Context-Aware Directed Acyclic Graph Network for Conversational Aspect-Based Sentiment Quadruple Analysis
Qiang Zhang, Jie Zeng, Runze Zhang, Dong Cui · IEEE Access · 2025
Conversational Aspect-based Sentiment Quadruple Analysis (DiaASQ) is a fine-grained sentiment analysis task that aims at extracting targets, aspects, opinions, and sentiments from multi-turn dialogues. Existing methods focus on token-level interaction modeling and neglect complex cross-utterance dependencies. To address this, we propose a context-aware directed acyclic graph network (CA-DAGNet). This model integrates syntax-aware context encoding and directed acyclic graph (DAG) modeling to capture intra-utterance syntactic structures and cross-utterance long-range dependencies. For global modeling, we construct the dialogue as a DAG and combine it with an information propagation mechanism, precisely capturing syntactic dependencies and semantic interactions while dynamically adjusting the scope of information propagation to avoid fixed-window limitations. In addition, we adopt a context filter to retain highly relevant information for the target utterance, suppress redundant noise, and improve the modeling of cross-utterance dependencies. Experiments conducted on Chinese and English datasets demonstrate that the proposed model achieves superior performance.