Cross encoder reranking

A cross-encoder reranker reads the query and a candidate passage together and outputs the probability that the passage is relevant: Nogueira and Cho fed BERT the query and passage as one input, fine-tuned it on MS MARCO relevant and non-relevant passages, and reached the top of the MS MARCO passage leaderboard, 27% (relative MRR@10) above the previous state of the art. Because relevance is judged on the pair, a passage that is on-topic but does not contain what the query asks for can be outranked after retrieval; that answer-first passages benefit is an inference, not a result of the paper.