This is the codebase of the paper: ERA-CoT: Improving Chain-of-Thought through Entity Relationship Analysis.

[2024/05/17]🔥 We are happy to announce ERA-CoT has been accepted to ACL 2024 main conference!

[2024/03/28]🔥 We are releasing the version 1.0 for running on Llama2 model.

Update your environment for the required dependency.

pip install -r requirement.txt

Get your Llama2 weight on https://huggingface.co/meta-llama/Llama-2-7b-chat-hf/tree/main, set up on default directory: /root/llama-2-7b-chat-hf (You can change the dir on era-cot/config.py)

Running the code:

python main.py --dataset gsm8k \
--engine llama2-7b \
--temperature 0.3

• Upload your json version of dataset on era-cot/dataset.

• Setting for loading your dataset on utils.

if args.dataset.lower() == 'commonsenseqa':
      json_res = decoder.raw_decode(line)[0]
      choice = "Answer Choices:"
      for c in json_res["question"]["choices"]:
            choice += " ("
            choice += c["label"]
            choice += ") "
            choice += c["text"]
            q = json_res["question"]["stem"].strip() + " " + choice
            a = json_res["answerKey"]
            id = 'temp_{}'.format(idx)
      questions.append(q)
      answers.append(a)
      ids.append(id)

✨ERA-CoT

• Entity extraction on text and finds out all the explicit relationships mentioned in the text.
• Gradually infers the possible implicit relationships involved between entities based on the explicit relationships.
• Scores and filters the reliability of these implicit relationships.
• Finally answers questions based on these entity relationships.

If you found that ERA-CoT helps your work, please cite:

@article{liu2024era,
  title={ERA-CoT: Improving Chain-of-Thought through Entity Relationship Analysis},
  author={Liu, Yanming and Peng, Xinyue and Du, Tianyu and Yin, Jianwei and Liu, Weihao and Zhang, Xuhong},
  journal={arXiv preprint arXiv:2403.06932},
  year={2024}
}

For academic and non-commercial use only.The whole project is under the CC-BY-NC 4.0 license. See LICENSE for additional details.