Cross-Refine: Improving Natural Language Explanation Generation by Learning in Tandem

🍻2024-11-29: Accepted to COLING 2025

🗒️ arXiv link: https://arxiv.org/abs/2409.07123

⚓ Pipeline

Cross-Refine has the following pipeline as shown in the figure:

1. The generator outputs an initial explanation based on the given input (experiments/generator_explanation.py)
   1. The explanation will be judged by LLMs whether it needs to be improved further (experiments/generator_explanation_evaluation.py)
2. The critic takes the initial explanation and input into account and outputs feedback on the explanation generated by generator & suggested explanation (critic_feedback_explanation_generation.py)
3. The generator gets feedback and suggested explanation from critic and use them to refine the initial explanation (feedback_refinement.py)

⚙️Install the requirements

python -m pip install --upgrade pip
pip install -r requirements.txt

Example

Here is an example for an instance from ECQA dataset:

Citation

@misc{wang2024crossrefineimprovingnaturallanguage,
      title={Cross-Refine: Improving Natural Language Explanation Generation by Learning in Tandem}, 
      author={Qianli Wang and Tatiana Anikina and Nils Feldhus and Simon Ostermann and Sebastian Möller and Vera Schmitt},
      year={2024},
      eprint={2409.07123},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2409.07123}, 
}