caiy0220 / poe

Power of Explanations - PoE

Hi there, this repository contains the source code and necessary resources of the paper:
Power of Explanations: Towards automatic debiasing in hate speech detection
Yi Cai, Arthur Zimek, Gerhard Wunder, Eirini Ntoutsi (Preprint on arxiv)

Dependencies

Most of the experiments were carried out on Colab. The dependencies are therefore derived from the Colab's environment at the time the code was runned.

• Python: 3.7.13
• Pytorch: 1.12.1+cu113
• boto3: 1.24.68
• pyyaml: 6.0
• tqdm: 4.64.0

Dataset

The full Gab Hate Corpus can be downloaded via GHC under the ./Data folder, with the ghc_train.tsv containing the training set and ghc_test.tsv containing the test & validation set. A random split is performed on ghc_test.tsv for the preparation of the test & validation sets.

Files should be converted after the split into .jsonl format, namely train.jsonl, dev.jsonl, test.jsonl, and stored under the folder: ./data/majority_gab_dataset_25k/ (or any other path with the corresponding configuration updated).

Training

To train the model with MiD, simply run the following command, the other two supported modes are vanilla and soc, referring to the vanilla BERT and the baseline.

python launch.py --mode mid

Configurations on the whole framework can be found under the ./utils folder with the ending .yaml. Alter the parameters could also be done in the command line instead of editing the configuration file, e.g. change the random seed and the training mode:

python launch.py --seed 1 --mode vanilla

the same holds for other parameters in the 3 configuration files.

Detailed information during training will be stored as a list under the root of the workspace in a .pkl file, the list contains:

1. loss
2. regularized attribution penalty
3. debiasing lists over iterations
4. attribution and fpp of the corresponding token found by MiD over iterations in class AttrRecord
5. attribution and fpp of the corresponding token listed in baseline over iterations in class AttrRecord