结合五笔字形与上下文相关字向量的命名实体识别  

Named Entity Recognition Combining Wubi Glyphs with Contextualized Character Embeddings

作　　者：张栋[1] 王铭涛 陈文亮[1] ZHANG Dong;WANG Mingtao;CHEN Wenliang(School of Computer Science and Technology,Soochow University,Suzhou,Jiangsu 215006,China)

机构地区：[1]苏州大学计算机科学与技术学院,江苏苏州215006

出　　处：《计算机工程》2021年第3期94-101,共8页Computer Engineering

基　　金：国家自然科学基金(61876115)。

摘　　要：命名实体识别(NER)作为自然语言处理的重要部分,在信息抽取和知识图谱等任务中得到广泛应用。然而目前中文预训练语言模型通常仅对上下文中的字符进行建模,忽略了中文字符的字形结构。提出2种结合五笔字形的上下文相关字向量表示方法,以增强字向量的语义表达能力。第一种方法分别对字符和字形抽取特征并联合建模得到字向量表示,第二种方法将五笔字形作为辅助信息拼接到字向量中,训练一个基于字符和五笔字形的混合语言模型。实验结果表明,所提两种方法可以有效提升中文NER系统的性能,且结合五笔字形的上下文相关字向量表示方法的系统性能优于基于单一字符的语言模型。As a basic task of natural language processing,Named Entity Recognition(NER)is widely used in information extraction,knowledge graph and other tasks.However,the existing Chinese pre-trained language models usually only capture characters in the context,ignoring the structure of Chinese characters-glyphs.This paper proposes two kinds of contextualized character embeddings representation methods combined with Wubi to enhance the semantic representation of character embeddings.The first method is to present the character embeddings by jointly modeling the extracted character and glyph features.The second one is to splice the Wubi glyphs into the character embeddings for assistance,and on this basis train a language model combining characters and Wubi glyphs.Experimental results show that the proposed methods can significantly improve the performance of Chinese NER systems,which outperform the language models based on only characters.

关 键 词：语言模型 命名实体识别 五笔字形 上下文相关字向量 无标注语料 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]