Reduce the weight of abstract size

[michelole]

https://www.ncbi.nlm.nih.gov/pubmed/15085719 is top-10 ranked for topic 1, but does not even mention the gene.

However, it's probably well ranked due to its small size. Experiment reducing this weight.

[kasac]

“as fart as possible” :-D I would rank it max as partially relevant, I agree with u. LG Z

…

On 29 Jun 2017, at 14:55, Michel Oleynik ***@***.***> wrote: https://www.ncbi.nlm.nih.gov/pubmed/15085719 <https://www.ncbi.nlm.nih.gov/pubmed/15085719> is top-10 ranked for topic 1, but does not even mention the gene. However, it's probably well ranked due to its small size. Experiment reducing this weight. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#71>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AD_54SaAs1jUs0rpYmd1aMbi79dFgeuOks5sI56sgaJpZM4OJQy6>.

[michelole]

We have two options:

1. Disable field-length norm.
2. Use BM25 ([Experiment] Rank with Okapi BM25 #86) as similarity measure and set b to zero.

I would go with the second.

Any option to reduce the weight implies reindexing.

[michelole]

How to set b

[michelole]

Setting B = 0 in the Okapi BM25 ranking algorithm worsened results from 0,7693 to 0,6852.

Also, the documents are longer and take more time to evaluate.