An interactive multi-head self-attention capsule network model for aspect sentiment classification

The endpoint of aspect-level sentiment classification, a finely ground categorization task in sentiment analysis, is to gauge the polarity for various aspects in context. However, traditional attentional mechanisms still under-explore the relationship between aspect terms and context, and have difficulty in recognizing the overlapping features that arise when expressing multiple sentiment polarities to efficiently obtain deeper semantic representations. To solve these issues, we propose an interactive multi-head self-attention capsule network model (IMHSACap) for aspect sentiment classification. We design Local Context Mask to attenuate the influence of non-local contexts that are far away from the aspect terms, while expanding the influence of local contexts. Then the long-range intrinsic dependencies of global and local contexts are obtained by the interactive attention mechanism, which consists of two parts, Global2Local and Local2Global. The routing algorithm and activation function of the capsule network are optimized to improve the classification accuracy. Hence, experiments on three publicly available datasets are carried out to demonstrate that the IMHSACap model outperforms other baseline approaches for aspect sentiment classification.