Converting Scores to Probability

[bkj]

Is there a natural way to convert LID confidence scores to probabilities, or do they not really have a probabalistic interpretation? I imagine this may vary depending on the classifier used -- I'm specifically interested in the tweetRecallNorm.mod.gz at the moment.

Thanks
Ben

[JosephPCampbell]

Hi Ben,

That’s a simple question with a messy answer. I don’t know if Bill or Charlie have already answered. If you could explain what you wish to do, that might help us help you. Thanks for using our stuff!

Joe

From: Ben Johnson [mailto:notifications@github.com]
Sent: Tuesday, September 27, 2016 12:29 PM
To: mitll/LLClass
Subject: [mitll/LLClass] Converting Scores to Probability (#2)

Is there a natural way to convert LID confidence scores to probabilities, or do they not really have a probabalistic interpretation? I imagine this may vary depending on the classifier used -- I'm specifically interested in the tweetRecallNorm.mod.gz at the moment.

Thanks
Ben

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub #2 , or mute the thread https://github.com/notifications/unsubscribe-auth/AG2Owl_HFWnp7fVbp5yAQVz0kiMJ2FT_ks5quURAgaJpZM4KH4NX . https://github.com/notifications/beacon/AG2OwoNUJHXqX8xOVxEYSzK5xK8Rgwl0ks5quURAgaJpZM4KH4NX.gif

[gvidaver]

Let me look into this and get back to you.

[swadey]

@gvidaver in the original text.scala code there was a mechanism for this. I can point you to it if needed.

[bkj]

Hi All -- any updates on this? If not I can look at implementing it myself (possibly just as a postprocessing step). I think Charlie had said that you had used Platt scaling for this in the past?