo Zhang",
"Boxing Chen",
"Derek F. Wong",
"Lidia S. Chao"
],
"abstract": "Translation quality evaluation plays a crucial role in machine translation.\nAccording to the input format, it is mainly separated into three tasks, i.e.,\nreference-only, source-only and source-reference-combined. Recent methods,\ndespite their promising results, are specifically designed and optimized on one\nof them. This limits the convenience of these methods, and overlooks the\ncommonalities among tasks. In this paper, we propose UniTE, which is the first\nunified framework engaged with abilities to handle all three evaluation tasks.\nConcretely, we propose monotonic regional attention to control the interaction\namong input segments, and unified pretraining to better adapt multi-task\nlearning. We testify our framework on WMT 2019 Metrics and WMT 2020 Quality\nEstimation benchmarks. Extensive analyses show that our \\textit{single model}\ncan universally surpass various state-of-the-art or winner methods across\ntasks. Both source code and associated models are available at\nhttps://github.com/NLP2CT/UniTE.",
"comments": "ACL2022"
},
"raw_data": {
"context_before_exp": [
"\n",
"\\pdfoutput=1\n",
"\n",
"\n",
"\\documentclass[11pt]{article}\n",
"\n",
"\n",
"\n",
"\\usepackage{acl}\n",
"\n",
"\n",
"\\usepackage{times}\n",
"\\usepackage{latexsym}\n",
"\n",
"\n",
"\\usepackage[T1]{fontenc}\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\\usepackage[utf8]{inputenc}\n",
"\n",
"\n",
"\n",
"\n",
"\\usepackage{microtype}\n",
"\n",
"\\usepackage{amsmath}\n",
"\\usepackage{amssymb}\n",
"\\usepackage{xcolor}\n",
"\\usepackage{arydshln}\n",
"\\usepackage{tikz}\n",
"\\usetikzlibrary{arrows, decorations.text, shapes.geometric, positioning, decorations.pathreplacing, calligraphy}\n",
"\\usepackage{pgfplots}\n",
"\\usepackage{pgfmath}\n",
"\\usepackage{pgffor}\n",
"\\pgfplotsset{compat=1.17}\n",
"\\usepackage{subcaption}\n",
"\\usepackage{mathtools}\n",
"\\usepackage{multirow}\n",
"\\usepackage{booktabs}\n",
"\\usepackage{pifont}\n",
"\\usepackage{ifthen}\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\\newcommand{\\cref}{\\colorbox{blue!10}{\\textsc{Ref}}}\n",
"\\newcommand{\\csrc}{\\colorbox{red!10}{\\textsc{Src}}}\n",
"\\newcommand{\\csrcref}{\\colorbox{green!10}{\\textsc{Src+Ref}}}\n",
"\\newcommand{\\ccref}{\\fboxsep0.5ex\\colorbox{blue!10}{\\textsc{Ref}}}\n",
"\\newcommand{\\ccsrc}{\\fboxsep0.5ex\\colorbox{red!10}{\\textsc{Src}}}\n",
"\\newcommand{\\ccsrcref}{\\fboxsep0.5ex\\colorbox{green!10}{\\textsc{Src+Ref}}}\n",
"\n",
"\\newcommand{\\cmodel}{UniTE}\n",
"\n",
"\\title{\\cmodel: Unified Translation Evaluation}\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\\author{Yu Wan$^{a,b}$\\thanks{~~Work was done when Yu Wan was interning at DAMO Academy, Alibaba Group.}~~~Dayiheng Liu$^b$\\thanks{~~Dayiheng Liu and Derek F. Wong are co-corresponding authors.}~~~Baosong Yang$^b$~~~Haibo Zhang$^b$~~~Boxing Chen$^b$\\\\\\textbf{Derek F. Wong}$^{b\\dagger}$~~~\\textbf{Lidia S. Chao}$^a$ \\\\\n",
" $^a$NLP$^2$CT Lab,\n",
" University of Macau\\\\\n",
" {\\tt [email protected], \\{derekfw,lidiasc\\}@umac.mo} \\\\\n",
" $^b$Alibaba Group\\\\\n",
" {\\tt \\{liudayiheng.ldyh,yangbaosong.ybs,zhanhui.zhb,}\\\\\n",
" {\\tt boxing.cbx\\}@alibaba-inc.com}}\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\\begin{document}\n",
"\\maketitle\n",
"\\begin{abstract}\n",
"Translation quality evaluation plays a crucial role in machine translation. According to the input format, it is mainly separated into three tasks, \\textit{i.e.}, reference-only, source-only and source-reference-combined.\n",
"Recent methods, despite their promising results, are specifically designed and optimized on one of them. \n",
"This limits the convenience of these methods, and overlooks the commonalities among tasks.\n",
"In this paper, we propose \\cmodel, which is the first unified framework engaged with abilities to handle all three evaluation tasks.\n",
"Concretely, we propose monotonic regional attention to control the interaction among input segments, and unified pretraining to better adapt multi-task learning. \n",
"We testify our framework on WMT 2019 Metrics and WMT 2020 Quality Estimation benchmarks. Extensive analyses show that our \\textit{single model} can universally surpass various state-of-the-art or winner methods across tasks.\n",
"Both source code and associated models are available at \\href{https://github.com/NLP2CT/UniTE}{https://github.com/NLP2CT/UniTE}.\n",
"\n",
"\\end{abstract}\n",
"\n",
"\n",
"\\section{Introduction}\n",
"Automatically evaluating the translation quality with the given reference segment(s), is of vital importance to identify the performance of Machine Translation (MT) models~\\cite{freitag2020bleu,mathur2020tangled,zhao2020limitations,kocmi2021ship}.\n",
"Based on the input contexts, translation evaluation can be mainly categorized into three classes:\n",
"1) reference-only evaluation (\\cref) approaches like BLEU~\\cite{papineni2002bleu} and BLEURT~\\cite{sellam2020bleurt}, which evaluate the hypothesis by referring the golden reference at target side;\n",
"2) source-only evaluation (\\csrc) methods like YiSi-2~\\cite{lo2019yisi} and TransQuest~\\cite{ranasinghe2020transquest}, which are also referred as quality estimation (QE).\n",
"These methods estimate the quality of the hypothesis based on the source sentence without using references;\n",
"3) source-reference-combined evaluation (\\csrcref) works like COMET~\\cite{rei2020comet}, where the evaluation exploits information from both source and reference.\n",
"With the help of powerful pretrained language models~\\citep[PLMs,][]{devlin2019bert,conneau2020unsupervised}, model-based approaches (\\textit{e.g.}, BLEURT, TransQuest, and COMET) have shown promising results in recent WMT competitions~\\cite{ma2019results,mathur2020results,freitag2021results,fonseca2019findings,specia2020findings,specia2021findings}.\n",
"\n",
"Nevertheless, each existing MT evaluation work is usually designed for one specific task, \\textit{e.g.}, BLEURT is only used for \\cref~task and can not support \\csrc~and \\csrcref~tasks.\n",
"Moreover, those approaches preserve the same core -- evaluating the quality of translation by referring to the given segments.\n",
"We believe that it is valuable, as well as feasible, to unify the capabilities of all MT evaluation tasks (\\cref, \\csrc~and \\csrcref) into one model.\n",
"Among the promising advantages are ease of use and improved robustness through knowledge transfer across evaluation tasks.\n",
"To achieve this, two important challenges need to be addressed:\n",
"1) How to design a model framework that can unify all translation evaluation tasks?\n",
"2) How to make the powerful PLMs better adapt to the unified evaluation model?\n",
"\n",
"In this paper, we propose \\textbf{\\cmodel} - \\textbf{Uni}fied \\textbf{T}ranslation \\textbf{E}valuation, a novel approach which unifies the functionalities of \\cref, \\csrc~and \\csrcref~tasks into one model.\n",
"To solve the first challenge as mentioned above, based on the multilingual PLM, we utilize layerwise coordination which concatenates all input segments into one sequence as the unified input form.\n",
"To further unify the modeling of three evaluation tasks, we propose a novel Monotonic Regional Attention (MRA) strategy, which allows partial semantic flows for a specific evaluation task. For the second challenge, a multi-task learning-based unified pretraining is proposed.\n",
"To be concrete, we collect the high-quality translations and degrade low-quality translations of NMT models as synthetic data.\n",
"Then we propose a novel ranking-based data labeling strategy to provide the training signal.\n",
"Finally, the multilingual PLM is continuously pretrained on synthetic dataset with multi-task learning manner.\n",
"Besides, our proposed models, named \\cmodel-MRA and \\cmodel-UP respectively, can benefit from finetuning with human-annotated data over three tasks at once, not requiring extra task-specific training.\n",
"\n",
"Experimental results demonstrate the superiority of \\cmodel.\n",
"Compared to various strong baseline systems on each task, \\cmodel, which unifies \\cref, \\csrc~and \\csrcref~tasks into one \\textit{single model}, achieves consistently absolute improvements of Kendall's $\\tau$ correlations at 1.1, 2.3 and 1.1 scores on English-targeted translation directions of WMT 2019 Metric Shared task~\\cite{fonseca2019findings}, respectively.\n",
"Meanwhile, after introducing multilingual-targeted support for our unified pretraining strategy, a single model named \\cmodel-MUP also gives dominant results against existing methods on non-English-targeted translation evaluation tasks. \n",
"Furthermore, our method can also achieve competitive results over WMT 2020 QE task compared with the winner submission~\\cite{ranasinghe2020transquest}.\n",
"Ablation studies reveal that, the proposed MRA and unified pretraining strategies are both important for model performance, making the model preserve the outstanding performance and multi-task transferability concurrently.\n",
"\n",
"\n",
"\\section{Related Work}\n",
"In this section, we briefly introduce the three directions of translation evaluation.\n",
"\n",
"\\subsection{Reference-Only Evaluation}\n",
"\\cref~assesses the translation quality via comparing the translation candidate and the given reference. \n",
"In this setting, the two inputs are written in the same language, thus being easily applied in most of the metric tasks. \n",
"In the early stages, statistical methods are dominant solutions due to their strengths in wide language support and intuitive design. \n",
"These methods measure the surface text similarity for a range of linguistic features, including n-gram~\\citep[BLEU,][]{papineni2002bleu}, token~\\citep[TER,][]{snover2006a}, and character~\\citep[ChrF \\& ChrF++,][]{popovic2015chrf,popovic2017chrf++}.\n",
"However, recent studies pointed out that these metrics have low consistency with human judgments and insufficiently evaluate high-qualified MT systems~\\cite{freitag2020bleu,rei2020comet,mathur2020tangled}.\n",
"\n",
"Consequently, with the rapid development of PLMs, researchers have been paying their attention to model-based approaches.\n",
"The basic idea of these studies is to collect sentence representations for similarity calculation~\\citep[BERTScore,][]{zhang2020bertscore} or evaluating probabilistic confidence~\\citetext{\\citealp[PRISM-ref,][]{thompson2020automatic}; \\citealp[BARTScore,][]{yuan2021bartscore}}. \n",
"To further improve the model, \\newcite{sellam2020bleurt} pretrained a specific PLM for the translation evaluation (BLEURT), while \\newcite{lo2019yisi} combined statistical and representative features (YiSi-1). Both these methods achieve higher correlations with human judgments than statistical counterparts.\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\\subsection{Source-Only Evaluation}\n",
"\\csrc , which also refers to quality estimation\\footnote{Refer to ``quality estimation'' or ``reference-free metric'' in WMT (\\href{http://www.statmt.org/wmt19/qe-task.html}{http://www.statmt.org/wmt19/qe-task.html}, \\href{http://www.statmt.org/wmt21/metrics-task.html}{http://www.statmt.org/wmt21/metrics-task.html}).}, is an important translation evaluation task especially for the scenario where the ground-truth reference is unavailable.\n",
"It takes the source-side sentence and the translation candidate as inputs for the quality estimation.\n",
"To achieve this, the methods are required to model cross-lingual semantic alignments. \n",
"Similar to reference-only evaluation, statistical-based~\\citep{ranasinghe2020transquest}, model-based~\\citetext{\\citealp[TransQuest,][]{ranasinghe2020transquest};~\\citealp[PRISM-src,][]{thompson2020automatic}}, and feature combination~\\citetext{\\citealp[YiSi-2,][]{lo2019yisi}} are typical and advanced methods in this tasks.\n",
"\n",
"\\subsection{Source-Reference-Combined Evaluation}\n",
"Aside from the above tasks that only consider either source or target side at one time, \\csrcref~takes both source and reference sentences into account. \n",
"In this way, methods in this context can evaluate the translation candidate via utilizing the features from both sides. \n",
"As a rising paradigm among translation evaluation tasks, \\csrcref~also profits from the development of cross-lingual PLMs. For example, \n",
"finetuning PLMs over human-annotated datasets~\\citep[COMET,][]{rei2020comet} achieves new state-of-the-art results among all evaluation approaches in WMT 2020~\\cite{mathur2020results}.\n",
"\n",
"\n",
"\\begin{figure*}\n",
" \\centering\n",
" \\scalebox{0.9}\n",
" {\n",
" \\begin{tikzpicture}\n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \\draw[rectangle, draw=orange, fill=orange!10, very thick] (-2.875, 0) rectangle node{} (2.875, 3.5);\n",
" \\draw[rectangle, color=orange, draw=orange!30, fill=orange!20] (-2.625, 0.25) rectangle node{Pretrained Language Model} (2.625, 1.0);\n",
" \\draw[rectangle, color=orange, draw=orange!30, fill=orange!20] (-2.625, 1.25) rectangle node{Pooling Layer} (2.625, 2.0);\n",
" \\draw[rectangle, color=orange, draw=orange!30, fill=orange!20] (-2.625, 2.25) rectangle node{Feedforward Network} (2.625, 3.0);\n",
" \n",
" \\draw[color=orange!80!black] (-2, 3.25) node{\\textbf{\\cmodel}};\n",
" \n",
" \\draw[rectangle, draw=gray, fill=gray!10, very thick] (-2.875, -1.5) rectangle node(){} (2.875, -0.25);\n",
" \n",
" \\draw[rectangle, draw=gray!30, fill=gray!20] (-2.625, -1.375) rectangle node(){\\small{$\\mathbf{x}_1$}} (-2.0, -0.875);\n",
" \n",
" \\draw[rectangle, draw=gray!30, fill=gray!20] (-1.75, -1.375) rectangle node(){\\small{$\\mathbf{x}_2$}} (-1.125, -0.875);\n",
" \n",
" \\draw[rectangle, draw=gray!30, fill=gray!20] (-0.875, -1.375) rectangle node(){\\small{$\\mathbf{x}_3$}} (-0.25, -0.875);\n",
" \n",
" \\draw[rectangle, draw=gray!30, fill=gray!20] (0.0, -1.375) rectangle node(){\\small{$\\mathbf{x}_4$}} (0.625, -0.875);\n",
" \n",
" \\draw (1.3125, -1.125) node{$\\cdots$};\n",
" \n",
" \\draw[rectangle, draw=gray!30, fill=gray!20] (2.0, -1.375) rectangle node(){\\small{$\\mathbf{x}_L$}} (2.625, -0.875);\n",
" \n",
" \\draw[color=gray!80!black] (-1.25, -0.5) node{\\textbf{Input Sequence}};\n",
" \n",
" \\draw[rectangle, color=violet!80!black, draw=violet!50, fill=violet!10, very thick] (-0.75, 3.75) rectangle node{\\textbf{Score}} (0.75, 4.25);\n",
" \n",
" \\draw [->, color=black!50, very thick](0, -0.25) -- (0, 0.0);\n",
" \n",
" \\draw [->, color=black!50, very thick](0, 3.5) -- (0, 3.75);\n",
" \n",
" \\draw [rectangle, draw=black!50, dashed, very thick] (-3.0, -1.75) rectangle node{} (3.0, 4.5);\n",
" \n",
" \\filldraw[draw=blue!30, fill=blue!10, very thick] (3.5, 4.0) -- (5.0, 4.0) -- (5.0, 4.25) -- (5.5, 3.75) -- (5.0, 3.25) -- (5.0, 3.5) -- (3.5, 3.5) -- cycle;\n",
" \\draw (4.5, 3.75) node{\\textsc{Ref}};\n",
" \n",
" \\filldraw[draw=red!30, fill=red!10, very thick] (3.5, 1.625) -- (5.0, 1.625) -- (5.0, 1.875) -- (5.5, 1.375) -- (5.0, 0.875) -- (5.0, 1.125) -- (3.5, 1.125) -- cycle;\n",
" \\draw (4.5, 1.375) node{\\textsc{Src}};\n",
" \n",
" \\filldraw[draw=green!30, fill=green!10, very thick] (3.5, -0.75) -- (5.0, -0.75) -- (5.0, -0.5) -- (5.5, -1.0) -- (5.0, -1.5) -- (5.0, -1.25) -- (3.5, -1.25) -- cycle;\n",
" \\draw (4.5, -1) node{\\textsc{Src+Ref}};\n",
" \n",
" \n",
" \\filldraw[draw=blue!30, fill=blue!10, very thick] ;\n",
" \\draw (4.5, 3.75) node{\\textsc{Ref}};\n",
" \n",
" \\draw[rectangle, draw=blue!50, very thick, dashed] (6.0, 2.5) rectangle node(){} (9.0, 4.5);\n",
" \n",
" \\draw[rectangle, draw=orange!10, fill=orange!5, very thick, dashed] (6.125, 3.0) rectangle node{} (8.875, 4.0);\n",
" \n",
" \\draw[rectangle, draw=gray!50, fill=gray!30, very thick] (6.125, 2.625) rectangle node{\\tiny{Hyp; Ref}}(8.875, 2.875);\n",
" \n",
" \\draw[rectangle, draw=violet!10, fill=violet!5, very thick, dashed] (7.125, 4.125) rectangle node{} (7.875, 4.375);\n",
" \n",
" \\draw[rectangle, color=orange, draw=orange!50, fill=orange!30, very thick] (6.25, 3.125) rectangle node{\\tiny{PLM}} (8.75, 3.375);\n",
" \n",
" \\foreach \\x in {0,...,3}\n",
" \\foreach \\y in {0,...,3}\n",
" {\n",
" \\draw[draw=gray!50, fill=white] (10+0.25*\\x, 3+0.25*\\y) circle (0.1);\n",
" }\n",
" \n",
" \\draw [color=brown, draw=brown!30, decorate, decoration = {brace}, very thick] (9.9, 3.875) -- (10.35, 3.875) node[pos=0.5, above=-0.025]{\\tiny{Hyp}};\n",
" \\draw [color=cyan, draw=cyan!30, decorate, decoration = {brace}, very thick] (10.4, 3.875) -- (10.85, 3.875) node[pos=0.5, above=0.025]{\\tiny{Ref}};\n",
" \n",
" \\draw [color=brown, draw=brown!30, decorate, decoration = {brace}, very thick] (9.875, 3.4) -- (9.875, 3.85) node[pos=0.5, left=0.025]{\\tiny{Hyp}};\n",
" \\draw [color=cyan, draw=cyan!30, decorate, decoration = {brace}, very thick] (9.875, 2.9) -- (9.875, 3.35) node[pos=0.5, left=0.025]{\\tiny{Ref}};\n",
" \n",
" \\draw [color=black] (12.5, 3.6) node{\\small{Full Attention}};\n",
" \n",
" \\draw[rectangle, draw=red!50, very thick, dashed] (6.0, 0.375) rectangle node(){} (9.0, 2.375);\n",
" \n",
" \\draw[rectangle, draw=orange!10, fill=orange!5, very thick, dashed] (6.125, 0.875) rectangle node{} (8.875, 1.875);\n",
" \n",
" \\draw[rectangle, draw=gray!50, fill=gray!30, very thick] (6.125, 0.5) rectangle node{\\tiny{Hyp; Src}}(8.875, 0.75);\n",
" \n",
" \\draw[rectangle, draw=violet!10, fill=violet!5, very thick, dashed] (7.125, 2.0) rectangle node{} (7.875, 2.25);\n",
" \n",
" \\draw [color=black!80] (10.625, 0.4) node{\\tiny{Key}};\n",
" \\draw [color=black!80] (9.25, -0.875) node{\\rotatebox{90}{\\tiny{Query}}};\n",
" \n",
" \n",
" \\draw[rectangle, color=orange, draw=orange!50, fill=orange!30, very thick] (6.25, 1.0) rectangle node{\\tiny{PLM}} (8.75, 1.25);\n",
" \n",
" \\foreach \\x in {0,...,3}\n",
" \\foreach \\y in {0,...,3}\n",
" {\n",
" \\draw[draw=gray!50, fill=white] (10+0.25*\\x, 0.875+0.25*\\y) circle (0.1);\n",
" }\n",
" \n",
" \\draw [color=brown, draw=brown!30, decorate, decoration = {brace}, very thick] (9.9, 1.75) -- (10.35, 1.75) node[pos=0.5, above=-0.025]{\\tiny{Hyp}};\n",
" \\draw [color=magenta, draw=magenta!30, decorate, decoration = {brace}, very thick] (10.4, 1.75) -- (10.85, 1.75) node[pos=0.5, above=0.025]{\\tiny{Src}};\n",
" \n",
" \\draw [color=brown, draw=brown!30, decorate, decoration = {brace}, very thick] (9.875, 1.275) -- (9.875, 1.725) node[pos=0.5, left=0.025]{\\tiny{Hyp}};\n",
" \\draw [color=magenta, draw=magenta!30, decorate, decoration = {brace}, very thick] (9.875, 0.775) -- (9.875, 1.225) node[pos=0.5, left=0.025]{\\tiny{Src}};\n",
" \n",
" \\draw [color=black] (12.5, 1.375) node{\\small{Full Attention}};\n",
" \n",
" \\draw[rectangle, draw=green!50, very thick, dashed] (6.0, -1.75) rectangle node(){} (9.0, 0.25);\n",
" \n",
" \\draw[rectangle, draw=orange!10, fill=orange!5, very thick, dashed] (6.125, -1.25) rectangle node{} (8.875, -0.25);\n",
" \n",
" \\draw[rectangle, draw=gray!50, fill=gray!30, very thick] (6.125, -1.625) rectangle node{\\tiny{Hyp; Src; Ref}}(8.875, -1.375);\n",
" \n",
" \\draw[rectangle, draw=violet!10, fill=violet!5, very thick, dashed] (7.125, -0.125) rectangle node{} (7.875, 0.125);\n",
" \n",
" \\draw[rectangle, color=orange, draw=orange!50, fill=orange!30, very thick] (6.25, -1.125) rectangle node{\\tiny{PLM}} (8.75, -0.875);\n",
" \n",
" \\foreach \\x in {0,...,5}\n",
" \\foreach \\y in {0,...,5}\n",
" {\n",
" \\ifthenelse{\\(\\x<2 \\AND \\y>1 \\AND \\y<4\\) \\OR \\(\\x<4 \\AND \\y<2\\)}\n",
" {\n",
" \\draw[draw=gray!50, fill=white] (10+0.25*\\x, -1.5+0.25*\\y) circle (0.1) node {\\textcolor{red}{\\tiny{\\ding{55}}}};\n",
" }\n",
" {\n",
" \\draw[draw=gray!50, fill=white] (10+0.25*\\x, -1.5+0.25*\\y) circle (0.1);\n",
" }\n",
" }\n",
" \n",
" \\draw [color=brown, draw=brown!30, decorate, decoration = {brace}, very thick] (9.9, -0.125) -- (10.35, -0.125) node[pos=0.5, above=-0.025]{\\tiny{Hyp}};\n",
" \\draw [color=magenta, draw=magenta!30, decorate, decoration = {brace}, very thick] (10.4, -0.125) -- (10.85, -0.125) node[pos=0.5, above=0.025]{\\tiny{Src}};\n",
" \\draw [color=cyan, draw=cyan!30, decorate, decoration = {brace}, very thick] (10.9, -0.125) -- (11.35, -0.125) node[pos=0.5, above=0.025]{\\tiny{Ref}};\n",
" \n",
" \\draw [color=brown, draw=brown!30, decorate, decoration = {brace}, very thick] (9.875, -0.6) -- (9.875, -0.15) node[pos=0.5, left=0.025]{\\tiny{Hyp}};\n",
" \\draw [color=magenta, draw=magenta!30, decorate, decoration = {brace}, very thick] (9.875, -1.1) -- (9.875, -0.65) node[pos=0.5, left=0.025]{\\tiny{Src}};\n",
" \\draw [color=cyan, draw=cyan!30, decorate, decoration = {brace}, very thick] (9.875, -1.6) -- (9.875, -1.15) node[pos=0.5, left=0.025]{\\tiny{Ref}};\n",
" \n",
" \\node[draw, color=black, draw=none, fill=none, align=left] at (12.5, -0.75) () {\\small{Monotonic} \\\\ \\small{Regional} \\\\ \\small{Attention}};\n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \\end{tikzpicture}\n",
" }\n",
" \\caption{Illustration of \\cmodel. Our model can give predictions for different data items formatted as \\fboxsep0.5ex\\cref, \\fboxsep0.5ex\\csrc, or \\fboxsep0.5ex\\csrcref~setting, unifying all evaluation tasks into one single model without additional modifications. For \\csrcref, we show the hard design for monotonic regional attention. \\textcolor{red}{\\ding{55}} denotes the masked attention logits.}\n",
" \\label{fig.architecture}\n",
"\\end{figure*}\n",
"\n",
"\n",
"\\section{Methodology}\n",
"As mentioned above, massive methods are proposed for different automatic evaluation tasks.\n",
"On the one hand, it is inconvenient and expensive to develop and employ different metrics for different evaluation scenarios.\n",
"On the other hand, separate models absolutely overlook the commonalities among these evaluation tasks, of which knowledge potentially benefits all three tasks.\n",
"In order to fulfill the aim of unifying the functionalities on \\cref, \\csrc, and \\csrcref~into one model, in this section, we introduce \\textbf{\\cmodel} (Figure~\\ref{fig.architecture}).\n",
"\n",
"\n",
"\\subsection{Model Architecture}\n",
"\\label{sec:model}\n",
"\n",
"By receiving a data example composing of hypothesis, source, and reference segment, \\cmodel~first modifies it into concatenated sequence following the given setting as \\cref, \\csrc, or \\csrcref:\n",
"\n",
"\\begin{align}\n",
" \\mathbf{x}_\\textsc{Ref} & = \\texttt{Concat} (\\mathbf{h}, \\mathbf{r}) \\in \\mathbb{R}^{(l_h + l_r)}, \\notag \\\\\n",
" \\mathbf{x}_\\textsc{Src} & = \\texttt{Concat}(\\mathbf{h}, \\mathbf{s}) \\in \\mathbb{R}^{(l_h + l_s)}, \\\\\n",
" \\mathbf{x}_\\textsc{Src+Ref} & = \\texttt{Concat}(\\mathbf{h}, \\mathbf{s}, \\mathbf{r}) \\in \\mathbb{R}^{(l_h + l_s + l_r)}, \\notag\n",
"\\end{align}\n",
"where $\\mathbf{h}$, $\\mathbf{s}$ and $\\mathbf{r}$ are hypothesis, source and reference segments, with the corresponding sequence lengths being $l_h$, $l_s$ and $l_r$, respectively.\n",
"The input sequence is then fed to PLM to derive representations $\\tilde{\\mathbf{H}}$.\n",
"Take \\cref~as an example:\n",
"\\begin{align}\n",
" \\tilde{\\mathbf{H}}_\\textsc{Ref} = \\texttt{PLM}(\\mathbf{x}_\\textsc{Ref}) \\in \\mathbb{R}^{(l_h + l_r) \\times d},\n",
"\\end{align}\n",
"where $d$ is the model size of PLM.\n",
"According to~\\newcite{ranasinghe2020transquest}, we use the first output representation as the input of feedforward layer. \n",
"\n",
"Compared to existing methods~\\cite{zhang2020bertscore,rei2020comet} which take sentence-level representations for evaluation, the advantages of our architecture design are as follows.\n",
"First, our \\cmodel~model can benefit from layer-coordinated semantical interactions inside every one of PLM layers, which is proven effective on capturing diverse linguistic features ~\\cite{he2018layer,lin2019open,jawahar2019what,tenney2019bert,rogers2020primer}.\n",
"Second, for the unified approach of our model, the concatenation provides the unifying format for all task inputs, turning our model into a more general architecture.\n",
"When conducting different evaluation tasks, our model requires no further modification inside.\n",
"Note here, to keep the consistency across all evaluation tasks, as well as ease the unified learning, $\\mathbf{h}$ is always located at the beginning of the input sequence.\n",
"\n",
"After deriving $\\tilde{\\mathbf{H}}_{\\textsc{Ref}}$, a pooling block is arranged after PLM which gives sequence-level representations $\\mathbf{H}_{\\textsc{Ref}}$.\n",
"Finally, a feedforward network takes $\\mathbf{H}_{\\textsc{Ref}}$ as input, and gives a scalar $p$ as prediction:\n",
"\\begin{align}\n",
" \\mathbf{H}_\\textsc{Ref} & = \\texttt{Pool}(\\tilde{\\mathbf{H}}_\\textsc{Ref}) \\in \\mathbb{R}^{d}, \\\\\n",
" p_\\textsc{Ref} & = \\texttt{FeedForward}(\\mathbf{H}_\\textsc{Ref}) \\in \\mathbb{R}^{1}.\n",
"\\end{align}\n",
"For training, we encourage the model to reduce the mean squared error with respect to given score $q$:\n",
"\\begin{align}\n",
" \\mathcal{L}_\\textsc{Ref} = (p_\\textsc{Ref} - q) ^ 2.\n",
"\\end{align}\n",
"\n",
"\n",
"However, for the pretraining of most PLMs~\\citep[e,g., XLM-R,][]{conneau2020unsupervised}, the input patterns are designed to receive two segments at most.\n",
"Thus there exists a gap between the pretraining of PLM and the joint training of \\cmodel~where the concatenation of three fragments is used as input.\n",
"Moreover, previous study~\\cite{takahashi2020automatic} shows that directly training over \\csrcref~by following such design leads to worse performance than \\cref~scenario.\n",
"To alleviate this issue, we propose two strategies: \\textbf{Monotonic Regional Attention} as described in~\\S\\ref{sec:mra} and \\textbf{Unified Pretraining} in~\\S\\ref{sec:up}.\n",
"\n",
"\\subsection{Monotonic Regional Attention}\n",
"\\label{sec:mra}\n",
"To fill the modeling gap between the pretraining of PLM and the joint training of three downstream tasks, a natural idea is to unify the number of involved segments when modeling semantics for \\csrc, \\cref~and \\csrcref~tasks.\n",
"\n",
"Following this, we propose to modify the attention mask of \\csrcref~to simulate the modeling of two segments in \\csrc~and \\cref.\n",
"Specifically, when calculating the attention logits, semantics from a specific segment are only allowed to derive information from two segments at most.\n",
"Considering the conventional attention module:\n",
"\\begin{align}\n",
" \\mathbf{A} = \\mathrm{Softmax}(\\frac{\\mathbf{Q}\\mathbf{K}^\\top}{\\sqrt{d}}) \\in \\mathbb{R}^{L \\times L},\n",
"\\end{align}\n",
"where $L$ is the sequential length for input, $\\mathbf{Q}, \\mathbf{K} \\in \\mathbb{R}^{L \\times d}$ are query and key representations, respectively.\\footnote{For simplicity, we omit the multi-head mechanism.}\n",
"As to monotonic regional attention (MRA), we simply add a mask $\\mathbf{M}$ to the softmax logits to control attention flows:\n",
"\\begin{align}\n",
" \\mathbf{A} & = \\mathrm{Softmax}(\\frac{\\mathbf{Q}\\mathbf{K}^\\top}{\\sqrt{d}} + \\mathbf{M}) \\in \\mathbb{R}^{L \\times L}, \\\\\n",
" \\mathbf{M}_{ij} & = \n",
" \\begin{cases}\n",
" -\\infty & (i,j) \\in \\mathbf{U}, \\\\\n",
" 0 & \\text{otherwise},\n",
" \\end{cases}\n",
"\\end{align}\n",
"where $\\mathbf{U}$ stores the index pairs of all masked areas. \n",
"\n",
"Following this idea, the key of MRA is how to design the matrix $\\mathbf{U}$.\n",
"For the cases where interactions inside each segment, we believe that these self-interactions are beneficial to the modeling.\n",
"For other cases where interactions are arranged across segments, three patterns are included: hypothesis-reference, source-reference, and hypothesis-source.\n",
"Intuitively, the former two parts are beneficial for model training, since they might contribute the monolingual signals and cross-lingual disambiguation to evaluation, respectively.\n",
"\n",
"This leaves the only case, where our experimental analysis also verifies (see \\S\\ref{sec.ablation.mra}), that interaction between hypothesis and source leads to the performance decrease for \\csrcref~task, thus troubling the unifying.\n",
"\n",
"\\tikzset{\n",
"semi/.style={\n",
" semicircle,\n",
" draw,\n",
" }\n",
"}\n",
"\n",
"\\begin{figure}\n",
" \n",
" \\centering\n",
" \\scalebox{0.7}{\n",
" \\begin{tikzpicture}\n",
" {\n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \\node[fill=brown!20, draw=brown!50, circle, minimum size=30pt] (hyp) at (0, 0) {$\\mathbf{h}$};\n",
" \\node[fill=magenta!20, draw=magenta!50, circle, minimum size=30pt] (src) at (-1.25, -1.5) {$\\mathbf{s}$};\n",
" \\node[fill=cyan!20, draw=cyan!50, circle, minimum size=30pt] (ref) at (1.25, -1.5) {$\\mathbf{r}$};\n",
" \n",
" \\draw[brown!80, thick, ->] (0.35, 0.45) arc (-30:210:0.4);\n",
" \\draw[magenta!80, thick, ->] (-1.7, -1.15) arc (60:300:0.4);\n",
" \\draw[cyan!80, thick, ->] (1.7, -1.15) arc (120:-120:0.4);\n",
" \n",
" \\draw[cyan!80, thick, ->] (0.65, -1.5) -- (-0.65, -1.5);\n",
" \\draw[cyan!80, thick, ->] (0.95, -1.0) -- (0.45, -0.35);\n",
" \\draw[magenta!80, thick, ->] (-0.95, -1.0) -- (-0.45, -0.35);\n",
" \n",
" \n",
" \n",
" \\node[fill=brown!20, draw=brown!50, circle, minimum size=30pt] (hyp) at (6, 0) {$\\mathbf{h}$};\n",
" \\node[fill=magenta!20, draw=magenta!50, circle, minimum size=30pt] (src) at (4.75, -1.5) {$\\mathbf{s}$};\n",
" \\node[fill=cyan!20, draw=cyan!50, circle, minimum size=30pt] (ref) at (7.25, -1.5) {$\\mathbf{r}$};\n",
" \n",
" \\draw[brown!80, thick, ->] (6.35, 0.45) arc (-30:210:0.4);\n",
" \\draw[magenta!80, thick, ->] (4.3, -1.15) arc (60:300:0.4);\n",
" \\draw[cyan!80, thick, ->] (7.7, -1.15) arc (120:-120:0.4);\n",
" \n",
" \\draw[cyan!80, thick, ->] (6.65, -1.45) -- (5.35, -1.45);\n",
" \\draw[magenta!80, thick, ->] (5.35, -1.55) -- (6.65, -1.55);\n",
" \n",
" \\draw[cyan!80, thick, ->] (6.98, -0.95) -- (6.48, -0.3);\n",
" \\draw[brown!80, thick, ->] (6.42, -0.4) -- (6.92, -1.05);\n",
" \n",
" \\draw[magenta!80, thick, ->] (5.05, -1.0) -- (5.55, -0.35);\n",
" }\n",
" \n",
" \\end{tikzpicture}\n",
" }\n",
" \n",
" \\caption{Attention flows in monotonic regional attention. $\\mathbf{h}$, $\\mathbf{s}$ and $\\mathbf{r}$ are hypothesis, source and reference, respectively. We prevent specified interactions in \\fboxsep0.5ex\\colorbox{green!10}{\\textsc{Src+Ref}}~training via modifying the attention mask with regional properties. We show the hard (left) and soft design (right, no $\\mathbf{h}\\rightarrow\\mathbf{s}$) in this figure.}\n",
" \\label{fig.monotonic}\n",
"\\end{figure}\n",
"To give more fine-grained designs, we propose two approaches for \\cmodel-MRA, which apply the MRA mechanism into \\cmodel~model (Figure~\\ref{fig.monotonic}):\n",
"\n",
"\\begin{itemize}\n",
" \\item Hard MRA. Only monotonic attention flows are allowed. Interactions between any two segments are strictly unidirectional through the entire PLM, where\n",
" $\\mathbf{U}$ stores the index pairs of unidirectional interactions of $\\mathbf{h}\\rightarrow\\mathbf{r}$, $\\mathbf{s}\\rightarrow\\mathbf{r}$ and $\\mathbf{h}\\rightarrow\\mathbf{s}$, where ``$\\rightarrow$'' denotes the direction of attention flows.\n",
" \\item Soft MRA. Specific attention flows are forbidden inside each attention module.\n",
" The involved two segments may interact inside a higher layer.\n",
" In practice, index pairs which denoting $\\mathbf{h}\\rightarrow\\mathbf{s}$ or $\\mathbf{s}\\rightarrow\\mathbf{h}$ between source and hypothesis are stored in $\\mathbf{U}$.\n",
"\\end{itemize}\n",
"\n",
"Note that, although the processing in source and reference may be affected because their positions are not indexed from the start, related studies on positional embeddings reveal that, PLM can well capture relative positional information~\\cite{wang2020what}, which dispels this concern.\n",
"\n",
"\\subsection{Unified Pretraining}\n",
"\\label{sec:up}\n",
"To further bridge the modeling gap between PLM and the joint training of \\cmodel~mentioned in~\\S\\ref{sec:model}, we propose a unified pretraining strategy including the following main stages: 1) collecting and downgrading synthetic data; 2) labeling examples with a novel ranking-based strategy; 3) multi-task learning for unified pretraining and finetuning.\n",
"\n",
"\\paragraph{Synthetic Data Collection}\n",
"As our approach aims at evaluating the quality of translations, generated hypotheses with NMT models are ideal synthetic data.\n",
"To further improve the diversity of synthetic data quality,\n",
"we follow existing experiences~\\cite{sellam2020bleurt,wan2021robleurt} to apply the word and span dropping strategy to downgrade a portion of hypotheses.\n",
"The collected data totally contains $N$ triplets composing of hypothesis, source and reference segments, which is formed as $\\mathcal{D}' = \\{\\langle\\mathbf{h}^i, \\mathbf{s}^i, \\mathbf{r}^i\\rangle\\}_{i=1}^{N}$.\n",
"\n",
"\\paragraph{Data Labeling}\n",
"After obtaining the synthetic data, the next step is to augment each data pair with a label which serves as the signal of unified pretraining.\n",
"To stabilize the model training, as well as normalize the distributions across all score systems and languages, we propose a novel ranking-based approach.\n",
"This method is based on the idea of Borda count~\\cite{ho1994decision,emerson2013original}, which provides more precise and well-distributed synthetic data labels than Z-score normalization.\n",
"\n",
"Specifically, we first use available approaches to derive the predicted score $\\hat{q}^i$ for each item, yielding labeled synthetic quadruple examples formed as $\\mathcal{D}'' = \\{\\langle\\mathbf{h}^{i}, \\mathbf{s}^{i}, \\mathbf{r}^{i}, \\hat{q}^i\\rangle\\}_{i=1}^{N}$.\n",
"Then, we tag each example with its rank index $\\tilde{q}^i$ referring to $\\hat{q}^i$:\n",
"\\begin{align}\n",
" \\tilde{q}^i = \\texttt{IndexOf}(\\hat{q}^i, \\mathcal{Q}),\n",
"\\end{align}\n",
"where $\\mathcal{Q}$ is the list storing all the sorted $\\hat{q}^i$ descendingly.\n",
"Then, we use the conventional Z-score strategy to normalize the scores:\n",
"\\begin{align}\n",
" q^i = \\frac{\\tilde{q}^i - \\mu}\\sigma,\n",
"\\end{align}\n",
"where $\\mu$ and $\\sigma$ are the mean and the standard deviation of values in $\\mathcal{Q}$, respectively.\n",
"The dataset thus updates its format to $\\mathcal{D} = \\{\\langle\\mathbf{h}^i, \\mathbf{s}^i, \\mathbf{r}^i, q^i\\rangle\\}_{i=1}^{N}$.\n",
"Note here that, an example with higher $\\hat{q}^i$ is assigned with higher $\\tilde{q}^i$, thus a larger value of $q^i$.\n",
"\n",
"Compared to related approaches which apply Z-score normalization~\\cite{bojar2018findings}, or leave the conventional labeled scores as signals for learning~\\citep[\\textit{i.e.}, knowledge distillation,][]{kim2016sequence,phuong2019towards}, our approach can alleviate the bias of chosen model for labeling and prior distributional disagreement of scores.\n",
"For example, different methods may give scores with different distributions.\n",
"Especially for translation directions of low-resource, scores may follow skewed distribution~\\cite{sellam2020bleurt}, which has a disagreement with rich-resource scenarios.\n",
"Our method can unify the distribution of all labeling data into the same scale, which can also be easily applied by the ensembling strategy.\n",
"\n",
"\n",
"\\paragraph{Multi-task Pretrainig and Finetuning}\n",
"To unify all evaluation scenarios into one model, we apply multi-task learning for both pretraining and finetuning.\n",
"For each step, we arrange three substeps for all input formats, yielding $\\mathcal{L}_\\textsc{Ref}$, $\\mathcal{L}_\\textsc{Src}$, and $\\mathcal{L}_\\textsc{Src+Ref}$, respectively.\n",
"The final learning objective is to reduce the summation of all losses:\n",
"\\begin{align}\n",
" \\mathcal{L} = \\mathcal{L}_\\textsc{Ref} + \\mathcal{L}_\\textsc{Src} + \\mathcal{L}_\\textsc{Src+Ref}.\n",
"\\end{align}\n",
"\n",
"\n",
"\\begin{table*}[t]\n",
" \\small\n",
" \\centering\n",
" \\scalebox{1.0}\n",
" \n",
" {\n",
" \\begin{tabular}{lcccccccc}\n",
" \\toprule\n",
" \n",
" \\multirow{2}{*}{\\textbf{Model}} & \\multicolumn{4}{c}{\\textbf{High-resource}} & \\multicolumn{3}{c}{\\textbf{Zero-shot}} & \\multirow{2}{*}{\\textbf{Avg.}} \\\\\n",
" \\cmidrule(l{2pt}r{2pt}){2-5}\\cmidrule(l{2pt}r{2pt}){6-8}\n",
" & \\textit{\\underline{De-En}} & \\textit{\\underline{Ru-En}} & \\textit{\\underline{Zh-En}} & \\underline{Fi-En} & Gu-En & Kk-En & Lt-En & \\\\\n",
" \n",
" \n",
" \n",
" \\midrule\n",
" \n",
" \\multicolumn{9}{c}{\\textit{Reference-only Evaluation}} \\\\\n",
" \\cdashline{1-9}\\noalign{\\vskip 0.1ex}\n",
" \n",
" $^\\heartsuit$BLEU~\\cite{papineni2002bleu} & ~~5.4 & 11.5 & 32.1 & 23.6 & 19.4 & 27.6 & 24.9 & \\colorbox{blue!10}{20.6} \\\\\n",
" $^\\spadesuit$ChrF~\\cite{popovic2015chrf} & 12.3 & 17.7 & 37.1 & 29.2 & 24.0 & 32.3 & 30.4 & \\colorbox{blue!10}{26.1} \\\\\n",
" $^\\heartsuit$BERTScore~\\cite{zhang2020bertscore} & 19.0 & 22.1 & 43.0 & 35.4 & 29.2 & 35.1 & 38.1 & \\colorbox{blue!10}{31.7} \\\\\n",
" $^\\heartsuit$BLEURT~\\cite{sellam2020bleurt} & 17.4 & 22.0 & 43.6 & 37.4 & 31.3 & 37.2 & 38.8 & \\colorbox{blue!10}{32.5} \\\\\n",
" $^\\spadesuit$YiSi-1~\\cite{lo2019yisi} & 16.4 & 21.7 & 42.6 & 34.7 & 31.2 & \\textbf{44.0} & 37.6 & \\colorbox{blue!10}{32.6} \\\\\n",
" $^\\heartsuit$PRISM-ref~\\cite{thompson2020automatic} & 20.4 & \\textbf{22.5} & 43.8 & 35.7 & 31.3 & 43.4 & 38.2 & \\colorbox{blue!10}{33.6} \\\\\n",
" $^\\heartsuit$BARTScore~\\cite{yuan2021bartscore} & 23.8 & 21.9 & 44.7 & 37.4 & 31.8 & 37.6 & 38.6 & \\colorbox{blue!10}{33.7} \\\\\n",
" $^\\diamondsuit$XLM-R+Concat~\\cite{takahashi2020automatic} & 24.5 & 21.8 & 45.8 & 37.0 & 31.5 & 37.4 & 39.5 & \\colorbox{blue!10}{33.9} \\\\\n",
" $^\\diamondsuit$RoBERTa+Concat~\\cite{takahashi2020automatic} & 25.1 & 22.4 & 46.4 & 36.2 & 30.8 & 38.0 & \\textbf{40.0} & \\colorbox{blue!10}{34.1} \\\\\n",
" \\cdashline{1-9}\\noalign{\\vskip 0.1ex}\n",
" \n",
" \\cmodel-MRA & 25.2 & 22.4 & 46.4 & 36.5 & 31.6 & 38.4 & 39.1 & \\colorbox{blue!10}{34.2} \\\\\n",
" \\cmodel-UP & \\textbf{25.9} & 21.9 & \\textbf{46.7} & \\textbf{37.9} & \\textbf{32.2} & 38.7 & \\textbf{40.0} & \\colorbox{blue!10}{\\textbf{34.8}} \\\\\n",
" \n",
" \\midrule\n",
" \n",
" \\multicolumn{9}{c}{\\textit{Source-only Evaluation}} \\\\\n",
" \\cdashline{1-9}\\noalign{\\vskip 0.1ex}\n",
" \n",
" $^\\spadesuit$YiSi-2~\\cite{lo2019yisi} & ~~6.8 & ~~5.3 & 25.3 & 12.6 & ~-0.1 & ~~9.6 & ~~7.5 & \\colorbox{red!10}{~~9.5} \\\\\n",
" $^\\heartsuit$PRISM-src~\\cite{thompson2020automatic} & 10.9 & \\textbf{17.8} & 33.6 & 30.0 & 10.2 & \\textbf{39.1} & 35.6 & \\colorbox{red!10}{25.3} \\\\\n",
" $^\\heartsuit$MTransQuest~\\cite{ranasinghe2020transquest} & 11.1 & 14.0 & 32.1 & 29.7 & 27.2 & 31.6 & 30.7 & \\colorbox{red!10}{25.2} \\\\\n",
" $^\\diamondsuit$MTransQuest~\\cite{ranasinghe2020transquest} & 17.0 & 17.3 & 37.6 & 29.2 & 26.5 & 31.9 & 34.2 & \\colorbox{red!10}{27.7} \\\\\n",
" $^\\diamondsuit$XLM-R+Concat~\\cite{takahashi2020automatic} & 16.9 & 17.6 & 38.1 & 29.1 & 26.2 & 31.6 & 34.3 & \\colorbox{red!10}{27.7} \\\\\n",
" \\cdashline{1-9}\\noalign{\\vskip 0.1ex}\n",
" \n",
" \\cmodel-MRA & 17.4 & 17.7 & 41.0 & \\textbf{34.3} & 29.0 & 32.7 & \\textbf{36.2} & \\colorbox{red!10}{29.7} \\\\\n",
" \\cmodel-UP & \\textbf{19.3} & 16.9 & \\textbf{41.4} & 34.0 & \\textbf{29.7} & 33.6 & 35.4 & \\colorbox{red!10}{\\textbf{30.0}} \\\\\n",
" \n",
" \\midrule\n",
" \n",
" \\multicolumn{9}{c}{\\textit{Source-Reference-Combined Evaluation}} \\\\\n",
" \\cdashline{1-9}\\noalign{\\vskip 0.1ex}\n",
" \n",
" \n",
" \n",
" $^\\diamondsuit$XLM-R+Concat~\\cite{takahashi2020automatic} & 24.0 & 22.0 & 44.7 & 35.7 & 30.4 & 37.2 & 38.9 & \\colorbox{green!10}{33.4} \\\\\n",
" $^\\diamondsuit$COMET~\\cite{rei2020comet} & 23.4 & 20.7 & 45.8 & 36.2 & 30.9 & 37.9 & 40.3 & \\colorbox{green!10}{33.6} \\\\\n",
" \\cdashline{1-9}\\noalign{\\vskip 0.1ex}\n",
" \n",
" \\cmodel-MRA & 25.6 & \\textbf{22.9} & 46.9 & 37.6 & 31.6 & 38.5 & \\textbf{40.5} & \\colorbox{green!10}{34.8} \\\\\n",
" \\cmodel-UP & \\textbf{26.0} & 22.0 & \\textbf{47.2} & \\textbf{37.7} & \\textbf{32.3} & \\textbf{39.4} & 40.0 & \\colorbox{green!10}{\\textbf{35.0}} \\\\\n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \\bottomrule\n",
" \\end{tabular}\n",
" }\n",
" \\caption{Kendall's Tau correlation (\\\n",
" \\label{table.main_results_metric_x-en}\n",
"\\end{table*}\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
" \n",
"\n",
"\n",
" \n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
" \n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
" \n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
" \n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
" \n",
"\n",
"\n",
"\n",
"\n",
"\n",
" \n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
" \n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
" \n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
" \n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
" \n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\\begin{table*}[t]\n",
" \n",
" \\centering\n",
" \n",
" \\scalebox{0.675}{\n",
" \\begin{tabular}{lcccccccccccc}\n",
" \\toprule\n",
" \n",
" \n",
" \n",
" \\multirow{2}{*}{\\textbf{Model}} & \\multicolumn{5}{c}{\\textbf{High-resource}} & \\multicolumn{6}{c}{\\textbf{Zero-shot}} & \\multirow{2}{*}{\\textbf{Avg.}} \\\\\n",
" \\cmidrule(l{2pt}r{2pt}){2-6}\\cmidrule(l{2pt}r{2pt}){7-12}\n",
" & \\textit{\\underline{En-Cs}} & \\textit{\\underline{En-De}} & \\textit{\\underline{En-Ru}} & \\textit{\\underline{En-Zh}} & \\underline{En-Fi} & En-Gu & En-Kk & En-Lt & De-Cs & De-Fr & Fr-De & \\\\\n",
" \n",
" \\midrule\n",
" \n",
" \\multicolumn{13}{c}{\\textit{Reference-only Evaluation}} \\\\\n",
" \\cdashline{1-13}\\noalign{\\vskip 0.1ex}\n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" $^\\heartsuit$BLEU~\\cite{papineni2002bleu} & 36.4 & 24.8 & 46.9 & 23.5 & 39.5 & 46.3 & 36.3 & 33.3 & 22.2 & 22.6 & 17.3 & \\colorbox{blue!10}{31.7} \\\\\n",
" $^\\spadesuit$ChrF~\\cite{popovic2015chrf} & 44.4 & 32.1 & 54.8 & 24.1 & 51.8 & 54.8 & 51.0 & 43.8 & 34.1 & 28.7 & 27.4 & \\colorbox{blue!10}{40.6} \\\\\n",
" $^\\heartsuit$BERTScore~\\cite{zhang2020bertscore} & 50.0 & 36.3 & \\textbf{58.5} & 35.6 & 52.7 & 56.8 & 54.0 & 46.4 & 35.8 & 32.9 & 30.0 & \\colorbox{blue!10}{44.5} \\\\\n",
" $^\\spadesuit$YiSi-1~\\cite{lo2019yisi} & 47.5 & 35.1 & \\textbf{58.5} & 35.5 & 53.7 & 55.1 & 54.6 & 47.0 & 37.6 & 34.9 & 31.0 & \\colorbox{blue!10}{44.6} \\\\\n",
" $^\\spadesuit$BLEURT~\\cite{sellam2020learning} & 60.3 & 42.2 & 49.2 & 33.7 & 61.5 & 57.7 & 55.8 & 58.4 & 46.1 & 44.9 & \\textbf{42.7} & \\colorbox{blue!10}{50.2} \\\\\n",
" $^\\diamondsuit$XLM-R+Concat~\\cite{takahashi2020automatic} & 60.2 & 43.0 & 58.1 & 41.0 & 60.2 & 60.8 & 60.1 & 58.8 & 47.0 & 45.1 & 40.9 & \\colorbox{blue!10}{52.3} \\\\\n",
" \n",
" \\cdashline{1-13}\\noalign{\\vskip 0.1ex}\n",
" \n",
" \\cmodel-UP & 60.1 & 44.4 & 50.7 & \\textbf{45.3} & 62.2 & 62.1 & 61.1 & \\textbf{61.5} & 48.3 & \\textbf{47.3} & 42.3 & \\colorbox{blue!10}{53.2} \\\\\n",
" \\cmodel-MUP & \\textbf{62.1} & \\textbf{45.6} & 52.2 & 44.8 & \\textbf{62.5} & \\textbf{63.0} & \\textbf{61.9} & 61.4 & \\textbf{49.1} & 46.9 & 42.3 & \\colorbox{blue!10}{\\textbf{53.8}} \\\\\n",
" \n",
" \\midrule\n",
" \n",
" \\multicolumn{13}{c}{\\textit{Source-only Evaluation}} \\\\\n",
" \\cdashline{1-13}\\noalign{\\vskip 0.1ex}\n",
" \n",
" $^\\heartsuit$MTransQuest~\\cite{ranasinghe2020transquest} & 35.8 & 28.4 & 31.1 & 29.0 & 50.8 & 52.7 & 56.3 & 43.9 & 35.7 & 23.7 & ~~9.4 & \\colorbox{red!10}{36.1} \\\\\n",
" $^\\diamondsuit$MTransQuest~\\cite{ranasinghe2020transquest} & 40.2 & 33.1 & \\textbf{32.9} & 32.8 & 54.2 & 57.2 & 60.2 & 49.1 & 40.4 & 29.8 & 17.9 & \\colorbox{red!10}{40.7} \\\\\n",
" $^\\diamondsuit$XLM-R+Concat~\\cite{takahashi2020automatic} & 53.5 & 38.0 & 30.2 & 34.0 & 53.9 & 55.9 & 53.5 & 53.8 & 35.7 & 32.5 & 31.5 & \\colorbox{red!10}{42.9} \\\\\n",
" \n",
" \\cdashline{1-13}\\noalign{\\vskip 0.1ex}\n",
" \n",
" \\cmodel-UP & 52.3 & 41.7 & 27.3 & \\textbf{40.7} & 60.7 & 59.1 & 60.4 & 56.8 & 40.7 & \\textbf{37.0} & 32.1 & \\colorbox{red!10}{46.3} \\\\\n",
" \\cmodel-MUP & \\textbf{55.9} & \\textbf{43.8} & 28.7 & 40.6 & \\textbf{61.9} & \\textbf{60.5} & \\textbf{61.1} & \\textbf{59.3} & \\textbf{41.4} & 35.6 & \\textbf{36.7} & \\colorbox{red!10}{\\textbf{47.8}} \\\\\n",
" \n",
" \\midrule\n",
" \n",
" \\multicolumn{13}{c}{\\textit{Source-Reference-Combined Evaluation}} \\\\\n",
" \\cdashline{1-13}\\noalign{\\vskip 0.1ex}\n",
" \n",
" $^\\diamondsuit$XLM-R+Concat~\\cite{takahashi2020automatic} & 60.9 & 43.3 & 53.3 & 40.8 & 60.4 & 60.1 & 59.1 & 59.3 & 46.4 & 44.9 & 40.5 & \\colorbox{green!10}{51.7} \\\\\n",
" $^\\diamondsuit$COMET~\\cite{rei2020comet} & 61.0 & 44.6 & \\textbf{58.3} & 42.3 & 62.3 & 60.7 & 59.0 & 60.6 & 45.7 & 46.8 & 38.8 & \\colorbox{green!10}{52.7} \\\\\n",
" \n",
" \\cdashline{1-13}\\noalign{\\vskip 0.1ex}\n",
" \n",
" \\cmodel-UP & 60.0 & 44.9 & 49.7 & \\textbf{45.6} & 62.7 & 62.6 & 62.0 & 61.0 & 48.0 & 45.5 & \\textbf{42.4} & \\colorbox{green!10}{53.1} \\\\\n",
" \\cmodel-MUP & \\textbf{62.2} & \\textbf{46.0} & 54.6 & 44.9 & \\textbf{63.2} & \\textbf{63.2} & \\textbf{63.0} & \\textbf{61.8} & \\textbf{48.7} & \\textbf{47.5} & \\textbf{42.4} & \\colorbox{green!10}{\\textbf{54.3}} \\\\\n",
"\n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \\bottomrule\n",
" \\end{tabular}\n",
" }\n",
" \\caption{Kendall's Tau correlation (\\\n",
" \n",
" }\n",
" \\label{table.main_results_metric_en-x}\n",
"\\end{table*}\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
" \n",
"\n",
"\n",
" \n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
" \n",
"\n",
"\n",
"\n",
"\n",
"\n"
],
"context_after_exp": [
"\\section{Experiments}\n",
"\\subsection{Experimental Settings}\n",
"\\paragraph{Benchmarks} Following~\\newcite{rei2020comet, yuan2021bartscore}, we examine the effectiveness of the propose method on WMT 2019 Metrics~\\cite{ma2019results}.\n",
"For the former, we follow the common practice in COMET\\footnote{\\href{https://github.com/Unbabel/COMET}{https://github.com/Unbabel/COMET}}~\\cite{rei2020comet} to collect and preprocess the dataset.\n",
"The official variant of Kendall's Tau correlation~\\cite{ma2019results} is used for evaluation.\n",
"We evaluate our methods on all of \\cref, \\csrc~and \\csrcref~scenarios.\n",
"For \\csrc~scenario, we further conduct results on WMT 2020 QE task~\\cite{specia2020findings} referring to~\\newcite{ranasinghe2020intelligent} for data collection and preprocessing.\n",
"Following the official report, the Pearson's correlation is used for evaluation.\n",
"\n",
"\\paragraph{Model Pretraining}\n",
"As mentioned in \\S\\ref{sec:up}, we continuously pretrain PLMs using synthetic data.\n",
"The data is constructed from WMT 2021 News Translation task, where we collect the training sets from five translation tasks.\n",
"Among those tasks, the target sentences are all in English (En), and the source languages are Czech (Cs), German (De), Japanese (Ja), Russian (Ru), and Chinese (Zh).\n",
"Specifically, we follow~\\newcite{sellam2020bleurt} to use \\textsc{Transformer}-base~\\cite{vaswani2017attention} MT models to generate translation candidates, and use the checkpoints trained via \\cmodel-MRA approach for synthetic data labeling.\n",
"We pretrain two kinds of models, one is pretrained on English-targeted language directions, and the other is a multilingual version trained using bidirectional data.\n",
"Note that, for a fair comparison, we filter out all pretraining examples that are involved in benchmarks.\n",
"\n",
"\\paragraph{Model Setting}\n",
"We implement our approach upon COMET~\\cite{rei2020comet} repository and follow their work to choose XLM-R~\\cite{conneau2020unsupervised} as the PLM.\n",
"The feedforward network consists of 3 linear transitions, where the dimensionalities of corresponding outputs are 3,072, 1,024, and 1, respectively.\n",
"Between any two adjacent linear modules inside, hyperbolic tangent function is arranged as activation.\n",
"During both pretraining and finetuning phrases, we divided training examples into three sets, where each set only serves one scenario among \\cref, \\csrc~and \\csrcref~to avoid learning degeneration.\n",
"During finetuning, we randomly extracting 2,000 training examples from benchmarks as development set.\n",
"Besides \\cmodel-MRA and \\cmodel-UP which are derived with MRA (\\S~\\ref{sec:mra}) and Unified Pretraining (\\S~\\ref{sec:up}), we also extend the latter with multilingual-targeted unified pretraining, thus obtaining \\cmodel-MUP model.\n",
"\n",
"\n",
"\\paragraph{Baselines}\n",
"As to \\cref~approaches, we select BLEU~\\cite{papineni2002bleu}, ChrF~\\cite{popovic2015chrf}, YiSi-1~\\cite{lo2019yisi}, BERTScore~\\cite{zhang2020bertscore}, BLEURT~\\cite{sellam2020bleurt}, PRISM-ref~\\cite{thompson2020automatic}, BARTScore~\\cite{yuan2021bartscore}, XLM-R+Concat~\\cite{takahashi2020automatic}, and RoBERTa+Concat~\\cite{takahashi2020automatic} for comparison.\n",
"For \\csrc~methods, we post results of both metric and QE methods, including YiSi-2~\\cite{lo2019yisi}, XLM-R+Concat~\\cite{takahashi2020automatic}, PRISM-src~\\cite{thompson2020automatic} and multilingual-to-multilingual MTransQuest~\\cite{ranasinghe2020transquest}.\n",
"For \\csrcref, we use XLM-R+Concat~\\cite{takahashi2020automatic} and COMET~\\cite{rei2020comet} as strong baselines.\n",
"\n",
"\n",
"\n",
"\\subsection{Main Results}\n",
"\\paragraph{English-Targeted} Results on English-targeted metric task are conducted in Table~\\ref{table.main_results_metric_x-en}.\n",
"Among all involved baselines, for \\cref~methods, BARTScore~\\cite{yuan2021bartscore} performs better than other statistical and model-based metrics.\n",
"As to \\csrc~scenario, MTransQuest~\\cite{ranasinghe2020transquest} gives dominant performance.\n",
"Further, COMET~\\cite{rei2020comet} performs better than XLM-R+Concat~\\cite{takahashi2020automatic} on \\csrcref~scenario.\n",
"\n",
"As for our methods, we can see that, \\cmodel-MRA achieves better results on all tasks, demonstrating the effectiveness of monotonic attention flows for cross-lingual interactions.\n",
"Moreover, the proposed model \\cmodel-UP, which unifies \\cref, \\csrc, and \\csrcref~learning on both pretraining and finetuning, yields better results on all evaluation settings.\n",
"Most importantly, \\cmodel-UP is a \\textit{single model} which surpasses all the different state-of-the-art models on three tasks, showing its dominance on both convenience and effectiveness.\n",
"\n",
"\\begin{table*}[t]\n",
" \\small\n",
" \\centering\n",
" \\begin{tabular}{lcccccccc}\n",
" \\toprule\n",
"\n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \\textbf{Model} & \\textit{\\underline{De-En}} & \\textit{\\underline{Ru-En}} & \\textit{\\underline{Zh-En}} & \\underline{Fi-En} & Gu-En & Kk-En & Lt-En & \\textbf{Avg.} \\\\\n",
" \n",
" \\midrule\n",
" \n",
" \\multicolumn{9}{c}{\\textit{Reference-only Evaluation}} \\\\\n",
" \\cdashline{1-9}\\noalign{\\vskip 0.1ex}\n",
" \\cmodel-MUP & 25.5 & 21.3 & 46.6 & 37.0 & \\textbf{32.2} & \\textbf{39.1} & 38.6 & \\colorbox{blue!10}{34.3} \\\\\n",
" \\cmodel-UP & \\textbf{25.6} & \\textbf{21.9} & \\textbf{46.7} & \\textbf{37.9} & \\textbf{32.2} & 38.7 & \\textbf{40.0} & \\colorbox{blue!10}{\\textbf{34.8}} \\\\\n",
" \n",
" \\midrule\n",
" \n",
" \\multicolumn{9}{c}{\\textit{Source-only Evaluation}} \\\\\n",
" \\cdashline{1-9}\\noalign{\\vskip 0.1ex}\n",
" \\cmodel-MUP & 18.0 & 16.3 & 41.0 & 33.9 & 29.6 & \\textbf{34.7} & \\textbf{35.7} & \\colorbox{red!10}{29.9} \\\\\n",
" \\cmodel-UP & \\textbf{19.3} & \\textbf{16.9} & \\textbf{41.4} & \\textbf{34.0} & \\textbf{29.7} & 33.6 & 35.4 & \\colorbox{red!10}{\\textbf{30.0}} \\\\\n",
" \n",
" \\midrule\n",
" \n",
" \\multicolumn{9}{c}{\\textit{Source-Reference-Combined Evaluation}} \\\\\n",
" \\cdashline{1-9}\\noalign{\\vskip 0.1ex}\n",
" \\cmodel-MUP & 25.2 & 20.9 & 46.9 & 37.0 & 32.0 & 38.5 & 38.8 & \\colorbox{green!10}{34.2} \\\\\n",
" \\cmodel-UP & \\textbf{26.0} & \\textbf{22.0} & \\textbf{47.2} & \\textbf{37.7} & \\textbf{32.3} & \\textbf{39.4} & \\textbf{40.0} & \\colorbox{green!10}{\\textbf{35.0}} \\\\\n",
" \\bottomrule\n",
" \\end{tabular}\n",
" \\caption{Kendall's Tau correlation (\\\n",
" \\label{table.appendix_model-m_english-targeted}\n",
"\\end{table*}\n",
"\\paragraph{Multilingual-Targeted}\n",
"As seen in Table~\\ref{table.main_results_metric_en-x}, the multilingual-targeted \\cmodel-MUP gives dominant performance than all strong baselines on \\cref, \\csrc~and \\csrcref, demonstrating the transferability and effectiveness of our approach.\n",
"Besides, the \\cmodel-UP also gives dominant results, revealing an improvement of 0.6, 0.3 and 0.9 averaged Kendall's $\\tau$ correlation scores, respectively.\n",
"\n",
"However, we find that \\cmodel-MUP outperforms strong baselines but slightly worse than \\cmodel-UP on English-targeted translation directions (see Table~\\ref{table.appendix_model-m_english-targeted}).\n",
"We think the reason lies in the curse of multilingualism and vocabulary dilution~\\cite{conneau2020unsupervised}.\n",
"\n",
"\n",
"\n",
"\\begin{table*}[t]\n",
" \\small\n",
" \\centering\n",
" \\begin{tabular}{lcccccccc}\n",
" \\toprule\n",
" \n",
" \n",
" \n",
" \\textbf{Model} & \\textit{{En-De}} & \\textit{En-Zh} & \\textit{{Ru-En}} & Et-En & Ne-En & Ro-En & Si-En & \\textbf{Avg.} \\\\\n",
" \n",
" \\midrule\n",
" \\noalign{\\vskip -0.5ex}\n",
" \n",
" \n",
" \n",
" OpenKiwi~\\cite{kepler2019openkiwi} & 14.6 & 19.0 & 54.8 & 47.7 & 38.6 & 68.5 & 37.4 & \\colorbox{red!10}{40.1} \\\\\n",
" mBERT~\\cite{devlin2019bert} & 37.7 & 39.8 & 66.6 & 62.3 & 64.5 & 83.5 & - & \\colorbox{red!10}{-} \\\\\n",
" \\noalign{\\vskip -0.5ex}\n",
" TransQuest-m~\\cite{ranasinghe2020transquest} & 44.2 & 46.5 & \\textbf{75.2} & 75.7 & \\textbf{75.8} & \\textbf{88.6} & \\textbf{65.3} & \\colorbox{red!10}{67.3} \\\\\n",
" \\cdashline{1-9}\\noalign{\\vskip 0.1ex}\n",
" \\cmodel-MUP & \\textbf{52.5} & \\textbf{50.5} & 64.4 & \\textbf{79.1} & 75.6 & 88.3 & 64.3 & \\colorbox{red!10}{\\textbf{67.8}} \\\\\n",
" \n",
" \\bottomrule\n",
" \\end{tabular}\n",
" \n",
" \\caption{Pearson correlation (\\\n",
" \\label{table.main_results_qe}\n",
"\\end{table*}\n",
"\\paragraph{Quality Estimation}\n",
"\n",
"The results for \\cmodel~approach on WMT 2020 QE task are concluded in Table~\\ref{table.main_results_qe}.\n",
"As seen, it achieves competitive results on QE task compared with the winner submission~\\cite{ranasinghe2020transquest}. \n",
"\n",
"\n",
"\n",
"\\section{Ablation Studies}\n",
"In this section, we conduct ablation studies to investigate the effectiveness of regional attention patterns (\\S\\ref{sec.ablation.mra}), unified training (\\S\\ref{sec:unift}), and ranking-based data labeling (\\S\\ref{sec:rdl}).\n",
"All experiments are conducted by following English-targeted setting.\n",
"\n",
"\\subsection{Regional Attention Patterns}\n",
"\\label{sec.ablation.mra}\n",
"\n",
"\\begin{table}[t]\n",
" \\centering\n",
" \\small\n",
" \\scalebox{1.0}\n",
" {\n",
" \\begin{tabular}{lcr}\n",
" \\toprule\n",
" \\textbf{Model} & \\textbf{Avg. $\\tau$} (\\\n",
" \\midrule\n",
" Full attention & 34.1 & -- \\\\\n",
" \\cdashline{1-3}\\noalign{\\vskip 0.3ex}\n",
" \n",
" \n",
" ~~no H$\\rightarrow$S (Soft) & \\textbf{34.8} & \\textbf{+0.7} \\\\\n",
" ~~no S$\\rightarrow$H (Soft) & 34.6 & +0.5 \\\\\n",
" \\cdashline{1-3}\\noalign{\\vskip 0.3ex}\n",
" ~~no H$\\rightarrow$S, H$\\rightarrow$R \\& S$\\rightarrow$R (Hard) & 34.3 & +0.2 \\\\\n",
" \\cdashline{1-3}\\noalign{\\vskip 0.3ex}\n",
" ~~no R$\\rightarrow$S & 34.0 & -0.1 \\\\\n",
" ~~no S$\\rightarrow$R & 33.9 & -0.2 \\\\\n",
" ~~no R$\\rightarrow$H & 33.6 & -0.5 \\\\\n",
" ~~no H$\\rightarrow$R & 34.0 & -0.1 \\\\\n",
" \\bottomrule\n",
" \\end{tabular}\n",
" }\n",
" \\caption{Averaged Kendall's Tau correlation (\\\n",
" \\label{table.ablation_monotonic}\n",
"\\end{table}\n",
"To investigate the effectiveness of MRA, we further collect experiments in Table~\\ref{table.ablation_monotonic}.\n",
"As seen, MRA can give performance improvements than full attention, and preventing the interactions between hypothesis and source segment can improve the performance most.\n",
"We think the reasons behind are twofold.\n",
"First, the source side is formed with a different language, whose semantic information is rather weak than the reference side.\n",
"Second, by preventing direct interactions between source and hypothesis, semantics inside the former must be passed through reference, which is helpful for disambiguation.\n",
"Besides, not allowing the source to derive information from the hypothesis is better than the opposite direction.\n",
"\\newcite{wang2020what} found that the positional embeddings in PLM are engaged with strong adjacent information.\n",
"We think the reason why S$\\rightarrow$H performs worse than H$\\rightarrow$S lies in the skipping of indexes, which corrupts positional similarities in alignment calculation.\n",
"\n",
"Additionally, when we combined two methods together, \\textit{i.e.}, unified pretraining and finetuning with \\csrcref~\\cmodel-MRA setting, model performance drops to 34.9 over English-targeted tasks on average.\n",
"We think that both methods all intend to solve the problem of unseen \\csrcref~input format, and MRA may not be necessary if massive data examples can be obtained for pretraining.\n",
"Nevertheless, \\cmodel-MRA has its advantage on wide application without requiring pseduo labeled data.\n",
"\n",
"\\subsection{Unified Training}\n",
"\\label{sec:unift}\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
" \n",
"\n",
"\n",
"\n",
" \n",
"\n",
"\n",
"\n",
" \n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\\begin{table}[t]\n",
" \\centering\n",
" \\small\n",
" \\scalebox{1.0}{\n",
" \\begin{tabular}{cccccr}\n",
" \\toprule\n",
" \\textbf{Unified} & \\textbf{Unified} & \\multicolumn{3}{c}{\\textbf{Avg. $\\tau$} (\\\n",
" \\textbf{Pretrain} & \\textbf{Finetune} & \\ccref & \\ccsrc & \\ccsrcref \\\\\n",
" \\midrule\n",
" \n",
" \\ding{51} & \\ding{51} & \\textbf{34.8} & \\textbf{30.0} & \\textbf{35.0} \\\\\n",
" \n",
" \n",
" \n",
" \\cdashline{1-5}\\noalign{\\vskip 0.3ex}\n",
" \\ding{51} & \\ding{55} & 33.8 & 29.1 & 33.9 \\\\\n",
" \n",
" \n",
" \n",
" \\ding{55} & \\ding{55} & 31.9 & 27.7 & 32.6 \\\\\n",
" \n",
" \\bottomrule\n",
" \\end{tabular}\n",
" }\n",
" \\caption{Unified and task-specific training for \\cmodel-UP approach. As seen, combination of unified pretraining and finetuning gives best performances, meanwhile requires only one unified model.}\n",
" \\label{table.ablation_unified_specific_training}\n",
"\\end{table}\n",
"\\begin{table}[t]\n",
" \\centering\n",
" \\small\n",
" \\scalebox{1.0}\n",
" {\n",
" \\begin{tabular}{lcr}\n",
" \\toprule\n",
" \\textbf{Method} & \\textbf{Avg.} $\\tau$ (\\\n",
" \\midrule\n",
" Rank-Norm, Ens & \\textbf{35.0} & - \\\\\n",
" Rank-Norm & 34.7 & -0.3 \\\\\n",
" \\cdashline{1-3}\\noalign{\\vskip 0.3ex}\n",
" Z-Norm, Ens & 33.5 & -1.5 \\\\\n",
" Z-Norm & 34.2 & -0.8 \\\\\n",
" \n",
" \\bottomrule\n",
" \\end{tabular}\n",
" }\n",
" \\caption{Pseudo-data labeling with different methods. Ranking-based normalization (Rank-Norm) performs better than conventional Z-score approach (Z-Norm). Besides, ensembling (Ens) ranking-based normalized scores can give higher result, while conventional Z-Norm performs worse after ensembling.}\n",
" \\label{table.ablation_pseudo_labeling}\n",
"\\end{table}\n",
"\n",
"Experiments for comparing unified and task-specific training are concluded in Table~\\ref{table.ablation_unified_specific_training}.\n",
"As seen, when using the unified pretraining checkpoint to finetune over the specific task, performance over three models reveals performance drop consistently, indicating that the unified finetuning is helpful for model learning.\n",
"This also verifies our hypothesis, that the cores of \\cref, \\csrc, and \\csrcref~tasks are identical to each other.\n",
"Moreover, unified pretraining and finetuning are complementary to each other.\n",
"Also, utilizing task-specific pretraining instead of unified one reveals worse performance.\n",
"To sum up, unifying both pretraining and finetuning only reveals one model, showing its advantage on the generalization on all tasks, where one united model can cover all functionalities of \\cref, \\csrc~and \\csrcref~tasks concurrently.\n",
"\n",
"\n",
"\\subsection{Ranking-based Data Labeling}\n",
"\\label{sec:rdl}\n",
"To verify the effectiveness of ranking-based labeling, we collect the results of models applying different pseudo labeling strategies.\n",
"After deriving the original scores from the well-trained \\cmodel-MRA checkpoint, we use Z-score and proposed ranking-based normalization methods to label synthetic data.\n",
"For both methods, we also apply an ensembling strategy to assign training examples with averaged scores deriving from 3 \\cmodel-MRA checkpoints.\n",
"Results show that, Z-score normalization reveals a performance drop when applying score ensembling with multiple models.\n",
"Our proposed ranking-based normalization can boost the \\cmodel-UP model training, and its ensembling approach can further improve the performance.\n",
"\n",
"\n",
"\\section{Conclusion}\n",
"In the past decades, automatic translation evaluation is mainly divided into \\cref, \\csrc~and \\csrcref~tasks, each of which develops independently and is tackled by various task-specific methods.\n",
"We suggest that the three tasks are possibly handled by a unified framework, thus being ease of use and facilitating the knowledge transferring.\n",
"Contributions of our work are mainly in three folds:\n",
"(a) We propose a flexible and unified translation evaluation model UniTE, which can be adopted into the three tasks at once;\n",
"(b) Through in-depth analyses, we point out that the main challenge of unifying three tasks stems from the discrepancy between vanilla pretraining and multi-tasks finetuning, and fill this gap via monotonic regional attention (MRA) and unified pretraining (UP);\n",
"(c) Our single model consistently outperforms a variety of state-of-the-art or winner systems across high-resource and zero-shot evaluation in WMT 2019 Metrics and WMT 2020 QE benchmarks, showing its advantage of flexibility and convincingness.\n",
"We hope our new insights can contribute to subsequent studies in the translation evaluation community.\n",
"\n",
"\\section*{Acknowledgements}\n",
"The authors would like to send great thanks to all reviewers and meta-reviewer for their insightful comments.\n",
"This work was supported in part by the Science and Technology Development Fund, Macau SAR (Grant No. 0101/2019/A2), the Multi-year Research Grant from the University of Macau (Grant No. MYRG2020-00054-FST), National Key Research and Development Program of China (No. 2018YFB1403202), and Alibaba Group through Alibaba Research Intern Program.\n",
"\n",
"\n",
"\n",
"\\bibliographystyle{acl_natbib}\n",
"\\bibliography{anthology}\n",
"\n",
"\n",
"\\appendix\n",
"\n",
"\n",
"\\section{Collection of Pretraining Data} \n",
"Considering the English-targeted model, we select Czech (Cz), German (De), Japanese (Ja), Russian (Ru), and Chinese (Zh) as source languages, and English (En) as target.\n",
"For each translation direction, we collect 1 million samples, finally yielding 5 million examples in total for unified pretraining. \n",
"As to the multilingual-targeted model, we further collect 1 million synthetic data for each language direction of En-Cz, En-De, En-Ja, En-Ru, and En-Zh. Finally, we construct 10 million examples for the pretraining of the multilingual version by adding the data of the English-targeted model.\n",
"Note that, for a fair comparison, we filter out all pretraining examples that are involved in benchmarks.\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\\begin{table*}[t]\n",
" \\small\n",
" \\centering\n",
" \\begin{tabular}{lcccccccc}\n",
" \\toprule\n",
"\n",
" \n",
" \n",
" \n",
" \n",
" \n",
" \\textbf{Model} & \\textit{\\underline{De-En}} & \\textit{\\underline{Ru-En}} & \\textit{\\underline{Zh-En}} & \\underline{Fi-En} & Gu-En & Kk-En & Lt-En & \\textbf{Avg.} \\\\\n",
" \n",
" \\midrule\n",
" \n",
" \\multicolumn{9}{c}{\\textit{Reference-only Evaluation}} \\\\\n",
" \\cdashline{1-9}\\noalign{\\vskip 0.1ex}\n",
" \\cmodel-MUP & 25.5 & 21.3 & 46.6 & 37.0 & \\textbf{32.2} & \\textbf{39.1} & 38.6 & \\colorbox{blue!10}{34.3} \\\\\n",
" \\cmodel-UP & \\textbf{25.6} & \\textbf{21.9} & \\textbf{46.7} & \\textbf{37.9} & \\textbf{32.2} & 38.7 & \\textbf{40.0} & \\colorbox{blue!10}{\\textbf{34.8}} \\\\\n",
" \n",
" \\midrule\n",
" \n",
" \\multicolumn{9}{c}{\\textit{Source-only Evaluation}} \\\\\n",
" \\cdashline{1-9}\\noalign{\\vskip 0.1ex}\n",
" \\cmodel-MUP & 18.0 & 16.3 & 41.0 & 33.9 & 29.6 & \\textbf{34.7} & \\textbf{35.7} & \\colorbox{red!10}{29.9} \\\\\n",
" \\cmodel-UP & \\textbf{19.3} & \\textbf{16.9} & \\textbf{41.4} & \\textbf{34.0} & \\textbf{29.7} & 33.6 & 35.4 & \\colorbox{red!10}{\\textbf{30.0}} \\\\\n",
" \n",
" \\midrule\n",
" \n",
" \\multicolumn{9}{c}{\\textit{Source-Reference-Combined Evaluation}} \\\\\n",
" \\cdashline{1-9}\\noalign{\\vskip 0.1ex}\n",
" \\cmodel-MUP & 25.2 & 20.9 & 46.9 & 37.0 & 32.0 & 38.5 & 38.8 & \\colorbox{green!10}{34.2} \\\\\n",
" \\cmodel-UP & \\textbf{26.0} & \\textbf{22.0} & \\textbf{47.2} & \\textbf{37.7} & \\textbf{32.3} & \\textbf{39.4} & \\textbf{40.0} & \\colorbox{green!10}{\\textbf{35.0}} \\\\\n",
" \\bottomrule\n",
" \\end{tabular}\n",
" \\caption{Kendall's Tau correlation (\\\n",
" \\label{table.appendix_model-m_english-targeted}\n",
"\\end{table*}\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\\begin{table*}[t]\n",
" \\small\n",
" \\centering\n",
" \\begin{tabular}{lcccccccc}\n",
" \\toprule\n",
" \n",
" \n",
" \n",
" \\textbf{Model} & \\textit{{En-De}} & \\textit{En-Zh} & \\textit{{Ru-En}} & Et-En & Ne-En & Ro-En & Si-En & \\textbf{Avg.} \\\\\n",
" \n",
" \\midrule\n",
" \\noalign{\\vskip -0.5ex}\n",
" \n",
" \n",
" \n",
" OpenKiwi~\\cite{kepler2019openkiwi} & 14.6 & 19.0 & 54.8 & 47.7 & 38.6 & 68.5 & 37.4 & \\colorbox{red!10}{40.1} \\\\\n",
" mBERT~\\cite{devlin2019bert} & 37.7 & 39.8 & 66.6 & 62.3 & 64.5 & 83.5 & - & \\colorbox{red!10}{-} \\\\\n",
" \\noalign{\\vskip -0.5ex}\n",
" TransQuest-m~\\cite{ranasinghe2020transquest} & 44.2 & 46.5 & \\textbf{75.2} & 75.7 & \\textbf{75.8} & \\textbf{88.6} & \\textbf{65.3} & \\colorbox{red!10}{67.3} \\\\\n",
" \\cdashline{1-9}\\noalign{\\vskip 0.1ex}\n",
" \\cmodel-MUP & \\textbf{52.5} & \\textbf{50.5} & 64.4 & \\textbf{79.1} & 75.6 & 88.3 & 64.3 & \\colorbox{red!10}{\\textbf{67.8}} \\\\\n",
" \n",
" \\bottomrule\n",
" \\end{tabular}\n",
" \n",
" \\caption{Pearson correlation (\\\n",
" \\label{table.main_results_qe}\n",
"\\end{table*}\n",
"\n",
"\n",
"\n",
"\n",
"\\section{Reproducibility}\n",
"All the models reported in this paper were finetuned on a single Nvidia V100 (32GB) GPU.\n",
"Specifically for \\cmodel-UP and \\cmodel-MUP, the pretraining is arranged on 4 Nvidia V100 (32GB) GPUs. \n",
"Our framework is built upon COMET repository~\\cite{rei2020comet}.\n",
"\n",
"For the contribution to the research community, we release both the source code of \\cmodel~framework and the well-trained evaluation models as described in this paper at \\href{https://github.com/NLP2CT/UniTE}{https://github.com/NLP2CT/UniTE}.\n",
"\n",
"\\end{document}\n"
],
"del_percentage": 0.04196
}
} |