Seminar Detail

Abstract: In this talk, I will introduce our experience of applying Large Language Models (LLMs) to biomedicine at NCBI/NLM/NIH. I will first briefly introduce some basics of LLMs, including auto-regressive language modeling, scaling, alignment, few-shot learning, and chain-of-though reasoning. I will share a case study on biomedical question answering for better understanding of these concepts. Despite their great successes, LLMs are known to hallucinate confident-sounding but inaccurate content. In the second part, I will introduce two approaches that augment LLMs to reduce hallucinations in biomedicine, namely retrieval augmentation and tool augmentation. For the former, I will talk about our perspective on how LLMs will impact information seeking from biomedical literature. For the latter, I will present our GeneGPT work for teaching LLMs to use NCBI Web APIs. Finally, with the knowledge gained from the first two parts, I will share our application research, TrialGPT, for patient-to-trial matching with LLMs.

Bio: Dr. Qiao Jin is a postdoc researcher at the BioNLP group under NCBI/NLM/NIH, working with Dr. Zhiyong Lu. He received his M.D. degree from Tsinghua University in 2022. Dr. Jin works on the intersection of medicine and AI, with a focus on language modeling and information retrieval. He developed BioELMo, one of the first pre-trained language models in biomedicine. He also designed PubMedQA, a widely used biomedical benchmark for evaluating large language models. His work EBM-Net won the best clinical NLP paper awarded by the International Medical Informatics Association in 2021. In addition, Dr. Jin led the teams that won the first Biobank Disease AI Challenge and two TREC competitions. Recently, he is interested in applying LLMs to biomedicine and has published several pilot studies such as GeneGPT and TrialGPT. Dr. Jin serves as a program committee member for NeurIPS, ICML, ICLR, ACL, EMNLP, etc., and is the Associate Editor for the Journal of Medical Internet Research.