中国科学院自动化研究所机构知识库

Extracting emotion cause and experiencer from text can help people better understand users’ behavior patterns behind expressed emotions. Machine reading comprehension framework explicitly introduces a task-oriented query to boost the extraction task. In practice, how to learn a good task-oriented representation, accurately locate the boundary, and extract multiple causes and experiencers are the key technical challenges. To solve the above problems, this paper proposes BERT-based Machine Reading Comprehension Extraction Model with Multi-Task Learning (BERT-MRC-MTL). It first introduces query as prior knowledge and obtains text representation via BERT. Then, boundary-based and tag-based strategies are designed to select characters to be extracted, so as to extract multiple causes or experiencers simultaneously. Finally, hierarchical multi-task learning structure with residual connection is adopted to combine the answer extraction strategies. We conduct experiments on two public Chinese emotion datasets, and the results demonstrate the efficacy of our proposed model.