Aspect-based Sentiment Classification via Syntactic-aware Generative Model

Aspect-based sentiment classification (ABSC) aims to identify the sentiment polarity towards the speicified aspects in a sentence. Most existing works investigate ABSC in a discrimination manner, ignoring the rich semantic and syntactic knowledge in ABSA problems. To address these issues, we propose a generative framework with syntactic-aware graph attention networks to resolve the ABSC task. The framework leverages graph attention networks to incorporate syntactic knowledge into a pretrained generative model and addresses ABSA in a generative manner. In addition, considering the importance of syntactic information for solving ABSC tasks, we construct a multi-layer syntactic-aware graph attention network and effectively integrate it into the pre-trained generative models. We conduct experiments on five commonly used real-world datasets for ABSC task and use accuracy and Macro-F₁ scores for evaluation. Experimental results evaluated by accuracy and Macro-F₁ demonstrate that our approach achieves better performance than previous methods.