What events count as feature for training Rasa core

Rasa Open Source
1

Hi there!

I am trying to better understand the internal mechanics of Rasa core. In the KerasPolicy for example, what is considered a feature? What will be part of the max_history?

My understanding is that we will have at least the following as features:

  • intents
  • featurized slots values
  • entities (?)

But how about actions, utterances, and forms? Or form{"name": "xxx"}? And the requested_slot? Will they be used to predict the next actions?

The reason I am asking is that we have quite a few notifications (ie: actions that we execute through the API instead of predicting them) and we want to better control what will happen after they are triggered. Also, I’d just like to better understand what is going on in general.

Thanks a lot for the info! Cheers, Nicolas

2

@nbeuchat You can read more about the featurization for core policies here - Featurization Both KerasPolicy and EmbeddingPolicies use the featurization methods described above. Architectures for both these policies can be found here - Policies

To be precise, yes previous actions and active form is part of the feature bag that’s passed to the model. Actual text of utterances is not part of the featurization.

1 Like
3

Hi @dakshvar22, thanks a lot for your response, this is very helpful in better understanding what is going when training, especially with manually crafted stories.