This repository contains codes for our papers "文表現の摂動正規化: 事前学習済みモデルのDebias手法" in ANLP2021 (言語処理学会第27回年次大会).
The codes is to fine-tune BERT for debias and to visualize bias metrics on pre-trained language models.

This repository has some experiments.

• Sentence Pertubation Normalizer (will be published at ANLP2021)
• Bias Subspace Transformation of the word embedding in BERT Embeddings
• etc...

Related with ANLP

• Analyze_GLUE.ipynb
• Analyze_crows_pairs.ipynb

Others

• Model_Analyze.ipynb

The environment of this repository is builded by poetry.
If you execute my experiments, you put appropriate data into /data, and you should run tasks in /tasks.
Please ask me at Issue if you have anything you don't know.