Our dataset follows previous works. For long texts, we follow Conwea. For short texts, we follow LOTClass.
We transform all their data into unified json format.

1. Download datasets from: https://drive.google.com/drive/folders/1D8E9T-vuBE-YdAd9OBy-yS4UW4AptA58?usp=sharing

   • Long text datasets(follow Conwea):

     • 20Newsgroup Fine(20NF)
     • 20Newsgroup Coarse(20NC)
     • NYT Fine(NYT_25)
     • NYT Coarse(NYT_5)

   • Short text datasets(follow LOTClass)

     • Agnews
     • dbpedia
     • imdb
     • amazon

2. Unzip data into './data/processed'

Another way to obtain data (Not recommended):
You can download long text data from Conwea and short text data from LOTClass and transform data into json format using our code. The code is located at 'preprocess_data/process_long.py (process_short.py) You need to edit the preprocess code to change the dataset path to your downloaded path and change the taskname. The processed data is located in 'data/processed'. We alse provide preprocess code for X-class, which is 'process_x_class.py'.

This project is based on python==3.8. The dependencies are as follow:

pytorch
DGL
yacs
visdom
transformers
scikit-learn
numpy
scipy

• Recommend to start visdom to show the results.

visdom -p 8888

Open the browser to the server_ip:8888 to show visdom panel.

• Train:

  • First edit 'task/pipeline.py' to specify to config file and CUDA devices you used.
    Some configuration files are provided in the config folder.

  • Start training:

    python task/pipeline.py
    

  • Our code is based on multi GPUs, may be unable to run on single GPU currently.

1. provide datasets to dir data/processed.

   • keywords.json
     keywords for each class. type: dict. key: class_index. value: list containing all keywords for this class. See provided datasets for details.

   • unlabeled.json
     unlabeled sentences in our paper. type: list. item: list with 2 items([sentence_i,label_i]).
     In order to facilitate the evaluation, we are similar to Conwea's settings, where labels of sentences are provided. The labels are only used for evaluation.

2. provide config to dir config. You can copy one of the existing config files and change some fields, like number_classes, classifier.type, data_dir_name etc.

3. Specify the config file name in pipeline.py and run the pipeline code.

Please cite the following paper if you find our code helpful! Thank you very much. Paper link: https://aclanthology.org/2021.emnlp-main.222/

@inproceedings{zhang-etal-2021-weakly, title = "Weakly-supervised Text Classification Based on Keyword Graph", author = "Zhang, Lu and Ding, Jiandong and Xu, Yi and Liu, Yingyao and Zhou, Shuigeng", booktitle = "Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing", month = nov, year = "2021", address = "Online and Punta Cana, Dominican Republic", publisher = "Association for Computational Linguistics", url = "https://aclanthology.org/2021.emnlp-main.222", pages = "2803--2813", abstract = "Weakly-supervised text classification has received much attention in recent years for it can alleviate the heavy burden of annotating massive data. Among them, keyword-driven methods are the mainstream where user-provided keywords are exploited to generate pseudo-labels for unlabeled texts. However, existing methods treat keywords independently, thus ignore the correlation among them, which should be useful if properly exploited. In this paper, we propose a novel framework called ClassKG to explore keyword-keyword correlation on keyword graph by GNN. Our framework is an iterative process. In each iteration, we first construct a keyword graph, so the task of assigning pseudo labels is transformed to annotating keyword subgraphs. To improve the annotation quality, we introduce a self-supervised task to pretrain a subgraph annotator, and then finetune it. With the pseudo labels generated by the subgraph annotator, we then train a text classifier to classify the unlabeled texts. Finally, we re-extract keywords from the classified texts. Extensive experiments on both long-text and short-text datasets show that our method substantially outperforms the existing ones.", }