Created by Minqian Liu, Lizhao Liu and Junyi Cao from South China University of Technology.

This repository contains the official PyTorch-implementation of our Neurocomputing 2022 paper Co-attention network with label embedding for text classification. We release the full code for training the model and the preprocessed AAPD data. Preparing other datasets is easy and we omit them for code release.

We place datasets under the directory:

.data/my_custom_dataset/<DATASET_NAME>

We preprocess all data into form

{"context": "The text to classified.", "question": "class1, class2, ..., classN (All labels from the dataset)", "answer": "The labels for the context."}

We provide some preprocessing scripts (for AG News, DP Pedia, Yelp Full) under the preprocess folder. For other datasets, just slightly modify the scripts.

For example, to train the model on Schema Dataset on GPU 0:

python train.py --train_tasks schema --device 0 --embeddings YOUR_EMBEDDING_PATH

You can resume the model with stored checkpoints using --load <PATH_TO_CHECKPOINT> and --resume. Note that the checkpoints, configurations and training logs are saved in the directory specified in --save. (./results as default)

You can evaluate the model with EVALUATION_TYPE as validation or test:

python predict.py --evaluate EVALUATION_TYPE --path PATH_TO_CHECKPOINT_DIRECTORY --device 0 --tasks schema

The best checkpoint on validation set will be loaded. If you want to specify a checkpoint to evaluate, use --checkpoint_name in your command. For example:

python predict.py --evaluate EVALUATION_TYPE --path PATH_TO_CHECKPOINT_DIRECTORY --device 0 --checkpoint_name iteration_1000.pth --tasks schema

Our project is highly based on the DecaNLP projects. Please go and show some support!

If you find this code helpful for your research, please consider citing

@article{liu2022co,
  title={Co-attention network with label embedding for text classification},
  author={Liu, Minqian and Liu, Lizhao and Cao, Junyi and Du, Qing},
  journal={Neurocomputing},
  volume={471},
  pages={61--69},
  year={2022},
  publisher={Elsevier}
}