Parallel Universes, Parallel Languages: A Comprehensive Study on LLM-based Multilingual Counterfactual Example Generation

2026-04-06: The paper is accepted at ACL 2026 main conference!

Note: human-cf folder contains the link to the human-annotated counterfactuals provided by Kaushik et al., 2020.

📥 Dataset

• XNLI: https://huggingface.co/datasets/facebook/xnli
• SIB200: https://github.com/dadelani/sib-200

Target languages includes Arabic, German, English, Spanish, Hindi and Swahili.

🚀 Environment Setup

conda create -n <name> python=3.10.0
conda activate <name>
pip install -r requirements.txt

📂 Project Structure

├── data
├── evaluation
│   └──  cf_prediction
├── results
│   ├── CDA
│   │    ├── translated_counterfactuals    
│   │    └── directly generated counterfactuals
│   ├── predictions
│   │    ├── cda    
│   │    └── dataset predictions
│   └── translated_counterfactual
│       ├── human_forward_translation    
│       ├── sib200    
│       └── XNLI
└── experiments

📝 Citation

@misc{wang2026paralleluniversesparallellanguages,
      title={Parallel Universes, Parallel Languages: A Comprehensive Study on LLM-based Multilingual Counterfactual Example Generation}, 
      author={Qianli Wang and Van Bach Nguyen and Yihong Liu and Fedor Splitt and Nils Feldhus and Christin Seifert and Hinrich Schütze and Sebastian Möller and Vera Schmitt},
      year={2026},
      eprint={2601.00263},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2601.00263}, 
}