Understanding Rasa Core supervised ML features

rasa-core

(Syed Rizvi) #1

Hello Team,

I am trying to understand Rasa Stack and create a chatbot.While rasa_nlu looks straightforward,i am more interested to know how machine learning is applicable to Rasa_core ? Understand that rasa core model training required stories and domain file.

My question is -

  1. How based on these two files features are being created for training the supervised ML algorithm ?
  2. Any NLU framework will provide intent and entities based on which rasa_core will take descisons.So how does the rasa_core use the intents and entities for prediction ? Basically i am not very clear with how rasa_core policy works.

Any articles,whitepapers suggestions around above queries will be appreciated !

Thanks.


(Vladimir Vlasov) #2

Hi, did you check out the docs? There is high-level architecture of rasa_core: http://www.rasa.com/docs/core/architecture/

If you want to go deeper into ML framework, here is our paper: https://arxiv.org/pdf/1712.05181.pdf


(Syed Rizvi) #3

Hi,

I have gone through the Rasa architecture and also the white paper mentioned.I was looking out for an example/tutorial in which based on the story and domain files,features are created and accordingly to see how model is being trained.I understand that some of the features which would go in would be intent,entities,slots,last actions etc.But would love to see a more detailed one rather than just high level picture.

Also - if you can confirm (in high level architecture) where is Machine Learning applicable ? Interpretor(NLU side) and Policy(Core side) ??

Also not clear how is policy deciding on the next action ? I believe it would be based on the input(entities,intent,slots) given to the trained model.No explicit statement on this.

What is the relationship of Rasa policy to Rasa ML model ?

Lets say for building the restaurant chatbot - once we train the Rasa Agent we get policy_0_MemoizationPolicy and policy_1_RestaurantPolicy folders each containing featurizer.json file containing some inputs.

Would love to hear some lines shedding some light on above like How are these policies related and how do they help to create the dialogue framework.

Thanks.


(Vladimir Vlasov) #4

Hi, in the docs there is a description of the policies: http://www.rasa.com/docs/core/policies/ and featurization process: http://www.rasa.com/docs/core/api/featurizer/. Please check this thread for high-level discussion: Stories and conversations - is my mental model right?