Improving context and syntactic dependency for aspect-based sentiment analysis using a fused graph attention network

Aspect-based sentiment analysis is a hot research issue, which aims to determine the polarity of sentiments in a particular aspect of a review. Some recent methods adopt attention-based neural networks or graph neural networks to connect aspects implicitly with opinion words, achieving better results. However, the existing methods are usually limited to one of these. Also, multiple aspects may be present in one review, and previous methods can often confuse the connections between aspects and opinions. To address these problems, we explore a method of integrating a dependency structure with a graph attention network for sentence representation learning. For this reason, we propose a Fused Graph Attention Network model, which employs a bidirectional gated recurrent unit to learn sentence representation features, and further enhances the embedding through a graph attention layer. The experimental results demonstrate that the proposed model can better establish the relationship between the aspects and opinion words in a sentence.