VOH.CoLAB at TREC 2020 Precision Medicine Track.

This paper describes our participation in the Scientific Abstracts task of the TREC 2020 Precision Medicine Track. We present our approach and the methods implemented, including both submitted runs and several post-mortem experiments using different methods. We performed experiments with Drugbank-based synonym expansion, Rocchio-based pseudo-relevance feedback, and neural re-ranking using the BioBERT biomedical pre-trained language models. In our evaluation, the Rocchio-based pseudo-relevance feedback method was the best performing method. Finally, we found that metadata and other textual fields in the document (e.g., journal name), are useful for retrieval and, when indexed, can improve recall-oriented metrics considerably leading to improvements in retrieval performance across the board.