CONSUMER HEALTH QUESTION ANSWERING WITH LLM-BASED SIMPLIFICATION AND SUMMARIZATION

Abstract

Considering the complicated nature of available healthcare data, there’s a huge necessity for rendering this information more comprehensible to all the consumers. Large Language Models (LLMs) can be used to answer consumers’ questions in more simple and concise manner. This thesis explores the influence of such LLMs such as ChatGPT and Gemini in refining consumer health question answering through means of summarization and simplification of scientific abstracts from authoritative resources such as PubMed, and evaluation of these pipelines through metrics as BERTScore, ROUGE and SARI scores respectively. The main objective of this study is evaluation of results of retrieval, summarization on BioASQ data’ subset and simplification on PLABA dataset, and comparison of used LLMs on metrics mentioned above. Through iterative experiments, it was identified that choice of prompt and LLM greatly impacts the final result of simplification and summarization of the healthcare information.