tjysdsg / 11830-hate-speech-detection-debiasing

11830 Group Project

CMU 11830 Group Project

Based on https://github.com/BrendanKennedy/contextualizing-hate-speech-models-with-explanations

The LM used by hierarchical explanation

SOC debiasing requires a pretrained LM. run_model.py will automatically train one if not found in runs/lm/. We provided a pre-trained LM.

General notes regarding training scripts

• Set --grandient_accumulation_steps if OOM.
• I removed Nvidia Apex cuz it's just not working well in 2023. I didn't bother to implement AMP using torch.amp as BERT-base is not that large.
• Remove the loop in training scripts if you don't need multiple runs with different seeds

Training

• scripts/fou_vanilla.sh trains a model without debiasing
• scripts/fou_soc.sh trains a model with SOC debiasing

Evaluation

Hate Speech Classification Performance

• Use scripts/test_toxigen.sh to get accuracy and F1 scores on Toxigen. It will also create a prediction file used for other tests below.

Measure biases on Founta

• Run scripts/test_bias_founta.sh to get runs/founta_*/founta_bias_eval.csv. The bias metrics are from https://arxiv.org/abs/2102.00086

Measure biases on Toxigen

• Run scripts/calc_target_group_fpr.sh to get FPR of each target group in Toxigen. The output is in runs/founta_*/toxigen_fpr.csv.

Get word weights of Toxigen samples using SOC explanation for all experiments

• scripts/explain_soc.sh will calculate the weight of every word in samples specified in data/toxigen_soc_line_numbers.csv