vocab/gpt_neox_japanese/README.md

vocab.txt

るのは
よね
写真,寫真,冩真,写眞,寫眞,冩眞
マイ
そん
女性,𠨰性,⼥性,女𧢱,𠨰𧢱,⼥𧢱
内容,內容,内㣑,内㝐,内彮,内𠕺,內㣑,內㝐,內彮,內𠕺

怎么还有不同写法？？

文本归一化

以下的normalization，在生成任务中并不好。

        self.content_repatter1 = re.compile(r"(https?|ftp)(:\/\/[-_\.!~*\'()a-zA-Z0-9;\/?:\@&=\+$,%#]+)")
        self.content_repatter2 = re.compile(r"[A-Za-z0-9\._+]*@[\-_0-9A-Za-z]+(\.[A-Za-z]+)*")
        self.content_repatter3 = re.compile(r"[\(]{0,1}[0-9]{2,4}[\)\-\(]{0,1}[0-9]{2,4}[\)\-]{0,1}[0-9]{3,4}")
        self.content_repatter4 = re.compile(
            r"([12]\d{3}[/\-年])*(0?[1-9]|1[0-2])[/\-月]((0?[1-9]|[12][0-9]|3[01])日?)*(\d{1,2}|:|\d{1,2}時|\d{1,2}分|\(日\)|\(月\)|\(火\)|\(水\)|\(木\)|\(金\)|\(土\)|㈰|㈪|㈫|㈬|㈭|㈮|㈯)*"
        )
        self.content_repatter5 = re.compile(
            r"(明治|大正|昭和|平成|令和|㍾|㍽|㍼|㍻|\u32ff)\d{1,2}年(0?[1-9]|1[0-2])月(0?[1-9]|[12][0-9]|3[01])日(\d{1,2}|:|\d{1,2}時|\d{1,2}分|\(日\)|\(月\)|\(火\)|\(水\)|\(木\)|\(金\)|\(土\)|㈰|㈪|㈫|㈬|㈭|㈮|㈯)*"
        )
        self.content_repatter6 = re.compile(
            r"((0|[1-9]\d*|[1-9]\d{0,2}(,\d{3})+)*億)*((0|[1-9]\d*|[1-9]\d{0,2}(,\d{3})+)*万)*((0|[1-9]\d*|[1-9]\d{0,2}(,\d{3})+)*千)*(0|[1-9]\d*|[1-9]\d{0,2}(,\d{3})+)*(千円|万円|千万円|円|千ドル|万ドル|千万ドル|ドル|千ユーロ|万ユーロ|千万ユーロ|ユーロ)+(\(税込\)|\(税抜\)|\+tax)*"
        )

    def clean_text(self, content):
        content = self.content_repatter1.sub("<URL>", content)
        content = self.content_repatter2.sub("<EMAIL>", content)
        content = self.content_repatter3.sub("<TEL>", content)
        content = self.content_repatter4.sub("<DATE>", content)
        content = self.content_repatter5.sub("<DATE>", content)
        content = self.content_repatter6.sub("<PRICE>", content)
        content = content.translate(self.content_trans1)
        while "<BLOCK><BLOCK>" in content:
            content = content.replace("<BLOCK><BLOCK>", "<BLOCK>")
        return content

    def tokenize(self, text, clean=False):
        text = text.replace(" ", "<SP>")
        text = text.replace("　", "<SP>")
        text = text.replace("\r\n", "<BR>")
        text = text.replace("\n", "<BR>")
        text = text.replace("\r", "<BR>")
        text = text.replace("\t", "<TAB>")
        text = text.replace("—", "ー")
        text = text.replace("−", "ー")
        for k, v in self.emoji["emoji"].items():
            if k in text:
                text = text.replace(k, v)
        if clean:
            text = self.clean_text(text)