DoSSIER at TREC 2023 Clinical Trials Track

Abstract

This paper describes the experimental setup and results of the DoSSIER team’s participation in the Clinical Trials Track at TREC 2023. The primary objective of this track was to identify clinical trials for which patients meet the eligibility criteria. Our approach uses pipeline-based models, including large language models (LLMs) for query expansion and entity extraction techniques to augment both queries and documents. In our pipelines, we tested two different first-stage retrieval models, followed by a neural re-ranking framework that leverages topical relevance and eligibility criteria. We add to the pipeline a GPT-3.5-based question-answering post-processing step. Our findings demonstrate that the neural reranking and subsequent LLM post-processing notably enhanced performance. Future research will focus on a comprehensive assessment of the impact of query and document representation strategies on retrieval efficacy.