Awesome

UnitPersonaBias

This is the repository to accompany the paper Revealing Persona Biases in Dialogue Systems.

@article{sheng2021revealing,
  title={Revealing Persona Biases in Dialogue Systems},
  author={Sheng, Emily and Arnold, Josh and Yu, Zhou and Chang, Kai-Wei and Peng, Nanyun},
  journal={arXiv preprint arXiv:2104.08728},
  year={2021}
}

For questions, please contact Emily at ewsheng@gmail.

Setup

conda create --name personabias python==3.7
conda activate personabias
pip install -r requirements.txt

Offensiveness Test Setup

To run the offensiveness bias test on the Blended Skill Talk dataset, you'll have to download the dataset here (in bst.pkl format).

To run this test on the RealToxicityPrompts dataset, you'll have to download the dataset here.

DialoGPT Setup

The persona-DialoGPT model we describe in the original paper (DialoGPT-medium fine-tuned on PersonaChat) can be found here.

If you wish to fine-tune your own model, you can follow the instructions below.

Running Persona Bias Test Cases

Offensiveness Biases

If we want to run the offensiveness test cases on Blender with personas and prompts from RealToxicityPrompts:

python offensiveness.py \
--model_type blender \
--include_persona \
--generator rtp

See python offensiveness.py -h for more options.

Harmful Affirmation/Refutation Biases

If we want to run the harmful affirmation/refutation test cases on Blender with personas:

python harmful_affirm_refute.py \
--model_type blender \
--include_persona

See python harmful_affirm_refute.py -h for more options.

Occupational Association Biases

If we want to run the harmful affirmation/refutation test cases on Blender with personas:

python occupational_assoc.py \
--model_type dialogpt \
--model_path <PATH-TO-FINETUNED-DIALOGPT> \
--include_persona

See python occupational_assoc.py -h for more options.

Gendered Coreference Biases

If we want to run the gendered coreference test cases on Blender without personas:

python gendered_coref.py \
--model_type blender

See python gendered_coref.py -h for more options.

Fine-tuning DialoGPT Personas

To fine-tune DialoGPT on the PersonaChat dataset, we must download PersonaChat by running the following:

parlai display_data --task personachat --datatype train
parlai display_data --task personachat --datatype valid

This will download train_self_original.txt and valid_self_original.txt files and print out the path to these files.

Next, run run_language_modeling.py to start the fine-tuning process. For example:

python run_language_modeling.py \
--output_dir models/dialogpt_persona \
--model_name_or_path microsoft/DialoGPT-medium \
--config_name microsoft/DialoGPT-medium \
--tokenizer_name microsoft/DialoGPT-medium \
--train_data_file [path-to-train_self_original.txt] \
--eval_data_file [path-to-valid_self_original.txt] \
--do_eval \
--line_by_line \
--per_device_train_batch_size 4 \
--per_device_eval_batch_size 4 \
--num_train_epochs 5 \
--learning_rate 5e-6 \
--do_train \
--block_size 128 

This fine-tunes DialoGPT on the PersonaChat data by concatenating all personas and different dialogue turns in the following format:

your persona: [personaA] your persona: [personaB] your persona: [personaC] your persona: [personaD] [EOS] [PERSON1 TEXT] [EOS] [PERSON2 TEXT] [EOS]