Bidirectional Edge-Enhanced Graph Convolutional Networks for Aspect-based Sentiment Classification

Aspect-based sentiment classification aims to predict the sentiment polarity of an aspect term in a sentence. It has been verified that syntactic dependent trees, especially integrating with graph convolutional networks (GCN) can provide crucial syntactic features for sentiment classification. However, it can not consider the dependency label information between the aspect and context in the sentence, which contains rich semantic information. In this paper, we propose a bidirectional edge-enhanced graph convolutional networks (BE-GCN), which combines the syntactic structure and the dependency label information effectively. Specifically, we design an edge-enhanced module to dynamically update the dependency edge according to dependency relation and contextual information. Then the dependency edge can update the word representation reversely, thereby selectively outputting sentiment features according to the given aspect. Comparison experiments demonstrate the effectiveness of our model using dependency label information and syntactic structure.