Shortcut Enhanced Syntactic and Semantic Dual-channel Network for Aspect-based Sentiment Analysis

Abstract

Aspect-based sentiment analysis (ABSA) is a fine-grained task that predicts the sentiment polarity of different aspects in the same sentence. The main challenge is how to build a strong dependency between aspects and sentiment. Recently, the graph neural network (GNN) has become the mainstream trend to extract syntactic dependency relations from the syntactic dependency tree. However, further improvements are hampered by the inherent mistake on the syntactic dependency tree. Consequently, this article presents a dual-channel model to investigate whether considering both syntactic dependency and semantic relevance can further improve the performance. Specifically, we propose a multi-head syntactic graph convolution network (MHGCN) module in the syntactic channel, focusing on different aspects of the syntactic flow in parallel. We also design a syntactic local attention mechanism (Syn-LFAM) and a semantic local attention mechanism (Sem-LFAM) to fully exploit the crucial local information, respectively. Moreover, we use the cross semantic-syntax interaction module and gate fusion mechanism to control the combination of semantic and syntax dynamically. The experimental results show that we utilize less resource consumption, and the final model outperforms the state-of-the-art methods on three of the four publicly available datasets.