EMGCN: Enhancement Graph and Multi-head Attention Graph Convolutional Networks for Aspect-based Sentiment Analysis
Jinhang Chen, Rong Yan · 2024
Aspect-based Sentiment Analysis (ABSA) task aims to predict the sentiment polarity of specific aspects within sentences, uncovering multiple sentiment-object pairs contained in review sentences. Graph neural networks based on dependency trees are popular for addressing ABSA with high effectiveness and practicality. However, how to better utilize the syntactic structure and semantics within dependency trees still remains big challenge. For this purpose, in this paper, we explore to enhance the ability of capture both syntactic dependencies and semantic information by using semantic-enhanced graphs, and propose a novel model named EMGCN (Enhancement Graph and Multi-head Attention Graph Convolutional Networks) model. Additionally, we further focus on enhancing the context awareness by capturing semantic dependencies through the multi-head self-attention mechanism. Experimental validation on five datasets demonstrate the effectiveness of our proposed model.