引入词集级注意力机制的中文命名实体识别方法  

Incorporating word-set attention into Chinese named entity recognition Method

作　　者：钟诗胜[1] 陈曦 赵明航 张永健 ZHONG Shi-sheng;CHEN Xi;ZHAO Ming-hang;ZHANG Yong-jian(School of Mechatronics Engineering,Harbin Institute of Technology,Harbin 150001,China;School of Ocean Engineering,Harbin Institute of Technology at Weihai,Weihai 26A2091 China)

机构地区：[1]哈尔滨工业大学机电工程学院,哈尔滨150001 [2]哈尔滨工业大学(威海)海洋工程学院,山东威海264209

出　　处：《吉林大学学报（工学版）》2022年第5期1098-1105,共8页Journal of Jilin University:Engineering and Technology Edition

基　　金：国家重点研发计划项目(2019YFB1705302);国家自然科学基金项目(52075120)。

摘　　要：目前许多研究通过引入外部词汇信息提高中文命名实体识别的效果,这些研究通常在词集内对词向量进行加权将词信息整合到字表示中,但忽略了各个词集的重要度不同。为此,本文提出了一种引入词集级注意力机制的中文命名实体识别方法,在引入外部词汇信息的基础上,根据字在词中的位置划分词集、获得词集向量,进而建立词集级别的注意力机制,强制性地将注意力集中在适当的词集,忽略不可靠的部分,从而提高中文命名实体识别效果。最后,通过多个数据集上的实验验证了该方法的有效性。Chinese word segmentation is necessary to provide word-level information for Chinese named entity recognition.Recently,many researchers have attempted to improve the performance of Chinese named entity recognition by incorporating word-level information into character representations.However,they ignored the different importance among word sets.Aiming at this issue,this paper focuses on wordset level relationship and proposes a new method,which incorporates word-set attention into Chinese named entity recognition,that adaptively recalibrates word-set level features by attention mechanism.This method makes the network perform better in Chinese named entity recognition by learning to selectively emphasize informative features and suppress useless ones.Finally,the effectiveness of the method is verified by experiments on multiple datasets.

关 键 词：人工智能 中文命名实体识别 词集级 注意力机制 外部词汇信息 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]