Recommend scoring hits with BM25(k1=0.9,b=0.4).

[jpountz]

Currently different engines use different parameters for BM25, e.g. Tantivy and Lucene use (k1=1.2,b=0.75) while PISA uses (k1=0.9,b=0.4). Robertson et al. had initially suggested that 1.2/0.75 would make good defaults for BM25 but Trotman et al. later suggested that 0.9/0.4 would make better defaults and this seems to be the consensus nowadays.

The ranking function matters because it affects which hits may be skipped via dynamic pruninng, which in-turn affects search performance.

Closes #45

[jpountz]

I believe that PISA does not require changes though it would be nice to make the BM25 configuration more explicit in the query logic, what do you think @amallia? I could use some help making that change as I'm not too familiar with the PISA API.

It looks like Tantivy supports configuring the ranking function, but I'm not proficient in Rust and could use some help there too.

[fulmicoton]

The PR does not make the change for tantivy does it?

[jpountz]

It does not indeed. I would like to change it but I am not familiar with Rust and unsure how to do it. I could use some help.

[jpountz]

As a counterpoint, @rmuir pointed me to the DFR paper which shows that BM25 with k1=1.2/b=0.75 happens to closely match the parameter-free I(n)L2 model, giving further evidence that k1=1.2/b=0.75 are good defaults for BM25.

[jpountz]

Separately I checked more search engines and IR toolkits:

• PISA, Anserini, JASS, ATIRE use 0.9/0.4
• Terrier, Vespa, Lucene, Tantivy use 1.2/0.75

So there doesn't really seem to be a consensus actually. The point from the DFR paper that theory meets practice with 1.2/0.75 is quite convincing. Unless I find more evidence that 0.9/0.4 is more effective, I am considering switching PISA to 1.2/0.75 instead of switching Lucene and Tantivy to 0.9/0.4.