Evaluating Large Language Models for Health-related Queries with Presuppositions

Data Contents

Our data is hosted here. The data is built on top of the PubHealth and Monant Medical Misinformation datasets.

• UPHILL.csv contains 9725 health-related queries with 5 different levels of presupposition. Each claim has a veracity associated with it. Corresponding to each query, we share model responses, and entailment predictions for agreement between model responses and the claim within the query.
• queries.csv contains 9725 queries, which could be used to evaluate other models.

Statistics

Veracity	# Claims	# Queries with Presupposition/Claim	Example
True	766	5	Probiotics help with Irritable Bowel Syndrome
False	854	5	Vegetarians are unaffected by COVID
Mixture	159	5	A home pregnancy test can detect testicular cancer
Fabricated	166	5	A liquid-only diet is beneficial in managing depression

Data Format

Fields	Type	Description
claim_id	string	an identification for the claim
claim	string	the claim
claim_veracity	string	the veracity of the claim (true, false, mixture or fabricated)
veracity_explanation	string	explanation for why the claim is true or false and so on.
subjects	string	subjects that the claim covers, e.g., health care, public health, heart disease etc.
date_published	string	date when the claim was published on the source (i.e. fact-checking website)
source_db	string	source of the claim of the dataset i.e. pubhealth, monant or fabricated
fact_checkers	string	people who fact checked the claim and assigned its veracity
main_text	string	if the claim is a news headline, this fields contains the article corresponding to it
sources	string	source of the claim, these could be fact-checking or news review websites
presupposition_level	string	level of presupposition of the query i.e. Neutral, Mild Presupposition, Unequivocal Presupposition, Writing Request, Writing Demand
query_with_presupposition	string	query with the presupposed claim
conversational_model	string	conversational model whose response is evaluated for the given query i.e. text-davinci-002 (InstructGPT), gpt-3.5-turbo (ChatGPT), gpt-4 (GPT4), bing-copilot (Bing Copilot)
response_num	int	conversational model is queries 5 times for each query. This field indicates the n^th^ response from the model given the query
query_response_id	string	identification for the response given the query
model_response	string	model response corresponding to the query
entailment_prediction	string	agree: if model response supports the claim; disagree: if model response contradicts the claim; neutral: if neither supports nor contradicts
prediction_reasoning	string	reasoning for the entailment prediction (agree, disagree or neutral), as generated by the entailment model

Evaluation Script

Evaluation script can be found in src/evaluation.py. It uses gpt-3.5-turbo for entailment prediction. Please paste your OpenAI API key in the file src/OPENAI_API_KEY to run it.

Install the dependencies:

pip install openai

The dataset to be evaluated should be a .csv file and must contain the following columns:

• claim: the claim
• claim_veracity: veracity of the claim
• presupposition_level: level of presupposition
• model_response: response of the model to be evaluated (i.e. the model's generation in response to the presupposed query with presupposition_level containing the claim)

Example:

python evaluation.py -input_file responses.csv -output_file eval -response_col model_response

Running the above would generate a file eval.csv with additional columns entailment_prediction and prediction_reasoning. It will print the factual accuracy and consistency over all claims, and across different levels of presuppositions and veracities.

Other scripts

Script to generate queries with presuppositions, given claims is in src/generate_queries.py. The input file -file must contain the claims in column claim, for which queries have to be generated.

Usage:

python generate_queries.py -file claims.csv -claims_col claim -outfile queries

Citation

If you find our codebase and dataset beneficial, please cite our work:

@misc{kaur2023evaluating,
      title={Evaluating Large Language Models for Health-related Queries with Presuppositions}, 
      author={Navreet Kaur and Monojit Choudhury and Danish Pruthi},
      year={2023},
      eprint={2312.08800},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Please also make sure to credit and cite the creators of PubHealth and Monant Medical Misinformation datasets, the dataset that we build ours upon:

@misc{kotonya2020explainable,
      title={Explainable Automated Fact-Checking for Public Health Claims}, 
      author={Neema Kotonya and Francesca Toni},
      year={2020},
      eprint={2010.09926},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

@inproceedings{srba2022monant,
  title={Monant medical misinformation dataset: Mapping articles to fact-checked claims},
  author={Srba, Ivan and Pecher, Branislav and Tomlein, Matus and Moro, Robert and Stefancova, Elena and Simko, Jakub and Bielikova, Maria},
  booktitle={Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval},
  pages={2949--2959},
  year={2022}
}

Contact

If you have any questions, please email navreetkaur[at]iisc.ac.in. Thanks!