Rasa NLU 0.13.0 is released!


(Juste) #1

Rasa NLU 0.13.0 is released!

Please roll the drums! :drum::drum::drum:

Rasa NLU 0.13.0 is now released! It includes lots of new features like:

  • language-agnostic NER
  • docs translated into Chinese by our amazing contributors!
  • lots of improvements on ner-crf

and much more so make sure to check out the changelog and give rasa nlu 0.13.0 a go!


(Asokolow) #2

Hey, are the changes available on the github ?


(Juste) #3

Hey @asokolow. Yes, the master branch on GitHub contains all the latest changes.


(Pierre Paul) #4

Nice! Good job! Sadly, this also means the 0.12 doc doesn’t work anymore :frowning_face:

http://rasa.com/docs/core/slots/0.10.2/slots.html


(Juste) #5

Hey @PierrePaul. Thanks for pointing this out - there’s definitely an issue with the links. While we fix this, you can switch to a different version of the docs from the index page of the version of your choice. So for example, you can change to 0.10.2 (or any other version) from http://rasa.com/docs/core, then you can switch to any other version from http://rasa.com/docs/core/0.10.2/index.html and so on.

We will fix this as soon as possible :slight_smile:


(Pierre Paul) #6

Awesome thank you!


(Anders) #7

“language-agnostic NER” - so duckling is obsolete? fingers crossed


(Souvik Ghosh) #8

No, language-agnostic NER is NER_CRF that is to train custom entity. previously the ner_crf was dependent on spacY’s tokens to be able to build a model for entity extraction. that means you will have to provide a spacy’s vectors in order to extract entities making it dependent on the languages provided by spaCy. However now, language-agnostic NER CRF actually uses whitespace tokens( tokenised by whitespace) in order to build the entity extraction model. so it does not depend on spacY anymore so you can build a NER on Klingnon if you like :smiley:

Duckling is a rule based entity extractor so very system like entity extraction like Date, time, duration, distance, currency. You can also add custom dimensions in duckling for countries, cities etc etc since they all can be predefined by a set of rules unlike CRF where entities are extracted from patterns.