French CharLM and NER tagger models

[mhham]

I have trained forward and backward french CharLMs based on a french wikipedia dump, as well as a french NER tagger using the CharLMs and french Fasttext embedding, on the WikiNER annotated corpus.

For the CharLMs I have reached the following validation perplexities:

• Forward : 3.15
• Backward : 3.17

For the NER tagger, I have reached the following performances :

Set	Precision	Recall
Validation	88.1%	87.9%
Test	87.9%	87.7%
spaCy (for comparison)	82.24%	81.61%

You can download them and include them in your framework :

• CharLMs backward : link
• CharLMs forward : link
• NER tagger : link

[alanakbik]

@mhham wow great results!

Thanks very much for making these models available - we'll include them in the next release!

[alanakbik]

@mhham which WikiNER corpus did you use? In the WikiNER Github, there is a wp2 and a wp3 version for French (and all other languages).

[mhham]

I used the aij-wikiner-fr-wp3 one from the WikiNER Github you mentioned

[alanakbik]

Ok, thanks!

[aschmu]

Hi, just a quick question, do we know what parameters were used to train the final model? I'm trying to reproduce the metrics using the wikiner dataset.

[ghost]

Good morning ,
i am having a strange error will loading flair french NER models.
it seems that the download link is no longer valide.

can you help please me ?

[alanakbik]

@sylvanusmahe - see #1831

[BouriMohammed]

Hello, Please how I can extract the the wp3 file ?