QuestGen

Automated fact-checking pipelines decompose a claim into questions, retrieve evidence for each, and predict veracity from the answers. The assumption going in was that a large language model would generate the best questions, but it did not.

I wrote experiments in Python with PyTorch, Transformers and LangChain, and ran experiements across five aggregated datasets including a custom synthetic one generated with GPT-3.5 (SoTA at the time). The work was published to ACM CIKM 2024 and I presented a poster at the conference.

Setty, Ritvik and Vinay Setty. "QuestGen: Effectiveness of Question Generation Methods for Fact-Checking Applications." CIKM 2024.