Hierarchical Interactive Network for joint aspect extraction and sentiment classification

Abstract

Aspect-based sentiment analysis (ABSA) aims at identifying the opinion aspects (aspect extraction) and sentiment polarities toward corresponding aspects (sentiment classification) from a sentence. Recently, some span-based methods, which first extract aspects by detecting aspect boundaries and then predict the span-level sentiments, have achieved promising results. However, the correlations between aspect extraction and sentiment classification have not been explicitly explored. For example, sentimental expressions can be better understood if specific aspects are given. In contrast, aspects can be better detected if we know where the sentimental expressions are located. Therefore, we propose a novel Hierarchical Interactive Network (HIN) to enhance the internal connections between aspect extraction and sentiment classification. To this end, the HIN jointly learns the aspect extractor and sentiment classifier across two layers hierarchically. The former learns some shallow-level interactions via a cross-stitch mechanism, and the latter learns deep-level interactions between two subtasks by using mutual information maximization technology. Extensive experiments on three real-world datasets demonstrate the HIN’s superior performance.