Aspect-Level Sentiment Classification with Dependency Rules and Dual Attention

Abstract

Aspect-level sentiment classification aims to predict the sentiment polarity towards the given aspects of sentences. Neural network models with attention mechanism have achieved great success in this area. However, existing methods fail to capture enough aspect information. Besides, it is hard for simple attention mechanism to model complex interaction between aspects and contexts. In this paper, we propose a Segment Model with Dual Attention (SegM-DA) to tackle these problems. We combine deep learning models with traditional methods by defining dependency rules to extract auxiliary words, which helps to enrich aspect information. In addition, in order to model structural relation between aspects and contexts, we introduce dependent attention mechanism. Coupled with standard attention mechanism, we establish the dual attention mechanism, which models the interaction from both word- and structure- dependency. We perform aspect-level sentiment classification experiments on two real datasets. The results show that our model can achieve the state-of-the-art performance.