Table 5.

Performance of the individual masked language models and their combination using reciprocal rank fusion (RRF).

Model	NDCG@20a	P@20b	Bprefc	MAPd	# rele
BERTf	0.6209	0.6430	0.5588	0.2897	6879
RoBERTag	0.6261	0.6440	0.5530	0.2946	6945
XLNet	0.6436	0.6570	0.5644	0.3064	6926
mlm + rrf	0.7716	0.7880	0.5680	0.3468	6963

^aNDCG@20: normalized discounted cumulative gain at 20 documents.

^bP@20: precision at 20 documents.

^cBpref: binary preference.

^dMAP: mean average precision.

^e# ref: total number of relevant documents retrieved by the model for the 50 queries.

^fBERT: Bidirectional Encoder Representations from Transformers.

^gRoBERTa: robustly optimized BERT approach.