A Joint Training Dual-MRC Framework for Aspect Based Sentiment Analysis

Aspect based sentiment analysis (ABSA) involves three fundamental subtasks: aspect term extraction, opinion and aspect-level classification. Early works only focused on solving one of these subtasks individually. Some recent work a combination two subtasks, e.g., extracting terms along with polarities or the pair-wisely. More recently, triple extraction task has been proposed, i.e., (aspect term, polarity) triples from sentence. However, previous approaches fail to solve all in unified end-to-end framework. In this paper, we propose complete solution for ABSA. We construct machine reading comprehension (MRC) problems, by joint training BERT-MRC models parameters sharing. conduct experiments results several benchmark datasets demonstrate effectiveness our proposed framework, which significantly outperforms existing state-of-the-art methods.