融合特征编码和短语交互感知的隐式篇章关系识别
Implicit Discourse Relation Recognition Integrating Feature Coding and Phrase Interaction Perception
王秀利 1金方焱2
作者信息
- 1. 中央财经大学信息学院,北京 102206;国家金融安全教育部工程研究中心,北京 102206
- 2. 中央财经大学信息学院,北京 102206
- 折叠
摘要
隐式篇章关系识别难度大、普遍性高.从论元编码和论元交互角度入手,提出了一种融合特征编码和短语交互感知的隐式篇章关系识别模型.该模型兼顾了论元本身特征和论元间交互特征的作用,并分别进行了优化.论元编码部分整合了双向长短时记忆网络和循环注意力卷积神经网络,能够更全面地捕获论元全局和局部特征;论元交互部分从短语层级考虑论元间的语义关系建模,构建了短语级交互注意力机制,并利用神经张量网络深入挖掘其中的关系模式,更能体现出论元间潜在的更深层次的关联关系.在宾州篇章树库数据集上的实验结果表明,该模型F1值均优于其他模型.
Abstract
Implicit discourse relation recognition is a challenging task because of its difficulty and universality.From the perspective of argument coding and argument interaction,an implicit discourse relation recognition model integrating feature coding and phrase interaction perception is proposed.The model considers both the characteristics of argument it-self and the interaction characteristics between arguments,and optimizes separately.The part of argument coding incorpo-rates bidirectional long short-term memory(BiLSTM)and recurrent attention convolution neural network(RACNN),which can capture global and local features of arguments in a more comprehensive way;in the part of argument interaction,the se-mantic relationship between arguments is modeled from phrase level,and a mechanism of phrase-level interactive attention is constructed.Also,neural tensor network(NTN)is used to dig into the relational pattern,which can better reflect the po-tential deeper relational relationship between arguments.Experimental results on penn discourse treebank(PDTB)dataset show that the F1 values of this model are superior to other comparison models.
关键词
隐式篇章关系识别/双向长短时记忆网络/循环注意力卷积神经网络/短语级交互注意力/神经张量网络Key words
implicit discourse relation recognition/bidirectional long short-term memory/recurrent attention convo-lution neural network/phrase-level interactive attention/neural tensor network引用本文复制引用
基金项目
教育部哲学社会科学研究重大课题攻关项目(22JZD011)
中央财经大学新兴交叉学科建设项目()
出版年
2024
国家科技期刊平台
NSTL
万方数据