Department of Computational and Data Sciences
Department Seminar
Speaker : Kundan Krishna, PhD candidate in the Language Technologies Institute at CMU
Title : “GenAudit: Fixing Factual Errors in Language Model Outputs with Evidence.”
Date & Time : May 27, 2024, 02:00 PM
Venue : # 102, CDS Seminar Hall
ABSTRACT
LLMs can generate factually incorrect statements even when provided access to reference documents. Such errors can be dangerous in high-stakes applications (e.g., document-grounded QA for healthcare or finance). In this talk, I would present GenAudit — a tool intended to assist fact-checking LLM responses for document-grounded tasks. GenAudit suggests edits to the LLM response by revising or removing claims that are not supported by the reference document, and also presents evidence from the reference for facts that do appear to have support. To power GenAudit, we trained models on a scpecially created dataset with high-quality human annotations. Comprehensive evaluation by human raters shows that GenAudit can detect errors in 8 different LLM outputs when summarizing documents from diverse domains, outperforming GPT-4 while also being much cheaper. GenAudit is available for public use at https://genaudit.org
BIOGRAPHY
Kundan Krishna is a PhD candidate in the Language Technologies Institute at CMU, advised by Professor Zachary Lipton and Professor Jeffrey Bigham. His research focuses on mitigating safety issues in deployment of language models by improving their factual accuracy and reducing reliance on web-scale scraped data for their pretraining. He has also worked extensively on various aspects of text summarization, including improving robustness to noise, producing structured summaries, generating topic-focused summaries etc. Prior to CMU, he graduated with a Bachelor’s degree from IIT Kanpur, and worked as a research engineer at Adobe Research
Host Faculty: Dr. Danish Pruthi
ALL ARE WELCOME
