This repo contains code for the paper, "Automatically Neutralizing Subjective Bias in Text".

Concretely this means algorithms for

• Identifying biased words in sentences.
• Neutralizing bias in sentences.

This repo was tested with python 3.7.7.

These commands will download data and a pretrained model before running inference.

$ cd src/
$ python3 -m venv venv
$ source venv/bin/activate
$ pip install -r requirements.txt
$ python
>> import nltk; nltk.download("punkt")
$ sh download_data_ckpt_and_run_inference.sh

You can also run sh integration_test.sh to further verify that everything is installed correctly and working as it should be.

Click this link to download (100MB, expands to 500MB).

If that link is broken, try this mirror: download link

Click this link to download a (Modular) model checkpoint. We used this command to train it.

harvest/: Code for making the dataset. It works by crawling and filtering Wikipedia for bias-driven edits.

src/: Code for training models and using trained models to run inference. The models implemented here are referred to as MODULAR and CONCURRENT in the paper.

Please see src/README.md for bias neutralization directions.

See harvest/README.md for making a new dataset (as opposed to downloading the one available above).

1. Get your data into the same format as the bias dataset. You can do this by making a tsv file with columns [id, src tokenized, tgt tokenized, src raw, tgt raw] and then adding POS tags, etc with this script.
2. Training the pretrained model on your custom dataset using either the train command used to build that model or follow the instructions in the src/ directory to write your own training commands.