A Multi-granularity Network for Emotion-Cause Pair Extraction via Matrix Capsule

The task of Emotion-Cause Pair Extraction (ECPE) aims at extracting the clause pairs with the corresponding causality from the text.Existing approaches emphasize their multi-task settings. We argue that the clause-level encoders are ill-suited to the ECPE task where text information has many granularity features. In this paper, we design a Matrix Capsule-based multi-granularity framework (MaCa) for this task. Specifically, we first introduce a word-level encoder to obtain the token-aware representations. Then, two sentence-level extractors are used to generate emotion prediction and cause prediction. Finally, to obtain more fine-grained features of clause pairs, the matrix capsule is introduced, which can cluster the relationship of each clause pair. The empirical results on the widely used ECPE dataset show that our framework significantly outperforms most current methodsin the Emotion-Cause Extraction (ECE) and the challenging ECPE task.