Check if text features are present

[tabergma]

Some users are getting intent predictions with high confidence for words that aren’t in their training data (e.g askdjflksjdf) since they upgraded to 1.6.1. They’re using word level featurization, it should be assigning the intent None.

They noticed that this check here : https://github.com/RasaHQ/rasa/blob/1.6.x/rasa/nlu/classifiers/embedding_intent_classifier.py#L825
doesn’t check whether there’s text features present any more like it used to:

rasa/rasa/nlu/classifiers/embedding_intent_classifier.py

Line 617 in f1520ec

if X.any() and intent_ids.size > 0:

Rasa version: 1.7.0
Python version: 3.6