Interactive Learning while using TwoStageFallbackPolicy

Rasa Open Source
1

So I am using TwoStageFallbackPolicy and as suggested, used Interactive learning to create stories for a form. Now TwoStageFallbackPolicy asks user to affirm the intent if the confidence is below a threshold value.

As anything I type will fall into one of the intents defined (however low the confidence might be) Let’s assume what I typed fell under user_email intent. Now if the confidence is really low, the TwoStageFallbackPolicy would perform action_default_ask_affirmation to ask user for affirmation.

Now my story looks something like this.

    * user_email
        - action_default_ask_affirmation
    * ....

I am confused about this. Won’t this cause my stories to become inconsistent? As this is not what I want as a story in-case user_email was indeed the correct intent to go with. Won’t my core model get confused with this as now I might have two different stories

* user_email
    - action_default_ask_affirmation
* ....

and the original

* user_email
    - action_user_mail
* ....
2

You don’t want to define it in your stories. Its a policy similar to mapping in the sense that it doesn’t need to use the stories and will instead predict the specified actions directly when the threshold is reached. Just keep the original

3

@Mappi As mentioned above I am not putting it in the stories. It’s getting put there as a story when I use interactive learning. Everytime I train my model interactively and if the TwoStageFallbackPolicy is invoked, it creates these weird stories.