Research map / Label-efficient and robust learning / Recognition and long-tail
Utterance-level Emotion Recognition in Conversation with Conversation-level Supervision
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 Dataless ERC (DERC), a weakly-supervised setting where each conversation gets a set of emotions instead of per-utterance labels, and DERC-PL, a progressive self-training method that trains ERC predictors on it.
Problem
Utterance-level emotion labeling in conversations is slow and expensive because each utterance's emotion depends on context; the paper asks whether ERC predictors can be trained from cheaper conversation-level emotion sets.
Approach
Initialize pseudo-utterance emotions from the conversation-level emotion set, then jointly update them with the ERC predictor's confident predictions (restricted to emotions in the set), progressively training from conversations with fewer emotions (lower noise) to those with more.
Key results
- DERC-PL with BERT-base+MLP reaches 58.74 Micro-F1 on MELD vs 56.44 for ItS2CLR and 53.67 for PLOT under weak supervision.
- With SACL as base predictor on IEMOCAP, DERC-PL attains 63.74 Micro-F1 versus 69.08 under full utterance-level supervision.
- Manual conversation-level labeling achieved 94.5% accuracy and roughly 5-11 s per conversation of 1-5 utterances, versus 10-15 s per single utterance for ERC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DERC-PL instead of the strongest reported weakly supervised baseline | DualGATs on IEMOCAP · IEMOCAP test | Micro-F1: 62.60 -> 64.37 | Helped |
| use DERC-PL instead of the strongest reported weakly supervised baseline | DualGATs on EmoryNLP · EmoryNLP test | Micro-F1: 34.41 -> 37.68 | Helped |
| use DERC-PL instead of the strongest reported weakly supervised baseline | DualGATs on MELD · MELD test | Macro-F1: 56.52 -> 59.32 | Helped |
| add pseudo-utterance emotion updating (ablation) | DERC-PL with BERTbase+MLP on IEMOCAP · IEMOCAP test | Macro-F1: 52.84 -> 58.07 | Helped |
| add pseudo-utterance emotion updating (ablation) | DERC-PL with BERTbase+MLP on EmoryNLP · EmoryNLP test | Macro-F1: 26.91 -> 30.28 | Helped |
| add pseudo-utterance emotion updating (ablation) | DERC-PL with BERTbase+MLP on MELD · MELD test | Macro-F1: 51.12 -> 53.54 | Helped |
| add progressive training subsets (ablation) | DERC-PL with BERTbase+MLP on IEMOCAP · IEMOCAP test | Macro-F1: 53.87 -> 58.07 | Helped |
| add progressive training subsets (ablation) | DERC-PL with BERTbase+MLP on EmoryNLP · EmoryNLP test | Macro-F1: 27.91 -> 30.28 | Helped |
| add progressive training subsets (ablation) | DERC-PL with BERTbase+MLP on MELD · MELD test | Macro-F1: 52.07 -> 53.54 | Helped |
| add progressive training subsets and pseudo-utterance emotion updating (ablation) | DERC-PL with BERTbase+MLP on IEMOCAP · IEMOCAP test | Macro-F1: 46.87 -> 58.07 | Helped |
| add progressive training subsets and pseudo-utterance emotion updating (ablation) | DERC-PL with BERTbase+MLP on EmoryNLP · EmoryNLP test | Macro-F1: 24.72 -> 30.28 | Helped |
| add progressive training subsets and pseudo-utterance emotion updating (ablation) | DERC-PL with BERTbase+MLP on MELD · MELD test | Macro-F1: 46.24 -> 53.54 | Helped |
Surprising findings
- Labeling whole conversations as emotion sets took about as long as labeling single utterances for short conversations, while staying accurate (94.5% vs 92.9%).
Limitations
- Gap to fully supervised ERC remains (2.1%-9.1% F1), and sensitive to confidence threshold alpha across datasets.
- Labeling-time evidence comes from a small preliminary study with 20 undergraduate volunteers on MELD only.
- Performance is weaker on imbalanced datasets such as EmoryNLP, where Micro-F1 can be dominated by Neutral.
Experimental setup
Datasets: MELD IEMOCAP EmoryNLP
Benchmarks: MELD IEMOCAP EmoryNLP
Metrics: Micro-F1 Macro-F1
Baselines: PLOT ItS2CLR supervised ERC with BERT+MLP RGAT SACL DualGATs
Builds on: PLOT ItS2CLR focal loss
Closest papers by meaning
- Beyond Single Emotion: Multi-label Approach to Conversational Emotion Recognition (AAAI 2025 · similarity 0.89)
- Adaptive Graph Learning for Multimodal Conversational Emotion Detection (AAAI 2024 · similarity 0.88)
- Dual-View Learning for Conversational Emotion Recognition Through Context and Emotion-Shift Modeling (AAAI 2025 · similarity 0.87)
- From Coarse to Fine: A Distillation Method for Fine-Grained Emotion-Causal Span Pair Extraction in Conversation (AAAI 2024 · similarity 0.85)
- Dynamic Interactive Bimodal Hypergraph Networks for Emotion Recognition in Conversations (AAAI 2025 · similarity 0.85)
- Emotion Rendering for Conversational Speech Synthesis with Heterogeneous Graph-Based Context Modeling (AAAI 2024 · similarity 0.83)
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.