gdeval and (py)trec_eval give different ndcg@k values

[seanmacavaney]

I believe the current configuration will never use gdeval by default for ndcg@k. But when the component is used, it produces different ndcg@k values than pytrec_eval. For instance, on a bm25 run on the COIVID dataset for ndcg@10, we get:

qid	gdeval	pytrec_eval
1	0.3842	0.4074
2	0.4826	0.4826
3	0.3749	0.3987
4	0.0366	0.0550
5	0.3832	0.4174
...
26	0.3963	0.4121
27	0.7602	0.7854
28	0.9148	0.9148
29	0.7467	0.7650
30	0.7305	0.7978
mean	0.5371	0.5618

This seems to be because gdeval uses a variant of ndcg that places higher importance on relevant documents.

This should be parameterized. Something like ndcg_exprel@k?