Aspect-Level Sentiment Analysis Based on Graph Attention Fusion Networks

Abstract

The goal of aspect-based sentiment analysis is to predict the sentiment polarity of various aspect terms. In aspect-based sentiment classification, the most advanced approach uses the graph convolutional network (GCN) based on integrating the syntactic and semantic structures of sentences. When dealing with word relations, however, GCN only apply the same weight to the edges between words. In contrast, graph attention network (GAT) can assign different weights to different edges according to the importance of the words through an attention mechanism. Therefore, we propose a Graph Attention Fusion Network: in order to establish word dependencies, a Syntactic Graph Attention Network (SyGAT) module with rich syntactic knowledge is designed. A Semantic Graph Attention Network (SeGAT) module with Self-attention is developed to capture the semantic associations of words. Syntactic and semantic features are exchanged through Transitional Graph Attention Network (TraGAT). SyGAT, SeGAT and TraGAT are fused into a new GAT. A convolutional layer is also added to increase the learning ability of n-gram words features. Finally, the fused GAT is merged with the convolutional layer features for sentiment prediction. Experimental results on three benchmark datasets demonstrate the effectiveness of the model in this paper.