基于多粒度认知的命名实体识别方法  

Named entity recognition method based on multi-granularity cognition

作　　者：李攀锋 陈樱珏 钟泠韵 林锋[1] LI Pan-Feng;CHEN Ying-Jue;ZHONG Ling-Yun;LIN Feng(College of Computer Science,Sichuan University,Chengdu 610065,China)

机构地区：[1]四川大学计算机学院,成都610065

出　　处：《四川大学学报（自然科学版）》2022年第2期58-64,共7页Journal of Sichuan University(Natural Science Edition)

基　　金：国家重点研发计划(2018YFC0832300,2018YFC0832303)。

摘　　要：在数据匮乏的领域,命名实体识别效果受限于欠拟合的字词特征表达,引入常规的多任务学习方法可以有所改善,但需要额外的标注成本.针对这一问题,提出了一种基于多粒度认知的命名实体识别方法,在不产生额外标注成本的前提下,增强字特征信息,提高命名实体识别效果.该方法从多粒度认知理论出发,以BiLSTM和CRF为基础模型,将字粒度下的命名实体识别任务与句子全局粒度下的实体数量预测任务相联合,共同优化字嵌入表达.三个不同类型的数据集上的多组实验表明,引入多粒度认知的方法有效地提升了命名实体识别效果.In the field of data scarcity,the performance of named entity recognition is limited by the expression of under-fitting word features.The named entity recognition effect can be improved by introducing conventional multi-task learning methods,but additional labeling costs are required.Aiming at addressing this problem,we propose a new named entity recognition method based on multi-granularity cognition,which can enhance the character feature information and improve the performance of named entity recognition without incurring additional tagging costs.In order to optimize the expression of word embedding,in this approach,we start from the multi-granularity cognition theory and use BiLSTM and CRF as the basic model,the task of named entity recognition under word granularity is combined with the task of entity number prediction under sentence global granularity.Multiple experiments on three different types of data sets show that the method of introducing multi-granularity cognition method can effectively improve the performance of named entity recognition.

关 键 词：命名实体识别 多粒度认知 多任务学习 自然语言处理 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]