Research map / Label-efficient and robust learning / Knowledge distillation
From Coarse to Fine: A Distillation Method for Fine-Grained Emotion-Causal Span Pair Extraction in Conversation
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.
Builds FG-RECCON, a 16-label fine-grained emotion and causal-span version of RECCON, and proposes KD-ECS, where a LLaMA teacher predicts causal connectives between utterances and distills them into a RoBERTa student that jointly extracts fine-grained emotions and causal spans.
Problem
Emotion-causal span pair extraction in conversations: existing work uses only six coarse emotion categories and utterance-level causes, which miss nuanced emotions and within-utterance causal spans.
Approach
Distill utterance-level causal-connective knowledge from a prompt-based LoRA-tuned LLaMA teacher (trained on coarse ECPE data) into a RoBERTa student via KL divergence on connective predictions, so the student jointly predicts 16 fine-grained emotions and causal spans.
Key results
- On FG-RECCON, KD-ECS reaches 48.26% emotion accuracy, 33.85 EM, 58.62 Pos. F1 and 74.16 macro F1, the best among non-ChatGPT systems (KEC: 30.19 EM, 56.94 Pos. F1).
- Ablation on FG-RECCON: removing KD drops EM from 33.85 to 30.49 and macro F1 from 74.16 to 72.74.
- On the coarse RECCON ECPE task, KD-ECS achieves 61.04 emotion accuracy and 76.69 macro F1, a small gain over KEC (76.11 macro F1).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use KD-ECS instead of ChatGPT | Emotion and causal span extraction on FG-RECCON · FG-RECCON | EM: 33.78 -> 33.85 | Helped |
| use KD-ECS instead of ChatGPT | Emotion and causal span extraction on FG-RECCON · FG-RECCON | Pos. F1: 58.57 -> 58.62 | Helped |
| use KD-ECS instead of ChatGPT | Emotion and causal span extraction on FG-RECCON · FG-RECCON | macro F1: 73.99 -> 74.16 | Helped |
| use KD-ECS instead of ChatGPT | Emotion and causal utterance extraction on RECCON · RECCON | Pos. F1: 60.44 -> 60.51 | Helped |
| use KD-ECS instead of KEC | Emotion and causal utterance extraction on RECCON · RECCON | macro F1: 76.11 -> 76.69 | Helped |
| add knowledge distillation from the teacher model (ablation) | KD-ECS on FG-RECCON · FG-RECCON | EM: 30.49 -> 33.85 | Helped |
| add knowledge distillation from the teacher model (ablation) | KD-ECS on FG-RECCON · FG-RECCON | Pos. F1: 56.27 -> 58.62 | Helped |
| add knowledge distillation from the teacher model (ablation) | KD-ECS on FG-RECCON · FG-RECCON | macro F1: 72.74 -> 74.16 | Helped |
| add causal connective-word prediction (ablation) | KD-ECS on FG-RECCON · FG-RECCON | EM: 29.13 -> 33.85 | Helped |
| add causal connective-word prediction (ablation) | KD-ECS on FG-RECCON · FG-RECCON | Pos. F1: 55.34 -> 58.62 | Helped |
| add causal connective-word prediction (ablation) | KD-ECS on FG-RECCON · FG-RECCON | macro F1: 71.59 -> 74.16 | Helped |
| add utterance-level causal-relation recognition (maskcau) (ablation) | KD-ECS on FG-RECCON · FG-RECCON | EM: 28.68 -> 33.85 | Helped |
Surprising findings
- ChatGPT achieves much higher fine-grained emotion accuracy (67.20%) than KD-ECS (48.26%), though it trails on causal span EM.
Limitations
- Fine-grained emotion labels were produced by ChatGPT and then manually reviewed, with ChatGPT accuracy of only 69.3% on validation.
- Emotion accuracy of KD-ECS lags behind ChatGPT on the fine-grained task.
- Teacher scale benefits saturate; the student cannot exploit very large teachers.
Experimental setup
Datasets: RECCON FG-RECCON
Benchmarks: FG-RECCON RECCON
Metrics: accuracy exact match positive F1 negative F1 macro F1
Baselines: RoBERTa SpanBERT ECPE-2D MGSAG UECA KEC LLaMA-7B LLaMA-13B ChatGPT
Builds on: RECCON ECPE prompt learning for implicit discourse relations knowledge distillation (Hinton et al.)
Closest papers by meaning
- Utterance-level Emotion Recognition in Conversation with Conversation-level Supervision (AAAI 2025 · similarity 0.85)
- Dual-View Learning for Conversational Emotion Recognition Through Context and Emotion-Shift Modeling (AAAI 2025 · similarity 0.85)
- Adaptive Graph Learning for Multimodal Conversational Emotion Detection (AAAI 2024 · similarity 0.85)
- Federated Dialogue-Semantic Diffusion for Emotion Recognition under Incomplete Modalities (NeurIPS 2025 · similarity 0.84)
- AffectGPT: A New Dataset, Model, and Benchmark for Emotion Understanding with Multimodal Large Language Models (ICML 2025 · similarity 0.83)
- Harnessing Holistic Discourse Features and Triadic Interaction for Sentiment Quadruple Extraction in Dialogues (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.