Minimally Supervised Structure Rich Text Categorization by Learning on Text-Rich Networks

Paper

Our paper can be accessed here.

Running the Experiments

Requirements

The code requires Python 3.7+ and the HuggingFace Transformers library transformers==4.1.0. The detailed requirements can be found in requirements.txt. Note that specific versions of torch_scatter, torch_sparse, torch might be needed to work with different Cuda versions.

Steps to Run the Experiments

Download data

The data can be accessed through Dropbox.

Run the training script

Edit the training scripts run_amazon.sh and run_books.sh to specify path to data and the output. Then execute the scripts to run the experiments.

Running on custom datasets

Please follow the given datasets to format your data. Then create a training script to run the experiments.

Citation

Please cite the following paper if you found our dataset or framework useful. Thanks!

@inproceedings{zhang2021ltrn,
  author = {Zhang, Xinyang and Zhang, Chenwei and Dong, Luna Xin and Shang, Jingbo and Han, Jiawei},
  title = {Minimally Supervised Structure Rich Text Categorization by Learning on Text-Rich Networks},
  year = {2021},
  booktitle = {Proceedings of The Web Conference 2021},
  location = {Ljubljana, Slovenia},
  series = {WWW '21}
}