Semantic enhanced bi-syntactic graph convolutional network for aspect-based sentiment analysis

Abstract

Previous work on fine-grained sentiment analysis focuses on establishing the semantic correlations between words by means of attention mechanisms. More recently, effects of syntax-based models, applying graph convolution operation over dependency trees, are highlighted due to their superiority. However, these methods still have deficiencies. For one thing, little aspect-specific information is considered during semantic modeling of contextual words, which introduces irrelevant noise toward the aspect. For another, current syntax-based approaches either ignore the syntactic constituent knowledge, or fail to maintain the syntactic information from the reconstructed constituent tree. As such, no relation among words, phrases and clauses is built. In this work, a Semantic Enhanced Bi-Syntax Graph Convolutional Network (SEBS-GCN) is proposed to enhance semantics of context to the aspect, and capture the sentiment relevance among words, phrases and clauses. Specifically, we devise an aspect-aware gated mechanism to obtain the aspect-aware feature, based on the semantic correlations between the specific aspect and its contexts. Furthermore, the syntax information of the constituent tree is sufficiently exploited to analyze the hierarchical structure and the logical relation among words, phrases and clauses, based on which to capture the sentiment clues of the aspect.