ECo-Concept

Official repository for "Enhancing the Comprehensibility of Text Explanations via Unsupervised Concept Discovery", which has been accepted by ACL 2025 Findings.

Chinese Blog

Dataset

You can use your own dataset by modifying the data_path in the corresponding scripts. Make sure your dataset includes train.json, val.json, and test.json. In every json file:

• the content identifies the content of the sample.
• the label identifies the corresponding label of the sample.

Code

Requirements

• python==3.9.19

• CUDA: 11.3

• Python Packages:

  pip install -r requirements.txt
  

Pretrained Models

You can download pretrained models (bert-base-uncased and roberta-base) and change paths (bert_path) in the corresponding scripts.

Run

Main Parameter Configuration:

• base_model_name: Name of the pretrained model used.
• task: Name of the target dataset.
• num_classes: Number of classes in the classification task.
• num_concepts: Number of concepts to extract.
• dist_weight: Weight for the distinctiveness loss.
• con_weight: Weight for the consistency loss.
• com_weight: Weight for the comprehensibility loss.
• vis_threshold: Threshold for concept visualization.
• model_path: Path to save the trained model.
• checkpoint_path: Path to load saved model checkpoints.
• prompt_path: Path for prompts and summaries results.
• simulation_path: Path to save simulation results.
• if_retrain: Whether to execute the Concept Comprehensibility Enhancement Stage.

How to Cite

@misc{sun2025enhancing,
  title={Enhancing the Comprehensibility of Text Explanations via Unsupervised Concept Discovery},
  author={Sun, Yifan and Wang, Danding and Sheng, Qiang and Cao, Juan and Li, Jintao},
  year={2025},
  eprint={2505.20293},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url = {https://arxiv.org/abs/2505.20293}
}