An Attention Mask Mechanism for Conversational Aspect-based Sentiment Quadruple Analysis
Yu Yang, Fang Li · 2025
Conversational Aspect Sentiment Quadruple analysis (DiaASQ) aims to extract quadruples of target, aspect, opinion, and sentiment from dialogue. Unlike Aspect-Based Sentiment Analysis (ABSA), the DiaASQ task involves extracting sentiment quadruples generated by the intersection of multiple utterance, which makes the weights of attention more dispersed during the extraction process. In this paper, we propose a masked attention mechanism customized for DiaASQ. The attention weights are diluted according to the number of threads, and then, through the Attention Mask Word (AM-Word) method, part of the weights of the multi-head attention are restricted, enabling the attention mechanism to focus more on the effective information. Experimental results show that the attention mechanism proposed in our project has improved the overall F1 scores by 5.36% and 3.99% respectively on two datasets compared with the baseline model.