Politeness Stereotypes and Attack Vectors: Gender Stereotypes in Japanese and Korean Language Models
Setup your python environment via:
pip install -r requirements.txt
We used Python 3.9.12 for this project.
Run representational bias experiments via:
python ./src/maskScores.py
--lang LANG Language (ja/ko)
--template TEMP Template to use (SN: rep. bias template, SNL: modified location template)
--id ID ID string used to label experiment
[--batchsize N] BatchSize for GPU (omit for CPU)
Run allocational experiments via:
python ./src/hateDetectScores.py
--lang LANG Language (ja)
--template TEMP Template to use (HN: rep. bias template, HNS: attack/test template, HNT: training template, HMF: gender_only template, H: tweet_only)
--application APP Application (CB: Cyberbullying)
--id ID ID string used to label experiment
[--batchsize N] BatchSize for GPU (omit for CPU)
ja-toxicity-dataset.jsonneeds to be downloaded from Surge AI