On the Effectiveness of Adversarial Robustness for Abuse Mitigation with Counterspeech

This work aims at developing counterspeech classifiers for mitigating online hate targeting at premium league footballers. It employs human-and-model-in-the-loop data collection over multiple rounds to test and improve model capabilities.

1. Premium League Footballers (PLF) Dataset: Authentic Counterspeech Collection

Collect abusive tweets and their replies

The scripts in the folder collection contain example codes we use to gather potential abusive tweets and their replies.

2. DynaCounter Dataset: Dynamic Adversarial Counterspeech Collection

Model training and evaluation

DynaCounter is collected over multiple iterations. The file can be downloaded HERE.

To train/evaluate a counterspeech classifier, run

python -m src.train.cs_classification \
    --report_to wandb \
    --model_name_or_path $MODEL_PATH \
    --train_file $TRAIN_FILE \
    --validation_file $VAL_FILE \
    --test_file $TEST_FILE \
    --run_name $RUN_NAME \
    --learning_rate 2e-5 \
    --num_train_epochs $NUM_EPOCHS \
    --do_train \
    --do_eval \
    --do_predict \
    --max_seq_length 256 \
    --per_device_train_batch_size 32 \
    --evaluation_strategy steps \
    --eval_steps 100 \
    --logging_steps 100 \
    --logging_first_step True \
    --evaluation_strategy steps \
    --load_best_model_at_end True \
    --metric_for_best_model 'f1' \
    --output_dir $OUTPUT_DIR

You can specify various training parameters when calling the script.

Evaluate models on a given dataset

python -m src.evaluation.cs_classification_inference \
    --model_name_or_path $MODEL_PATH \
    --test_file $TEST_FILE \
    --run_name $RUN_NAME \    
    --do_predict \
    --max_seq_length 256 \
    --pred_file $PRED_FILE \
    --output_dir $OUTPUT_DIR

3. Counterspeech generation

Model training and evaluation

To train/evaluate a counterspeech generator, run

python src/train/cs_generation.py \
    --report_to wandb \
    --model_name_or_path $MODEL_PATH \
    --train_file $TRAIN_FILE \
    --validation_file $VAL_FILE \
    --run_name $RUN_NAME \
    --learning_rate 2e-5 \
    --num_train_epochs $NUM_EPOCHS \
    --do_train \
    --per_device_train_batch_size 4 \
    --evaluation_strategy steps \
    --eval_steps 100 \
    --logging_steps 100 \
    --logging_first_step True \
    --save_strategy steps \
    --save_steps 2000 \
    --load_best_model_at_end True \
    --output_dir $OUTPUT_DIR

Evaluate models on a given dataset

python src/evaluation/cs_generation_test.py \
  --model_type $MODEL_TYPE \
  --model_name_or_path $MODEL_PATH \
  --test_file $TEST_FILE \
  --out_file_path $OUTPUT_FILE_PATH \
  --length 50 \
  --p 0.9 \
  --num_return_sequences 1  

Contact

For any questions or feedback, please contact Yi-Ling Chung

Citation

For more details on data analysis and experiments, please see our paper.

@inproceedings{chung-etal-2024-towards,
    title = "On the Effectiveness of Adversarial Robustness for Abuse Mitigation with Counterspeech",
    author = "Chung, Yi-Ling and Bright, Jonathan",
    booktitle = "Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies ",
    year = "2024",
    publisher = "Association for Computational Linguistics"
}