Identifying implicit emotions via hierarchical structure and rhetorical correlation

Implicit emotional expressions, without using explicit emotion words, usually depend on rhetorics to vividly show the user’s emotions. Sentences carved with specific rhetorics tend to express certain types of emotions. Moreover, a hierarchical structure between emotional pleasure valences and categories exists in psychological sciences from coarse to fine, which helps human understand textual emotions. However, existing implicit emotion identification models ignore the hierarchy structure and the correlations between emotions and rhetorics. In this paper, we propose an implicit emotion identification model via hierarchical structure and rhetorical correlation, which consists of two major layers. Specifically, a hierarchical layer is designed to leverage hierarchical structure and provide coarse-grained emotional valences for identifying emotions, and a correlation layer to learn the latent correlation between emotions and rhetorics. Finally, supported by two layers, a novel multi-task learning model is proposed to train three related identification tasks of pleasure valences, emotions and rhetorics simultaneously, thus improving the overall performance of the emotion identification problem. Experimental results on the implicit emotion dataset demonstrate that the proposed model achieves 89.78% and 88.74% in terms of micro-F1 and weight-F1 metric respectively, outperforming the state-of-the-art methods consistently.