Relational Sequence Encoding Graph Attention Network for Aspect-based Sentiment Analysis

Abstract

Aspect-based sentiment analysis (ABSA) aims to classify all the aspects’ sentiment polarity in a sentence. However, most of the related research didn’t take full advantage of the dependency relations between words, the interaction between the node and its corresponding dependency relations is not taken into consideration, which causes the aspect words mistakenly focus on other context word. To solve these problems, a graph attention network model with relational sequence encoding (RSE-GAT) is proposed in this paper. The model constructs an aspect-rooted syntactic dependency tree, which constructs the representation of virtual connections and relation between aspect words and other words by dependency sequence encoding. At the same time, the composition operator is applied to enhance the interaction between word embedding and relation embedding. Finally, GAT is utilized for feature extraction. The experimental results on three public datasets demonstrate that the proposed RSE-GAT model achieves better results. Compared with the classical GCN-based model ASGCN <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">[5]</sup> , the accuracy on RES14, LAP14 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">[9]</sup> and Twitter <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">[10]</sup> is increased by 3.05%, 2.05% and 3.68% respectively.