Flair: Predict top n most likely entity classes depending on confidence score

[IsabelMeraner]

Is there a way to return the top n most likely entity labels for NER?
I want to increase the coverage of my NER models and to predict the most likely entity label if the model is not >0.99 confident that the token should be tagged as "O" (OUTSIDE).

une un  D_ind   O   O   (0.9999514818191528)
course  course  N_C O   O   (0.999929666519165)
au  à   P   O   O   (0.9998410940170288)
Moléson Moléson N_P B-LABEL O   (0.27142760157585144)

Thank you!

[alanakbik]

Hello @IsabelMeraner this is currently not possible in Flair, but we recently added support for exactly this to the TextClassifier. This is a good feature to have, so I think we will add this to the SequenceTagger as well, hopefully soon.

[IsabelMeraner]

Hello @IsabelMeraner this is currently not possible in Flair, but we recently added support for exactly this to the TextClassifier. This is a good feature to have, so I think we will add this to the SequenceTagger as well, hopefully soon.

Hello @alanakbik
Thank you so much for your quick reply. Perfect, that sounds great.

[jantrienes]

This would be a nice feature. I have a similar use-case where we want to analyze the probability distribution for all NER tags for each token. Currently, Token.tag only returns the most likely tag (i.e., the one with highest score).

[alanakbik]

@jantrienes sounds interesting, we'll try to add this feature soon!

[Philipduerholt]

Hi,

I created a draft PR regarding this feature: https://github.com/zalandoresearch/flair/pull/782.
The PR builds upon https://github.com/zalandoresearch/flair/pull/642.

I'll be happy if you give it a look @alanakbik @jantrienes.

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.