Negative relevance scores and -J
seanmacavaney opened this issue · 2 comments
I stumbled upon what appears to be a bug, in which documents with negative relevance scores are removed when using -J
.
It's simple to reproduce:
qrels
Q1 0 D1 -1
Q1 0 D2 1
run
Q1 0 D1 -1 2 1 run
Q1 0 D2 1 1 2 run
$ trec_eval qrels run.1 -m P.1
P_1 all 0.0000
$ trec_eval qrels run.1 -m P.1 -J
P_1 all 1.0000
# I would expect the above to be 0.0000
I would expect negative relevance scores to be considered judged. For instance, in TREC WebTrack, -2 indicates the assessor regarded the page as "Junk".
It seems this is caused by the docno_info[i].rel >= 0
condition here: https://github.com/usnistgov/trec_eval/blob/master/form_res_rels.c#L219. Based on this condition, is this actually the desired behavior?
Historically, a qrel of -1 indicated pooled but not judged. I started using -2 to mark spam, not realizing this quirk of -J. Since -J is documented as "here be dragons" I'm mixed on whether to fix this or just document it.
Sounds fair to me!