Contextualized hate speech classification

Code for "Assessing the impact of contextual information in hate speech detection", Pérez, J. M., Luque, F., Zayat, D., Kondratzky, M., Moro, A., Serrati, P., Zajac, J., Miguel, P., Gravano, A. & Cotik, V. (2022).

Link to paper

Dataset

The full dataset can be found at the huggingface hub.

You can check also our demo of contextualized hate speech detection

Instructions

1. Get the dataset and put it under data/

2. Generate and preprocess dataset

python bin/create_dataset.py data/articles.json data/comments.json --train_path data/train.json --test_path data/test.json
python bin/preprocess_dataset.py

3. Train models

First, plain hate classifiers

# Train non-contextualized model
python bin/train_hate_classifier.py --context 'none' --output_path models/bert-non-contextualized-hate-speech-es/ --epochs 10
python bin/train_category_classifier.py --context 'none' --output_path models/bert-non-contextualized-hate-category-es/ --epochs 5

# Train contextualized model
python bin/train_hate_classifier.py --context 'title' --output_path models/bert-contextualized-hate-speech-es/ --epochs 10
python bin/train_category_classifier.py --context 'title' --output_path models/bert-contextualized-hate-category-es/ --epochs 5


# Train fully contextualized
# Check out notebooks/Hatespeech_Colab_TPU.ipynb
python bin/train_category_classifier.py --context 'title+body' --output_path models/bert-title+body-hate-speech-es/ --epochs 5 --batch_size 8 --eval_batch_size 8
python bin/train_hate_classifier.py --context 'title+body' --output_path models/bert-title+body-hate-speech-es/ --epochs 10

For more instructions, check TRAIN_EVALUATE.md

Finetuning

1. First, preprocess data

python bin/preprocess_finetune_data.py "/content/drive/Shareddrives/HateSpeech/data/hatespeech-data/" "/content/drive/MyDrive/data/finetune-news/finetune_data/" --num_workers 10

2. Run finetuning

python bin/xla_spawn.py --num_cores 8 bin/finetune_lm.py config/no_context_ft.json