DGoT: Dynamic Graph of Thoughts for Scientific Abstract Generation

This is the implementation of our LREC-COLING 2024 paper DGoT: Dynamic Graph of Thoughts for Scientific Abstract Generation.

Quick Start

Environment Setup

We strongly recommend that you use Docker images dgot_demo to run our programs.

# pull docker image
docker pull jaycening/dgot_demo:v1.0.0

# Clone this repository and mount it into the image container.
git clone https://github.com/JayceNing/DGoT.git
docker run --gpus all -it -d --privileged=true -v ./DGoT:/home/nxy/LLM/DGoT

# Enter the target folder.
docker exec -it dgot /bin/bash
cd /home/nxy/LLM/DGoT

The image is preconfigured with InternLM2 deployed under LMDeploy version 0.2.4.

If you want to configure the environment manually, see the documentation environment_setup.md

Data Preparation for PubMedCite Dataset

The citation graph records of PubMedCite come from the CitationSum repository. Here, we provide code for downloading datasets based on the official PubMed API.

python get_data.py --required_num 100

• required_num is the number of data entries required for the training and testing datasets to be downloaded.

Activate LLM's API service

Taking InternLM2 as an example.

cd /home/nxy/internlm2_chat_deploy
lmdeploy serve api_server ./workspace --cache-max-entry-count 0.2

Training Process

cd /home/nxy/LLM/DGoT
python generate_abstract.py --begin 0 --end 1 --mode "train" --model "internlm2" --task "default"

• begin and end indicate the starting and ending indices of the dataset being used.
• mode indicates whether the dataset being used is train dataset or test dataset.
• model represents the LLM used for inference.
• task represents the type of task being performed.

Reasoning Process

python generate_abstract.py --begin 0 --end 100 --mode "test" --model "internlm2" --task "default" --thresh_g 0.34 --thresh_a 0.35 --thresh_i 0.34

• thresh_g, thresh_a and thresh_i respectively represent the thresholds used for generating transformation, aggregating transformation, and boosting transformation in DGoT.

Tutorials

Here, we provide detailed tutorials on other aspects.

• Environment Setup Manually
• Appendix Experiment Procedure

Acknowledgement

This work is based on the following prompt framework, large language model, and model deployment toolkit. Thanks for the open source contribution!

• Graph of Thoughts (GoT)
• ChatGLM2-6B
• InternLM
• LMDeploy

This paper's computational resources are supported by the High-performance Computing Platform of BUPT.

Citations

If you find this repository valuable, please give it a star!

Using this in your work? Please reference us using the provided citation:

@misc{ning2024dgot,
      title={DGoT: Dynamic Graph of Thoughts for Scientific Abstract Generation}, 
      author={Xinyu Ning and Yutong Zhao and Yitong Liu and Hongwen Yang},
      year={2024},
      eprint={2403.17491},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Contributor

Jayce Ning

Home Page：https://jaycening.github.io/zh-cn/

Github：https://github.com/JayceNing

ZhiHu：https://www.zhihu.com/people/XinyuNing