Il 27/05/15 15:29, David Przybilla ha scritto:


    - What is the "confidence"? I thought that the value shown with the
    "n-best candidates" was the confidence, but I obtain this weird
    result,
    as follows.
    1.
       - Leave the text of the demo.
       - Leave the default confidence (0.5).
       - Check "n-best candidates".
       - Annotate
    You will see that the first word, "First" is not linked.
    2.
       - Leave the text of the demo.
       - Set the confidence to 0.1.
       - Check "n-best candidates".
       - Annotate
    You will see that the first word, "First", now is linked to WWI with
    confidence 1.
    Is it normal?



The linking takes place in two stages : spotting & disambiguating.
Spotting matches text to surface forms. Disambiguating chooses one topic among potential candidates (n-best candidates). As it currently stands the confidence parameter refers both to the spotter and the disambiguator. So that parameter is used to prune potential spots as well as potential topics. My guess is that by lowering the confidence you allowed the spotter to get an extra surface form match ( "First").
Dear David,
sorry for my late answer. I was strongly sure to have answered to this e-mail (and I was waiting for the answer, poor me!), but indeed I had not.

Anyway...

I understand that by lowering the confidence I obtain more results.
The problem is that the results are given with confidence 1. This is weird, because if "First" appears only when the confidence input value is really low, I expect that che confidence return value would be low too.

I'll tell what I really need, so that maybe you can help me.
I'd like to launch the API without confidence (or confidence = -1, like in some examples of the documentation), then filter a posteriori the various thresholds. Unfortunately, setting confidence = -1 will result in everything tagged (that is ok), but with really high confidence values, so that I cannot filter which tags are good and which are not.

I hope that now it's clear what I mean.

Thank you.
Best,
Alessio
------------------------------------------------------------------------------
_______________________________________________
Dbp-spotlight-users mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/dbp-spotlight-users

Reply via email to