how to train a french NER based on stanford-nlp Conditional Random Fields model?

Question

I discovered the tools of stanford-NLP and found it really interesting. I\'m a french dataminer / datascientist, fond of text analysis and would love to use your tools, but

Answer 1

NB: I am not a developper of the Stanford tools, nor a NLP expert. Just a lambda user that also needed such informations at some point. Also note that part of the information given below are from the official FAQ: http://nlp.stanford.edu/software/crf-faq.shtml#a

Here are the steps I followed to train my own NER:

1. Install java8

2. Create a train/test sample. It must take the form of .tsv files with the following format:

     Venez    O
     découvrir    O
     lundi    DAY
     le    O
     nouvel    O
     espace    O
     de    O
     vente    O
     ODHOJS    ORGANISATION
   

   Depending on the original format of your text, you can create this sample with SQL statement or other NLP tools. The labelling is the most complicated part as I don't know other ways to proceed than to do it by hand.

3. Train the model with this command:

   java -cp "stanford-ner.jar:lib/*" -mx4g edu.stanford.nlp.ie.crf.CRFClassifier -prop prop.txt
   

   where prop.txt is also described here.

   This should create a new .jar containing the newly trained model.

4. Test the model performances:

   java -cp "stanford-ner.jar:lib/*" edu.stanford.nlp.ie.crf.CRFClassifier -loadClassifier ner-model.ser.gz -testFile test.tsv > test.res
   

   The input test.tsv has the same format than the train.tsv file. The output in test.res has an extra column containing the NER predicted class. The last lines also show the summary in terms of precision, recall and F1.

5. Finally, you can use your NER on real data:

   java -cp "stanford-ner.jar:lib/*" -mx5g edu.stanford.nlp.ie.crf.CRFClassifier -loadClassifier ner-model.ser.gz  -textFile test.txt -outputFormat inlineXML > test.res
   

Hope it helps.