Multi-dimensional Feature Interaction for Conversational Aspect-Based Quadruple Sentiment Analysis

Conversational aspect-level quadruple sentiment analysis (DiaASQ) is proposed as a new task that aims to extract target-aspect-opinion-sentiment quadruples in dialogues. However, this task faces the problem of complex context matching and multiple utterance feature modeling, which creates difficulties in extracting quadruples from multiple intersecting utterances. To address this problem, this paper proposes a Multi-dimensional Dialogue Feature Interaction (MDFI) approach. This method models dialogue features through an interactive network structure to capture interactions between utterance features. The approach adds two layers of ResNet to achieve deep association fusion based on multi-head self-attention. It superimposes the associated features of replies, speakers, and dialogue threads layer by layer and enhances the capability of conversation representation through linear augmentation. Our model outperforms the DiaASQ benchmark model in global utterance, intra-utterance, and cross-utterance quadruple extraction. In particular, the ZH dataset shows an improvement of 7.42 in global utterance and 9.66 in cross-utterance.