Research map / Graph learning / Knowledge and temporal graphs
Harnessing Holistic Discourse Features and Triadic Interaction for Sentiment Quadruple Extraction in Dialogues
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes H2DT, a token-level reply-speaker heterogeneous graph with heterogeneous attention plus a triadic scorer for token-pair labeling, achieving state-of-the-art sentiment quadruple extraction on the DiaASQ benchmark (Chinese and English).
Problem
Dialogue aspect-based sentiment quadruple extraction (DiaASQ) requires modeling dialogue discourse (reply and speaker structure) and capturing interactions among the target, aspect, and opinion items so that the quadruple is cohesive rather than assembled from weak pairwise predictions.
Approach
Builds a token-level reply-speaker (R-S) heterogeneous graph with meta-path attention to inject discourse features into token representations, and a triadic scorer that routes each token pair's relation score through an intermediate third token via triaffine scores, strengthening cohesion within quadruples.
Key results
- On DiaASQ Chinese, quadruple F1 of 40.34 vs 34.94 for Meta-WP (+5.4).
- On DiaASQ English, quadruple F1 of 39.01 vs 33.31 for Meta-WP (+5.7).
- Removing the triadic scorer drops quadruple F1 by 2.71 (ZH) and 2.61 (EN); removing the R-S graph drops it by 3.32 (ZH) and 1.83 (EN).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use H2DT instead of Meta-WP | DiaASQ · DiaASQ Chinese | quadruple F1: +5.40 | Helped |
| use H2DT instead of Meta-WP | DiaASQ · DiaASQ English | quadruple F1: +5.70 | Helped |
| add the R-S graph (ablation) | H2DT · DiaASQ Chinese | quadruple F1: 37.02 -> 40.34 | Helped |
| add the R-S graph (ablation) | H2DT · DiaASQ English | quadruple F1: 37.18 -> 39.01 | Helped |
| add the self-link to the R-S graph (ablation) | H2DT · DiaASQ Chinese | quadruple F1: 39.10 -> 40.34 | Helped |
| add the self-link to the R-S graph (ablation) | H2DT · DiaASQ English | quadruple F1: 38.86 -> 39.01 | Helped |
| add first-order reply and speaker links to the R-S graph (ablation) | H2DT · DiaASQ Chinese | quadruple F1: 38.70 -> 40.34 | Helped |
| add first-order reply and speaker links to the R-S graph (ablation) | H2DT · DiaASQ English | quadruple F1: 38.44 -> 39.01 | Helped |
| add second-order reply-speaker paths to the R-S graph (ablation) | H2DT · DiaASQ Chinese | quadruple F1: 38.97 -> 40.34 | Helped |
| add second-order reply-speaker paths to the R-S graph (ablation) | H2DT · DiaASQ English | quadruple F1: 38.30 -> 39.01 | Helped |
| add the triadic interaction term (ablation) | H2DT · DiaASQ Chinese | quadruple F1: 37.63 -> 40.34 | Helped |
| add the triadic interaction term (ablation) | H2DT · DiaASQ English | quadruple F1: 36.40 -> 39.01 | Helped |
Surprising findings
- Model has fewer parameters than Meta-WP (111.81M vs 114.05M) while being faster in training throughput, despite added token-level interactions.
Limitations
- Item detection gains are modest, since the task is relatively mature.
- Triadic scorer gains are smaller for T-A pairs than for T-O and A-O pairs.
- Evaluation is limited to the single DiaASQ benchmark.
Experimental setup
Datasets: DiaASQ
Benchmarks: DiaASQ (Chinese) DiaASQ (English)
Metrics: precision recall F1
Baselines: CRF-Extract SpERT ParaPhrase Span-ASTE Meta-WP
Builds on: Meta-WP Heterogeneous Graph Attention Network Triaffine attention RoBERTa RoFormer
Closest papers by meaning
- SimRP: Syntactic and Semantic Similarity Retrieval Prompting Enhances Aspect Sentiment Quad Prediction (AAAI 2025 · similarity 0.85)
- A Novel Energy Based Model Mechanism for Multi-Modal Aspect-Based Sentiment Analysis (AAAI 2024 · similarity 0.83)
- You Only Read Once: Constituency-Oriented Relational Graph Convolutional Network for Multi-Aspect Multi-Sentiment Classification (AAAI 2024 · similarity 0.83)
- From Coarse to Fine: A Distillation Method for Fine-Grained Emotion-Causal Span Pair Extraction in Conversation (AAAI 2024 · similarity 0.83)
- DanceHA: A Multi-Agent Framework for Document-Level Aspect-Based Sentiment Analysis (AAAI 2026 · similarity 0.82)
- TextGT: A Double-View Graph Transformer on Text for Aspect-Based Sentiment Analysis (AAAI 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.