DTS: A Decoupled Task Specificity Approach for Aspect Sentiment Triplet Extraction

Abstract

Aspect Sentiment Triplet Extraction (ASTE) aims to extract aspect terms, opinion terms, and their corresponding sentiment polarities from customer reviews. Contemporary table-filling approaches prominently construct a task-sharing table for distinct subtasks to exploit the interaction between entities. However, the single table mechanism neglects the need for task-specific knowledge, inevitably causing feature confusion. In this paper, we introduce an innovative method named Decoupled Task Specificity (DTS) to address these issues. Specifically, this model builds a term expert table to learn semantically and syntactically enhanced knowledge for term extraction while constructing another sentiment expert table for sentiment classification by incorporating more comprehensive contextual knowledge. These two task expert tables learn task-specific knowledge to mitigate the gap between knowledge and subtasks. Moreover, two task-specific decoders, based on different region detection strategies, are designed for the decoding of expert tables. Overall, these modules jointly achieve subtask specificity throughout the whole process. Experimental results on public benchmark datasets show the effectiveness and exceptional performance of the DTS.