基于MUBTM的方面词情感三元组抽取方法研究  

Research on aspect sentiment triplet extraction based on MUBTM

作　　者：葛继科 程文俊 武承志 陈祖琴[2] 董焱 Ge Jike;Cheng Wenjun;Wu Chengzhi;Chen Zuqin;Dong Yan(School of Intelligent Technology&Engineering,Chongqing University of Science&Technology,Chongqing 401331,China;Library,Chongqing University of Science&Technology,Chongqing 401331,China)

机构地区：[1]重庆科技学院智能技术与工程学院,重庆401331 [2]重庆科技学院图书馆,重庆401331

出　　处：《计算机应用研究》2023年第5期1416-1421,共6页Application Research of Computers

基　　金：国家社会科学基金西部项目(19XTQ010)。

摘　　要：方面词情感三元组抽取旨在从评论文本中识别方面词、评论观点词及情感极性。针对现有研究容易忽略三元组中部分实体与其情感极性在标记空间中的关联关系,并且较易出现三元组边界识别错误的问题,提出一种基于表格填充的多类别统一及边界标记模型(MUBTM)。首先,采用双仿射注意力机制学习词对间的交互,并构建多类别统一标记空间;然后,利用标记空间的对称性结构化约束与级联二进制边界约束,限制词对间的概率分布;最后,根据子元素在统一标记空间中结构为正方形或矩形的特性,逐步解码生成方面词情感三元组。实验结果表明,与其他基线模型相比,本模型在方面词情感三元组抽取上的F1值有显著提升,表明了所提出方法的可行性。Aspect sentiment triplet extraction aims to identify aspect terms,comment opinion terms and sentiment polarities from comment texts.In view of the existing studies,it is easy to ignore the association relationship between some entities and their sentiment polarities in the triplet,and often occurs the errors in recognizing the boundary of the triplets.This paper proposed a multi-category unification and boundary tagging model based on table filling(MUBTM)by using the category of the sub-elements.Firstly,the model used biaffine attention mechanism to learn the interaction among word pairs,and constructed a multi-category unification tagging space.Next,it adopted the symmetry structural constraints and the cascading binary boundary constraints of the tag space to restrict the probability distribution among word pairs.Finally,according to the property that the sub-elements were square or rectangular in the unified tag spaces,gradually decoded and generated aspect sentiment triplets.Experimental results show that,compared with the baseline models,the proposed model significantly raises the F 1 value of aspect sentiment triplet extraction,indicating the feasibility of the proposed method.

关 键 词：情感三元组抽取 表格填充 级联二进制 双仿射注意力机制 标记空间 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]