Learning Emotion-Aware Contextual Representations for Emotion Cause Analysis

Anonymous

Abstract: Emotion Cause Analysis has been a key topic in natural language processing. Previous works focus on Emotion Cause Extraction (ECE), a clause-level classification task aimed at extracting causes of certain given emotion in text. The task has been expanded to Emotion Cause Pair Extraction (ECPE) that focus on extracting both emotions and corresponding causes in the context. Most existing methods for the ECPE task implement a joint model that performs extracting and matching of emotion and cause clauses simultaneously. However, we argue that different input features are needed for the two subtasks, thus sharing contextual representations may be suboptimal. In this work, we propose a pipelined approach that builds on two independent pre-trained encoders, in which the emotion extraction model only provide input features for the cause extraction model. Based on a series of careful experiments, we validate that our model can create distinct contextual representations according to specific emotional texts, and thus achieve state-of-the-art performance in both ECE and ECPE tasks, with the absolute F1 improvements of 1.5% and 4.72% over best previous works respectively. Besides, we apply a set of simple clause selection rules to extract multiple pairs in the document, strengthening the applicability of our approach in real world scenarios.