Reproduce “Multimodal Temporal Modeling for Continuous Group Emotion Recognition in Multi-party Dialogues” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Soma Iwata, Koji Inoue, Muyun Wu, Taiga Mori, Divesh Lala, Tatsuya Kawahara
The paper introduces one-second group-emotion annotations for Arousal, Valence, and Mixed states in 36 Japanese discussions involving 12 three-person groups. A temporal Transformer combines participants’ audio and video histories to estimate distributions. With 20-second input, it achieved Arousal CCC of 0.807 versus 0.747 for the tested LLM-based system; this comparison is limited, and Mixed cases had higher error.
Key insights
- Audio is the stronger unimodal signal, but combining audio and video provides the best continuous recognition performance, indicating that visual cues add complementary information.
- Extending context from five to twenty seconds yields only modest and metric-specific gains, suggesting that longer history does not uniformly improve group-emotion prediction.
- Recognition is substantially less reliable during high-divergence intervals, showing both the difficulty of assigning a single group emotion and the need for models that explicitly represent participant disagreement.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.