A multi-task learning framework for end-to-end aspect sentiment triplet extraction

Abstract

Aspect sentiment triplet extraction (ASTE) is a significant and challenging task in aspect-based sentiment analysis, which aims to summarize people’s opinions by extracting triplets consisting of opinion targets, opinion expressions, and sentiment polarities. In this paper, we propose a novel multi-task learning framework to achieve end-to-end ASTE. We decompose ASTE into three subtasks, namely target tagging, opinion tagging, and sentiment tagging. In target tagging and opinion tagging, we adopt the BIO tagging scheme to detect the boundaries of opinion targets and opinion expressions. In sentiment tagging, we introduce a target-aware tagging scheme, which utilizes a series of target-specific tag sequences to identify the correspondences between opinion targets and opinion expressions, and determine their sentiment polarities. We conduct extensive experiments on four benchmark datasets. The experimental results show that our framework achieves consistently superior results. Compared with existing methods, our method has better performance in extracting overlapping triplets and identifying long-range correspondences. Further analysis demonstrates the effectiveness of our framework.