add Jelinek-Mercer (Dirichlet) QL scoring

[tteofili]

As discussed in #465 I've re introduced the option to use Jelinek Mercer QL scoring (qljm) and added the redundant QL Dirichlet option (qld) rebased against master.

[lintool]

Thanks! lgtm, I'll merge after CI turns green.

BTW, how does JM compare against D smoothing?
Anything surprising from what the Zhai et al. paper concludes?

[tteofili]

Anything surprising from what the Zhai et al. paper concludes?

my experiments suggest 0.4 / 0.5 is best even for title queries to obtain best NDCG.

[lintool]

How does JM compare against D smoothing? Better or worse?

[tteofili]

How does JM compare against D smoothing? Better or worse?

here's a quick round of experiments on NYT (on average JM is worse, they're close enough with axiom reranking enabled):

model	reranking	ndcg	map
qld		0.3643	0.1585
qld	rm3	0.4082	0.1935
qld	ax	0.4120	0.2017
qljm		0.3179	0.1285
qljm	rm3	0.3867	0.1811
qljm	ax	0.4017	0.2019

[lintool]

Ah nice. Seems like Anserini is making such experiments easy to run! :)

So it doesn't seem like it's worth adding regressions for QL JM then.

[tteofili]

Ah nice. Seems like Anserini is making such experiments easy to run! :)

yeah :)

So it doesn't seem like it's worth adding regressions for QL JM then.

agreed.