PaperLM

Source code for the CIKM 2023 paper "PaperLM: A Pre-trained Model for Hierarchical Examination Paper Representation Learning"

Environment

Requirments

numpy
pandas
torch==1.11.0
transformers==4.26.0
edunlp==0.0.8
tqdm
sklearn
scipy

Usage

• Data preprocessing

  • Convert paper text into vector using pre-trained BERT
  • Build knowledge table

  cd src
  # train
  python data_preprocess.py

• Pre-train

  cd src
  # train
  python main.py --mode pretrain

• Test

  cd src
  
  # paper difficulty estimation
  python main.py --mode finetune --downstream_task diff
  
  # examination paper retrieval
  python main.py --mode finetune --downstream_task similarity
  
  # paper clustering
  python main.py --mode finetune --downstream_task cluster

For more running arguments, please refer to [src/utils.py].