A Context Enhanced Attention Network for Aspect-Based Sentiment Classification

Aspect-based sentiment classification is a fine-grained task in sentiment analysis, it aims to detect the sentiment polarity of an aspect in a given context. Existing approaches mostly focused on the processing of aspect words, which often fail to adequately model the context via aspect terms. However, the surrounding words in context usually have an important impact on the sentiment polarity of aspect words. This paper proposes a context enhanced attention network (CE-ATT) to enhance context modeling for aspect-based sentiment classification. Firstly, the text sequence was divided into left and right contexts by a given aspect. Then, the Bidirectional Long Short Term Memory (BiLSTM) was used to model the left context and aspect, the right context and aspect, and both sides context and aspect, respectively. After summing these contextual representations, an attention mechanism was used to concentrate on the parts that have an important influence on the aspect word. The experimental results on three benchmark datasets demonstrate the effectiveness of the CE-ATT method.