Harnessing Syntax GCN and Multi-View Interaction for Conversational Aspect-Based Quadruple Sentiment Analysis

Conversational Aspect-Based Sentiment Analysis (DiaASQ) aims to extract fine-grained sentiment quadruples {target, aspect, opinion, polarity} from multiple segments of dialogue. The composition of these quadruples is often not limited to a single utterance but may span across the entire conversation. Therefore, the analysis needs to consider both the syntactic structure of individual utterances and the interactions between different views. However, previous studies often lack modeling of dialogue features and overlook the interdependence between utterances. To address this, this paper introduces the Syntax and Consecutive Multi-View Network. This model captures the syntactic dependencies among utterances using Graph Convolutional Networks (GCN) and employs multi-view interactions along with three consecutive multi-head attention modules to construct contextual connections within the dialogue. Finally, a triple scorer is used to decode the quadruples, thereby enhancing the weakly labeled relationships within the quadruples. Extensive experiments on standard datasets demonstrate that the proposed model significantly outperforms baseline methods across multiple dimensions.