Rather than using the BM25 ranking algorithm, for each abstract generate word embeddings and then select the xth closest to query `q`. Dependency: - https://github.com/hanxiao/bert-as-service/