Clause Fusion-Based Emotion Embedding Model for Emotion-Cause Pair Extraction

Abstract

The goal of emotion-cause pair extraction (ECPE) is to simultaneously extract all emotion clauses and their corresponding cause clauses in a document. In most existing methods, emotion clause representations and cause clause representations are usually obtained separately and are then fed into neural networks. However, the close relationship between emotion and cause is ignored, resulting in an insufficient representation of the emotion clause and the cause clause. To address this problem, we propose a new model, called the clause fusion-based emotion embedding model, to make full use of emotion-related knowledge by utilizing an emotion embedding method when obtaining the representation of the cause clause. First, the emotion word embedding is processed by the emotion clause encoder to get the emotion feature. Second, in clause fusion based emotion embedding network, the emotion clause-level feature in the sliding-window is fused to fused emotion features. The fused emotion features, cause word-level, and emotion word-level feature representation are embedded to get emotion embedding. Third, the emotion embedding is processed to the cause clause feature representation by a bidirectional long short-term memory. Forth, each emotion clause-level feature representation was paired with each cause clause-level feature representation to produce candidate pairs representation. Finally, in the clause pair encoder, a graph convolutional network is applied to model the pair-level context, and then contextual features are extracted for the candidate pairs. Experimental results show that our model achieves state-of-the-art performance on the Chinese benchmark ECPE corpus.