Tantivy NDCG Benchmarking Information Retrieval (BEIR) about tantivy

[fulmicoton]

I don't have time to debug this thing. One thing you can do is pick one specific example where tantivy is outperformed by BM25, and use "explain".

The usual suspects are

• your evaluation code
• tokenization
• the way query are sanitized
• BM25 constants

Tantivy has an explain function telling precisely how tantivy came up with a given score.
(The formula should be the same a lucene.)

https://docs.rs/tantivy/latest/tantivy/query/trait.Query.html#method.explain

[triandco]

Thank you @fulmicoton, I've gone through each of the usual suspects and verify each. I also ran the task against Lucene which yielded more similar scores. Look like this is working as intended.

Dataset	Tantivy ndcg@10	Apache Lucene ndcg@10	Beir BM25 Flat ndcg@10
Scifact	0.6251573122952132	0.632431156289918	0.679
NFCorpus	0.20505084876906404	0.20712280950112716	0.322
TREC-COVID	0.0362915780899568	0.035369826134136535	0.595
NQ	0.2637953053727399	0.2803606345656689	0.306

Regarding why Beir got such highscore, their "BM25" retrieval task is just a wrapper around ElasticSearch. I'm evaluating ElasticSearch now, will update the result soon.

[triandco]

Updated result with ElasticSearch evaluation and increase the retrieval task complexity from single field to multifield. The current result look reasonable as ElasticSearch default to do a bit more than BM25. I'll contact Beir about their specifics test since their result look a bit too pretty. I'll close this issue. Thank you @fulmicoton!

Dataset	Tantivy	Apache Lucene	Beir BM25 Flat	Elastic Search
Scifact	0.6110550406527024	0.6105774540257333	0.679	0.6563018879997284
NFCorpus	0.20174488628325865	0.2021653197430468	0.322	0.2116375800036891
TREC-COVID	0.03640657024103224	0.03705072222267741	0.595	0.05433894833185797
NQ	0.30181710921729077	0.301753090384626	0.306	0.310128528137924