计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (15): 187-195.DOI: 10.3778/j.issn.1002-8331.2204-0487

• 模式识别与人工智能 • 上一篇    下一篇

面向方面级情感分析的交互关系图注意力网络

郑智雄,刘建华,孙水华,林鸿辉,徐戈   

  1. 1.福建省大数据挖掘与应用技术重点实验室,福州 350118
    2.福建工程学院 计算机科学与数学学院,福州 350118
    3.闽江学院 计算机与控制工程学院,福州 350108
  • 出版日期:2023-08-01 发布日期:2023-08-01

Interactive Relation Graph Attention Network Model for Aspect-Based Sentiment Analysis

ZHENG Zhixiong, LIU Jianhua, SUN Shuihua, LIN Honghui, XU Ge   

  1. 1.Fujian Provincial Key Laboratory of Big Data Mining and Applications, Fuzhou 350118, China
    2.College of Information Science and Engineering, Fujian University of Technology, Fuzhou 350118, China
    3.College of Computer and Control Engineering, Minjiang University, Fuzhou 350108, China
  • Online:2023-08-01 Published:2023-08-01

摘要: 方面级别情感分析旨在分析网络评论每个方面的情感极性,是一种细粒度的情感分析技术。已经有许多相关研究把语法依赖树与图注意力网络结合应用到该任务,取得了较好的成绩。针对以往研究忽略关系类型信息,没有充分挖掘关系类型所包含的潜在语义信息,以及忽略了依赖关系和关系类型之间的联系等问题,提出了一种基于图注意力网络的交互关系图注意力网络模型(interactive relation graph attention network,IRGAT)。该模型提取关系类型的特征信息,使其与图注意力网络提取的上下文特征信息交互学习,使它们相互联系,强化各自的特征表示能力。通过方面注意力机制融合特征,再使用分类器捕获情感分类结果。该模型在四个公开数据集上进行了实验,实验结果表明,与现有的方面级情感分析模型相比,IRGAT模型的预测准确率和MF1值分别平均提升了1.52和1.56个百分点。

关键词: 神经网络, 方面级情感分析, 语法依赖树, 交互注意力机制, 图注意力网络

Abstract: Aspect-level sentiment analysis aims to analysis the sentiment polarity of each aspect of online review, and it is a fine-grained sentiment analysis technique. There have been many related studies that have combined syntactic dependency trees with graph attention networks and have applied them to this task with good results. To address the problems that previous studies have ignored information about relation types, have not fully explored the potential semantic information contained in relation types, and have ignored the connection between dependency relations and relation types, an interactive relation graph attention network(IRGAT) model based on graph attention networks is proposed. The model extracts feature information of relational types and then makes them learn interactively with the contextual feature information extracted by the graph attention network, so that they are connected to each other and strengthen their respective feature representations. Finally, the features are fused through the aspect attention mechanism, and then a classifier is used to capture the sentiment classification results. The model is tested on four publicly datasets. The experimental results show that the IRGAT model improves the percent of prediction accuracy and MF1 values by an average of 1.52 and 1.56?percentage points respectively compared to existing aspect-level sentiment analysis models.

Key words: neural network, aspect-based sentiment analysis, dependency tree, interactive attention mechanism, graph attention network