Explainable Bengali Multiclass News Classification

This repository contains the code for Explainable Bengali Multiclass News Classification — fine-tuning BanglaBERT (large) to classify Bengali news articles into 9 categories, and interpreting the model's predictions with Captum.

Setup

The dependencies are provided as a conda environment:

conda env create -f environment.yml
conda activate explainable-news

The code runs on both GPU and CPU (CUDA is used automatically when available).

Project structure

Note: data/ and saved_weights/ are not tracked in this repository (they are large). Generate the data with data_preprocess.ipynb and the weights with train.py.

Data pre-processing

Follow the steps in data_preprocess.ipynb. After pre-processing, the final data is saved in the data/ directory as even_cleaned_data.csv (a class-balanced version of the cleaned dataset).

Training & testing

Both training and testing use the same dataset (data/even_cleaned_data.csv) and the same stratified split (test_size=0.2, random_state=2023). Training uses the 80% partition and testing uses the disjoint 20% partition, so there is no train/test leakage.

To train the model:

python3 train.py

Trained weights are written to saved_weights/. Before testing, set the weight file you want to evaluate at line 38 of test.py:

saved_model_path = f'saved_weights/<weight file name>'

Then run:

python3 test.py

This prints the per-class precision/recall/F1 and overall accuracy on the held-out test set.

Explainability

Follow Explainability.ipynb. Load the saved weights the same way as in testing (cell 4 of the notebook), then run the Captum-based attribution cells to see which tokens drive each prediction.

References

1. Dataset: https://www.kaggle.com/datasets/furcifer/bangla-newspaper-dataset
2. BanglaBERT: https://huggingface.co/csebuetnlp/banglabert_large
3. Explainability library: https://captum.ai/

Citation

If you use this repository, please cite our paper:

DOI: 10.1109/ICCIT60459.2023.10441218

@inproceedings{sikder2023explainable,
    title={Explainable Bengali Multiclass News Classification},
    author={Sikder, Md Fahim and Ferdous, Md and Afroz, Shraboni and Podder, Uzzal and Fatema, Kaniz and Hossain, Mohammad Nahid and Hasan, Md Tahmid and Baowaly, Mrinal Kanti},
    booktitle={2023 26th International Conference on Computer and Information Technology (ICCIT)},
    pages={1--6},
    year={2023},
    organization={IEEE}
}

License

This project is licensed under the MIT License — see the LICENSE file for details.