Research map / Graph learning / Knowledge and temporal graphs
Dynamic Interactive Bimodal Hypergraph Networks for Emotion Recognition in Conversations
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 DIB-HGCN, a bimodal hypergraph network that separates monologic and dialogic conversation interactions and applies dynamic versus fragment context windows to each, combined with an undirected graph branch for emotion recognition in conversations.
Problem
Emotion recognition in conversations (ERC) with existing graph methods that do not distinguish interaction patterns (monologue vs. dialogue) and use a single window partitioning mechanism despite differing emotional transfer rates.
Approach
Constructs two types of sub-hypergraphs: monologic sub-hypergraphs with dynamic windows over consistent same-speaker runs, and dialogic sub-hypergraphs with fragment windows to limit noise from frequent emotional transfers; fused with a parallel attention-based undirected graph branch.
Key results
- On IEMOCAP, DIB-HGCN reaches 72.58% weighted accuracy and 72.46% weighted F1, above the best listed baseline MA-CMU-SGR (72.4 / 71.6).
- On MELD, DIB-HGCN reaches 68.01% accuracy and 66.61% F1, above SCMFN (67.01 / 66.25) and M3NET reimplementation (67.51 / 65.97).
- Removing the undirected graph branch drops IEMOCAP accuracy to 67.41%, the largest ablation drop.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DIB-HGCN instead of MA-CMU-SGR | IEMOCAP · IEMOCAP | accuracy: 72.4 -> 72.58 | Helped |
| use DIB-HGCN instead of MA-CMU-SGR | IEMOCAP · IEMOCAP | F1: 71.6 -> 72.46 | Helped |
| use DIB-HGCN instead of M3NET | MELD · MELD | accuracy: 67.51 -> 68.01 | Helped |
| use DIB-HGCN instead of SCMFN | MELD · MELD | F1: 66.25 -> 66.61 | Helped |
| add the hypergraph branch (ablation) | DIB-HGCN · IEMOCAP | accuracy: 71.04 -> 72.58 | Helped |
| add the hypergraph branch (ablation) | DIB-HGCN · MELD | accuracy: 67.70 -> 68.01 | Helped |
| add the undirected graph branch (ablation) | DIB-HGCN · IEMOCAP | accuracy: 67.41 -> 72.58 | Helped |
| add the undirected graph branch (ablation) | DIB-HGCN · MELD | accuracy: 67.13 -> 68.01 | Helped |
| add sub-hypergraph construction (ablation) | DIB-HGCN · IEMOCAP | accuracy: 70.86 -> 72.58 | Helped |
| add sub-hypergraph construction (ablation) | DIB-HGCN · MELD | accuracy: 67.36 -> 68.01 | Helped |
| use dynamic windows for dialogic sub-hypergraphs (ablation) | DIB-HGCN · IEMOCAP | accuracy: 72.58 -> 69.66 | Hurt |
| use fragment windows for monologic sub-hypergraphs (ablation) | DIB-HGCN · IEMOCAP | accuracy: 72.58 -> 71.10 | Hurt |
Surprising findings
- Using dynamic windows in dialogic sub-hypergraphs or fragment windows in monologic sub-hypergraphs both reduced performance, supporting the window-to-interaction-pattern matching.
Limitations
- Evaluated only on complete multimodal data; does not handle missing modalities.
- Weak performance on minority classes such as fear and disgust in MELD.
- Gains over MA-CMU-SGR on IEMOCAP are small (72.58 vs 72.4 accuracy).
Experimental setup
Datasets: IEMOCAP MELD
Benchmarks: IEMOCAP MELD
Metrics: weighted accuracy weighted F1
Baselines: A-DMN COGMEN CORECT MM-DFN SCMFN M3NET DCGCN MA-CMU-SGRNet
Builds on: M3NET DCGCN MMGCN HGNN
Closest papers by meaning
- Adaptive Graph Learning for Multimodal Conversational Emotion Detection (AAAI 2024 · similarity 0.90)
- BIG-FUSION: Brain-Inspired Global-Local Context Fusion Framework for Multimodal Emotion Recognition in Conversations (AAAI 2025 · similarity 0.86)
- Revisiting Multimodal Emotion Recognition in Conversation from the Perspective of Graph Spectrum (AAAI 2025 · similarity 0.86)
- AffectGPT: A New Dataset, Model, and Benchmark for Emotion Understanding with Multimodal Large Language Models (ICML 2025 · similarity 0.86)
- Dual-View Learning for Conversational Emotion Recognition Through Context and Emotion-Shift Modeling (AAAI 2025 · similarity 0.85)
- Utterance-level Emotion Recognition in Conversation with Conversation-level Supervision (AAAI 2025 · similarity 0.85)
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.