Targeted sentiment classification with multi-attention network

Abstract

Targeted sentiment classification aims at recognising the sentiment polarity of specific targets. However, existing methods mainly depend on a crude attention mechanism, while neglecting the mutual effects between target and context. In order to solve this problem, this paper introduces a Multi-Attention Network (MAN) for aspect level sentiment classification. We jointly modelled intra-level and inter-level attentional components to capture the interaction between target and context. The former attention mechanism pays attention to the context relation, whereas the latter attention mechanism considers important parts in a sentence. The experimental conducted on laptop, restaurant and Twitter data sets indicate that our model surpasses the baseline model.