Official PyTorch implementation of Divide and Refine (DnR), a plug-and-play framework for improving multimodal representations in Emotion Recognition in Conversation (MERC) by explicitly modeling uniqueness, redundancy, and synergy across text, audio, and visual modalities.

📄 Accepted at WACV 2026 Divide and Refine: Enhancing Multimodal Representation and Explainability for Emotion Recognition in Conversation

⸻

Overview

DnR is a two-phase framework: • Divide — Decomposes each modality into: • Unique (U) • Redundant (R) • Synergistic (S) • Refine — Strengthens representations via redundancy-focused augmentation and contrastive learning.

The method is lightweight, model-agnostic, and can be integrated into existing MERC backbones.

⸻

Features • Plug-and-play with common models (MMGCN, DialogueGCN, MM-DFN, SDT, etc.) • Consistent gains on IEMOCAP and MELD • Improved robustness to missing/noisy modalities • Better interpretability of multimodal signals

⸻

Installation

git clone https://github.com/mattam301/DnR-WACV2026.git cd DnR-WACV2026 pip install -r requirements.txt

⸻

End2End training bash scripts/atv.sh

⸻

Citation

@article{mai2026divide, title={Divide and Refine: Enhancing Multimodal Representation and Explainability for Emotion Recognition in Conversation}, author={Mai, Anh-Tuan and Nguyen, Cam-Van Thi and Le, Duc-Trong}, journal={arXiv preprint arXiv:2601.14274}, year={2026} }

:::