Rasa: ner_crf with tensorflow word embedding feature

[ctrado18]

Hey,

wouldn't is be interesting to use the word vectors from tensorflow embedding (or spacy) just as a new feature for ner_crf?

Could you some how acces the the word vectors from tensorflow_embedding when you run this first in the pipeline?
If I would try this, do I have to take care of Rasa specific habits (besides just the standard ML tasks) when I add this vector to the feature list? Until now, I am not so deep into Rasa architecture.

[Ghostvv]

This is a good question. Word_vectors are persisted in this tensor: https://github.com/RasaHQ/rasa_nlu/blob/2f31d7b53c9a596b254503c3ad94f8a05dddc7b7/rasa_nlu/classifiers/embedding_intent_classifier.py#L232 In order to get a vector for a word you would need to run self.session on self.word_vectors and feed the word

[ctrado18]

Thanks. I will have a look on it. 😄

[jmrf]

@ctrado18 did you start looking into it?

I am also interested in using the tf_embedding + ner_crf. If you are interested I could give a hand in bringing this together, evaluate etc

Let me know

[ctrado18]

@jmrf No, but I would like to see your results!

[akelad]

I'm gonna close this issue for now, as this isn't really an issue, but more of a discussion. The right place for this is our forum. We'd be super interested to see your results there!