Research map / Probabilistic and scientific modeling / Causal discovery
Mastering Context-to-Label Representation Transformation for Event Causality Identification with Diffusion Models
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.
Introduces DiffusECI, the first diffusion model for event causality identification, which gradually transforms event context representations into causal label representations via a denoising process conditioned on the context.
Problem
Event causality identification (ECI) relies on simple feed-forward networks to map noisy, long-context event representations to causal label representations, which requires drastic representation changes and hurts prediction, especially for distant event pairs.
Approach
Replace the one-shot feed-forward context-to-label transformation with a DDPM-style diffusion model whose reverse denoising steps are conditioned on RoBERTa event context representations, plus an auxiliary loss that forces intermediate representations to predict the causal label.
Key results
- On ESL intra+inter-sentence ECI, DiffusECI reaches 63.5 F1, outperforming KADE by about 3 points (reported as 3%).
- On CTB intra-sentence ECI, DiffusECI reaches 75.4 F1, about 8 points above KADE.
- On MECI, DiffusECI achieves 69.2 F1 on English and outperforms RichGCN and KnowBERT across Danish, Spanish, Turkish, and Urdu.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use diffusion-based context-to-label representation transformation (ablation) | DiffusECI · ESL intra + inter-sentence | F1: 57.7 -> 63.5 | Helped |
| add explicit transformation from event context representations into causal label representations (ablation) | RoBERTa ECI · ESL intra + inter-sentence | F1: 42.2 -> 56.5 | Helped |
| include event mentions in the label-representation template (ablation) | DiffusECI · ESL intra + inter-sentence | F1: 58.7 -> 63.5 | Helped |
| use DiffusECI instead of DPJL | ESL intra-sentence ECI · ESL intra-sentence | F1: 67.9 -> 71.4 | Helped |
| use DiffusECI instead of KADE | ESL inter-sentence ECI · ESL inter-sentence | F1: 60.5 -> 60.9 | Helped |
| use DiffusECI instead of KADE | ESL intra + inter-sentence ECI · ESL intra + inter-sentence | F1: 59.8 -> 63.5 | Helped |
| use DiffusECI instead of KADE | CTB intra-sentence ECI · CTB intra-sentence | F1: 66.7 -> 75.4 | Helped |
| use DiffusECI instead of XLM-RoBERTa | MECI English · MECI English test | F1: 53.7 -> 69.2 | Helped |
| use DiffusECI instead of RichGCN | MECI Danish · MECI Danish test | F1: 38.9 -> 47.4 | Helped |
| use DiffusECI instead of RichGCN | MECI Spanish · MECI Spanish test | F1: 52.8 -> 55.8 | Helped |
| use DiffusECI instead of RichGCN | MECI Turkish · MECI Turkish test | F1: 56.7 -> 58.7 | Helped |
| use DiffusECI instead of RichGCN | MECI Urdu · MECI Urdu test | F1: 45.1 -> 55.2 | Helped |
Surprising findings
- Gains over the XLM-RoBERTa-based baseline are much larger for event pairs more than 15 words apart (e.g., +35.3 F1 on English) than for closer pairs.
Limitations
- Reverse sampling requires many steps (T=1000, L=100), adding inference cost.
- Gains on ESL inter-sentence and MECI non-English languages are less uniform; some baselines are close.
Experimental setup
Datasets: EventStoryLine (ESL) v0.9 Causal-TimeBank (CTB) MECI
Benchmarks: EventStoryLine (ESL) Causal-TimeBank (CTB) MECI
Metrics: precision recall F1
Baselines: LSTM Seq LR+ LIP ML KnowBERT KnowDis CauSeRL LearnDA RichGCN ERGO CF-ECI
Builds on: DDPM RoBERTa LearnDA GenECI RichGCN
Closest papers by meaning
- A Joint Framework with Heterogeneous-Relation-Aware Graph and Multi-Channel Label Enhancing Strategy for Event Causality Extraction (AAAI 2024 · similarity 0.86)
- Causal Temporal Representation Learning with Nonstationary Sparse Transition (NeurIPS 2024 · similarity 0.83)
- Measure-Theoretic Anti-Causal Representation Learning (NeurIPS 2025 · similarity 0.82)
- Learning Linear Causal Representations from General Environments: Identifiability and Intrinsic Ambiguity (NeurIPS 2024 · similarity 0.82)
- On the Identification of Temporal Causal Representation with Instantaneous Dependence (ICLR 2025 · similarity 0.82)
- Towards Identifiability of Hierarchical Temporal Causal Representation Learning (NeurIPS 2025 · similarity 0.82)
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.