docTransformer

Transformer for classification tasks that operates with document fragments

Features

• UPCOMING: Multi-GPU support using the Accelerate library
• Explainability measures
  • Currently output salient words and phrases in the classification using Layer Integrated Gradients attribution scores with the Captum library
• Gradient accumulation to train with larger effective batches using the Accelerate library
• An input data streaming implementation to allow training with large datasets
• Possibility to build document embeddings before classification both during training and classification
  • Document embedding built as the average of the [CLS] token of n parts of the document:
    • the document is divided in n parts with an equal number of characters. No sentence/fragment information is used
    • the document is divided in n parts with an approximate equal number of sentences. If the sentence_batch is larger than n, sentences are averaged individually
• Several possibilities to be considered as input to the classifier besides the standard [CLS] token
  • [CLS] + tanh
  • [CLS]
  • average pooling + tanh
  • average pooling

Requirements

• Python version >= 3.9
• PyTorch version >= 2.0.1
• Accelerate version >= 0.21.0
• Captum version 0.6.0

Example Usage

Slurm

Training

srun --ntasks 1 --gpus-per-task 1 python -u docClassifier.py --gradient_accumulation_steps 2

srun -p A100-80GB -t 3-0 --ntasks 1 --gpus-per-task 1 python -u docClassifier.py --eval_steps 10000 --lr 5e-6 -f modelb2a8sentence16A80 -o modelb2a8sentence16A80seed2.bin -b2 -a8 --sentence_batch_size 16 --split_documents True --seed 5678

Evaluation

srun -p RTXA6000 --ntasks 1 --gpus-per-task 1 python -u docClassifier.py --task evaluation -f modelb2a8sentence2V100 -o modelb2a8sentence2V100seed3_333.bin -b2 --sentence_batch_size 2 --split_documents True --test_dataset data/multivariant3all.test --plotConfusionFileName modelSplit2Seed3test.png

Classification

srun --ntasks 1 --gpus-per-task 1 python -u docClassifier.py --task classification -f modelb2a8sentence2V100 -o modelb2a8sentence2V100seed3_333.bin -b1 --sentence_batch_size 2 --split_documents True --test_dataset ../es/es_meta_part_1.jsonl.unk

Explanation

srun -p RTXA6000 --ntasks 1 --gpus-per-task 1 python -u docClassifier.py --task explanation -t data/testExample.mx -f modelb2a8fixV100 -o modelb2a8fixV100seed3_3.bin -b1 --split_documents False --xai_threshold_percentile 90

Further Usage

In order to use it for your own classification task with full functionalities:

• Prepare your training data with the class in the 2nd column and the text to classify in the 5th column. Otherwise, modify the line_mapper function in data.py
• Adapt the classification labels in the last functions of data.py

Citation

Please, use the following bibtex entries when citing this research work

@InProceedings{espana-bonet-barron-cedeno-2024,
    title = "Elote, Choclo and Mazorca: on the Varieties of Spanish",
    author = "Espa{\~n}a-Bonet, Cristina  and  Barr{\'o}n-Cede{\~n}o, Alberto",
    booktitle = "Proceedings of the 2024 Annual Conference of the North American Chapter of the Association for Computational Linguistics",
    month = jun,
    year = "2024",
    address = "Mexico City, Mexico",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/",
    pages = "--"
}

Version v1.0.1 without the document level functionality (--split_documents False) has been used in

@InProceedings{espana-bonet-2023-multilingual,
    title = "Multilingual Coarse Political Stance Classification of Media. The Editorial Line of a ChatGPT and Bard Newspaper",
    author = "Espa{\~n}a-Bonet, Cristina",
    editor = "Bouamor, Houda  and
      Pino, Juan  and
      Bali, Kalika",
    booktitle = "Findings of the 2023 Conference on Empirical Methods in Natural Language Processing",
    month = dec,
    year = "2023",
    address = "Singapore",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.findings-emnlp.787",
    doi = "10.18653/v1/2023.findings-emnlp.787",
    pages = "11757--11777"
}