This repository is for the paper, "MuCAL: Contrastive Alignment for Preference-Driven Knowledge Graph-to-Text Generation" (EMNLP 2025). We basically provide three parts:
- Training and evaluation of MuCAL (KG-Text alignment model)
- Data collection for preference data creation.
- DPO training script.
- BE-MPNet-Hard2 (Hard-MuCAL): A multilingual KG-Text alignment model across 6 languages (Arabic, English, Chinese, French, Spanish, Russian).
- CE-MPNet: A multilingual KG-Text reranker, which has better retrieval performance in common scenarios (without human-curated corruptions).
- CLS-MPNet: A multilingual KG-Text alignment model, trained as a binary classifier. (aligned/not aligned)
- EREDAT: An English KG-Text representation model.
- FactSpotter: A reference-less English KG-Text alignment metric.
- Data Quest-Eval (DQE): A reference-less English Data-to-Text metric based on QA.
- Yifei Song (CNRS, Loria, Université de Lorraine)
- Claire Gardent (CNRS, Loria)
If you find this repo useful, please cite: [TO ADD]