A Hybrid Semantic-Topic Co-encoding Network for Social Emotion Classification

Social emotion classification is to predict the distribution of readers' emotions evoked by a document (e.g., news article). Previous work has shown that both semantic and topical information can help improve classification performance. However, many existing topic-based neural models represent the topical feature of document with only topic probabilities, ignoring the fine-grained semantic feature of terms in each topic. Moreover, traditional RNN-based semantic networks often face the disadvantage of slow training In this paper, we propose a hybrid semantic-topic co-encoding network. It contains a semantics-driven topic encoder to compose topic embeddings, and also utilizes a forward self-attention network to exploit document semantics. Finally, the semantic and topical features of the document are adaptively integrated through a gate layer, which generates the document representation for social emotion classification. Experimental results on three public datasets show that the proposed model outperforms the state-of-the-art approaches in terms of higher accuracy and average Pearson correlation coefficient. Moreover, the proposed model runs fast and with better explainability.