Code for "We Are What We Repeatedly Do: Inducing and Deploying Habitual Schemas in Persona-Based Responses"

Benjamin Kane and Lenhart Schubert. We Are What We Repeatedly Do: Inducing and Deploying Habitual Schemas in Persona-Based Responses, EMNLP 2023 (to appear).

Dependencies

See requirements.txt. A valid openai API key is also required in _keys/openai.txt.

Data

The data directory contains our expansion of the PersonaChat test set containing schemas and responses generated by each of the methods described in our paper.

How to run

cd src

Schema generation

To estimate costs:

python3 generate_schemas.py --dataset "personachat" --percent 100

To generate:

python3 generate_schemas.py --dataset "personachat" --percent 100 --output-filename "../data/personachat-schemas.json" --checkpoint-iter 50 --generate

Response generation

To estimate costs:

python3 generate_response.py --dataset "../data/personachat-schemas.json" --percent 100

To generate:

python3 generate_response.py --dataset "../data/personachat-schemas.json" --percent 100 --output-filename "../data/personachat-schemas-responses.json" --checkpoint-iter 50 --generate

Citation

If you find our response generation methods useful for your research, please cite our paper:

@inproceedings{Kane2023Habitual,
  author    = {Benjamin Kane and Lenhart Schubert},
  title     = {We Are What We Repeatedly Do: Inducing and Deploying Habitual Schemas in Persona-Based Responses},
  booktitle = {EMNLP},
  year      = {2023},
  url       = {https://arxiv.org/abs/2310.06245v1},
}