Relational Graph Attention Network Based on Bi-Directional Attention Flow for Aspect-based Sentiment Analysis

Abstract

Aspect-based sentiment analysis(ABSA) aims to predict the sentiment orientation of a specific aspect. Most syntax-based Graph Neural Networks encode the whole tree to obtain syntactic information, Empirically, only part of the parse tree is relevant to this task. Besides, modeling syntactic structure destroys the original word order structure, which leads to partial semantic absence and fails to judge sentiment polarity of the aspects corresponding to the non-opinion word. To address the problems, we propose a Relational Graph Attention Network based on Bi-Directional Attention Flow(BAF-RGAT), Firstly, we generate an aspect-oriented tree structure rooted it at the aspect, Relational Graph Attention Network is applied to encode the new tree and aggregate directed edge information. Secondly, The Bi-Directional Attention Flow network is employed to calculate the bidirectional attention weights between aspect and context words, which comprehensively learns the semantic interactions to generate greater information representations. Then, we introduce an effective consistency training strategy to regularize dropout, alleviating the inconsistency between the training and inference stages caused by the randomness of exploiting dropout. Experimental results on five datasets illustrate the proposed model has comparable effectiveness and superiority.