Learning Emotion-Aware Contextual Representations for Emotion Cause Analysis
Ying Chen,Wenjun Hou,Shoushan Li,Caicong Wu,Sophia Yat,Mei Lee,J. Devlin,Ming-Wei Chang,Kenton Lee,Chuang Fan,Chaofa Yuan,Jiachen Du,Lin Gui,Jiannan Hu,Yulan He,Ruifeng Xu,Qin Lu,Dongyin Wu,Yu Zhou
2021-01-01
Abstract:Emotion Cause Analysis has been a key topic 001 in natural language processing. Previous 002 works focus on Emotion Cause Extraction 003 (ECE), a clause-level classification task aimed 004 at extracting causes of certain given emotion 005 in text. The task has been expanded to Emo-006 tion Cause Pair Extraction (ECPE) that focus 007 on extracting both emotions and correspond-008 ing causes in the context. Most existing meth-009 ods for the ECPE task implement a joint model 010 that performs extracting and matching of emo-011 tion and cause clauses simultaneously. How-012 ever, we argue that different input features are 013 needed for the two subtasks, thus sharing con-014 textual representations may be suboptimal. In 015 this work, we propose a pipelined approach 016 that builds on two independent pre-trained en-017 coders, in which the emotion extraction model 018 only provide input features for the cause ex-019 traction model. Based on a series of careful 020 experiments, we validate that our model can 021 create distinct contextual representations ac-022 cording to specific emotional texts, and thus 023 achieve state-of-the-art performance in both 024 ECE and ECPE tasks, with the absolute F1 im-025 provements of 1.5% and 4.72% over best pre-026 vious works respectively. Besides, we apply a 027 set of simple clause selection rules to extract 028 multiple pairs in the document, strengthening 029 the applicability of our model in real world sce-030 narios. 031