摘要
针对现有命名实体识别方法存在的语义信息获取不全面问题,提出基于Affix-Attention的命名实体识别语义补充方法。将句子和句子中每个单词对应的词缀输入到编码层,使用Bi-LSTM提取上下文特征。在编码层设计特征融合模块、建模文本特征与词缀特征的对应关系,使用Affix-Attention同时关注文本信息和词缀信息进行语义补充。解码层使用CRF层得到目标序列。在生物医学领域的JNLPBA-2004和BC2GM基准数据集上的试验结果综合评价指标F1达到81.73%、84.73%;在公共数据集CONLL-2003中试验结果综合评价指标F1达到91.35%。试验结果表明,本研究方法能够有效获取词的内部语义特征,融合文本信息和词缀信息,达到语义补充的效果,提升命名实体识别的性能。
Aiming at the problem of incomplete semantic information acquisition in existing named entity recognition methods,a semantic supplementary method for named entity recognition based on AffixAttention was proposed.The sentence and the affixes corresponding to each word in the sentence were input to the encoding layer,and BiLSTM was used to extract contextual features.A feature fusion module was designed at the coding layer to model the correspondence between text features and affix features,and used AffixAttention to pay attention to both text information and affix information for semantic supplementation.The decoding layer used the CRF layer to obtain the target sequence.The comprehensive evaluation index F1 of the test results on the JNLPBA2004 and BC2GM benchmark datasets in the biomedical field reached 81.73%and 84.73%.In the public dataset CONLL2003,the comprehensive evaluation index F1 of the test results reached 91.35%.The experimental results showed that this research method could effectively obtain the internal semantic features of words,integrate text information and affix information,achieve the effect of semantic supplementation,and improve the performance of named entity recognition.
作者
宋佳芮
陈艳平
王凯
黄瑞章
秦永彬
SONG Jiarui;CHEN Yanping;WANG Kai;HUANG Ruizhang;QIN Yongbin(State Key Laboratory of Public Big Data,Guiyang 550025,Guizhou,China;College of Computer Science and Technology,Guizhou University,Guiyang 550025,Guizhou,China)
出处
《山东大学学报(工学版)》
CAS
CSCD
北大核心
2023年第2期70-76,共7页
Journal of Shandong University(Engineering Science)
基金
国家自然科学基金资助项目(62166007)。
关键词
命名实体识别
语义补充
注意力机制
特征融合
深度学习
named entity recognition
semantic supplement
attention mechanism
feature fusion
deep learning