Re-ranking Biomedical Literature for Precision Medicine with Pre-trained Neural Models

We propose a biomedical literature retrieval approach that incorporates a domain-specific BERT model as an auxiliary re-ranker. Experiments on TREC Precision Medicine dataset show its effectiveness in improving retrieval performance by 6.2% in inferred NDCG and 6.8% in R-precision over the best-published results. The contribution of this study is to provide evidence of incorporating BERT in a biomedical literature retrieval system, which serves the overall goal to improve the information retrieval for precision medicine.