Interactive learning with ner_crf and pre trained extractors

akela
(Maurice Doepke) #1

During interactive training the intents found by pre-trained extractors + regex like extractors like spacy and duckling are automatically annotated as entitys. The problem is that spacy and duckling will not profit from this and it might even confuse the ner_crf extractor.

Is there a way to turn this behaviour off or am I wrong and it is in some way useful for entit extraction? Here an example, where time is extracted by duckling:

Is the NLU classification for 'I need an appointment [from next week wednesday](time:{‘from’: ‘2019-02-20T00:00:00.000+01 :00’, ‘to’: None})from next week wednesday with Intent classification…

0 Likes

(Akela Drissner) #2

Hmm, do they get saved to the file as well? I know in the json format of NLU, there’s an extractor field and the entities labeled with anything other than ner_crf are ignored.

If they do get saved to the file normally though, please raise an issue on Rasa Core

0 Likes