Bi-graph attention network for aspect category sentiment classification

Aspect category sentiment classification (ACSC) aims to determine the sentiment polarities of sentences under given aspect categories, which can be used to infer finer-grained information in text sequences. It is widely used in consumer services, healthcare, and elections. Most models ignore the interaction of global sequence context and syntactic structure information in sentences and fail to fully learn the rich relations between word nodes related to specific aspect categories. To tackle these problems, this paper introduces a bi-graph attention network (BiGAT) for ACSC, which constructs two graphs to describe the sequential context information and syntactic structure information in sentences. It utilizes the graph attention network to aggregate neighbor information from each node within a single graph and uses biaffine modules to coordinate heterogeneous information between the sequential- and syntactic-based graphs. The model uses the aspect-specific mask operation and retrieval-based attention mechanism to reduce the effect of noise created by useless information in sentences. Experimental results on the SemEval 2015, SemEval 2016, and MAMS datasets show that BiGAT outperforms other state-of-the-art ACSC models.