Table 3.

IR Results on epic-qa-dev. Ranking quality and recall oriented systems are evaluated with and without neural reranking, for both passage and document retrieval strategies. RR column (2nd) indicates if reranking is used or not and when used the value of $k$.

(a) IR Results on epic-qa-dev for passage retrieval using neural reranking over 5000 element ranking. Recall oriented (PRF $fbt=125$$fbd=30$) vs. ranking quality (NDCG) oriented (PRF $fbt=125$, $fbd=10$) systems.
PRF optimized for	RR	NDCG	R@500	R@1K	R@2K	R@3K	R@4K	R@5K
NDCG	no	0.3993	0.3991	0.5169	0.6252	0.6796	0.728	0.7599
R	no	0.3855	0.4002	0.5059	0.6118	0.6842	0.7253	0.7665
NDCG	0.9	0.4157	0.4612	0.5787	0.6724	0.7238	0.7604	0.76
R	0.9	0.403	0.456	0.5665	0.6705	0.7195	0.7555	0.76288

(b) IR Results on epic-qa-dev for document retrieval using neural reranking over 5000 element ranking. Recall oriented (PRF $fbt=30$, $fbd=40$) vs. ranking quality (NDCG) oriented (PRF $fbt=10$, $fbd=40$) systems.
PRF optimized for	RR	NDCG	R@500	R@1K	R@2K	R@3K	R@4K	R@5K
NDCG	no	0.5781	0.6731	0.761	0.8214	0.8363	0.8487	0.8652
R	no	0.5596	0.6651	0.7675	0.8255	0.8446	0.8598	0.8712
NDCG	0.1	0.5807	0.7041	0.7867	0.8342	0.8538	0.8655	0.8686
R	0.2	0.5691	0.7041	0.7661	0.8271	0.8538	0.8655	0.8709