Hybrid Re-ranking for Biomedical Information Retrieval at the TREC 2021 Clinical Trials Track
Journal
30th Text REtrieval Conference, TREC 2021 - Proceedings
Date Issued
2021-01-01
Author(s)
Abstract
This paper presents our methodology for the task of TREC 2021 Clinical Trials Track, which requires a system to retrieve and return the most relevant biomedical articles after giving queries. We propose a novel approach to biomedical information retrieval by leveraging the term-based and the embedding-based retrieval models with a re-ranking strategy. In our hybrid framework, all the documents will be indexed by using a term-based, efficient search engine. For the given query, a smaller set of candidate results are retrieved from the search engine. The ranking is determined not only by the term-based ranking score but also by the term relationships labeled by the Amazon Comprehend service1 for refinement. Then, the candidate results are further scored by using the embedding-based model. We represent the document and the query with bioBERT and compute the cosine similarity between a pair of the document embedding and the query embedding as their relevance score. The final score is a linear combination of the term-based and the embedding-based scores. Experimental results show that our hybrid re-ranking method improves both Precision@k and R-precision scores on the 2016 Clinical Decision Support Track and 2021 Clinical Trials Track dataset.
Type
conference paper