Inconsistences in Policy predictions

Rasa Open Source
1

I have trained stories with two similar use cases, with actions of the type “out_of_scope” and “chitchat” (using response selector). The two actions can be performed in isolation or as part of an unhappy path within a form, for which a question was established to know whether or not the user wants to exit the form. When verifying the paths with VISUALIZE, it perfectly detects the next action in all cases, but when running with RUN it only predicts one of the cases well with TED_policy and the other does not (through Mappings predict Action_listen), I do not understand why this happens.

2

sorry, could you please explain what is the problem in more detail?

3

Stories by example:

story_01

  • start_session
    • action_session_start
    • slot{“nombre_usuario”:“John”}
    • slot{“apellido_usuario”:“West”}
  • search_product
    • form_product
    • form{“name”: “form_product”}
  • out_of_scope
    • action_out_of_scope
    • action_ask_continue
  • affirm
    • utter_continue
  • form_product
  • chitchat
    • respond_chitchat
    • action_ask_continue
  • affirm
    • utter_continue
  • form_product
    • form{“name”: null}
    • slot{“pointer”:2}
  • bye
    • utter_bye
    • action_experience
  • save_note
    • action_save_experience

story_02

  • start_session
    • action_session_start
    • slot{“nombre_usuario”:“John”}
    • slot{“apellido_usuario”:“West”}
  • out_of_scope
    • action_out_of_scope
  • chitchat
    • respond_chitchat
  • bye
    • utter_bye
    • action_experience
  • save_note
    • action_save_experience

when i run RASA TEST and then RASA VISUALIZE all is perfect.

but when running RASA RUN , for stoty 01 , only the intent out_of_scope works well and for chichat never ask to continue or not because predicts Action_listen instead of ask_continue

4

story_01 (correct indentation)

  • start_session

    • action_session_start
    • slot{“nombre_usuario”:“John”}
    • slot{“apellido_usuario”:“West”}

  • search_product

    • form_product
    • form{“name”: “form_product”}

  • out_of_scope

    • action_out_of_scope
    • action_ask_continue

  • affirm

    • utter_continue

  • form_product

  • chitchat

    • respond_chitchat
    • action_ask_continue

  • affirm

    • utter_continue

    • form_product

    • form{“name”: null}

    • slot{“pointer”:2}

  • bye

    • utter_bye
    • action_experience

  • save_note

    • action_save_experience
5

story_01

  • start_session

    • action_session_start
    • slot{“nombre_usuario”:“John”}
    • slot{“apellido_usuario”:“West”}

  • search_product

    • form_product
    • form{“name”: “form_product”}

  • out_of_scope

    • action_out_of_scope
    • action_ask_continue

  • affirm

    • utter_continue

    • form_product

  • chitchat

    • respond_chitchat
    • action_ask_continue

  • affirm

    • utter_continue

    • form_product

    • form{“name”: null}

    • slot{“pointer”:2}

  • bye

    • utter_bye
    • action_experience

  • save_note

    • action_save_experience
6

do you predict respond_chitchat using MappingPolicy? which policy predicts action_listen?

7

In both cases using policy = MappingPolicy. Both the Chitchat and OutofScope intents have triggers to the correspondent action in domain.yml Instead with another attempt called Stop, it works fine, and that intent doesn’t have any triggers. Could it be this? What puzzles me is that in VISUALIZE the same does not happen

8

Solved! I remove triggers in domain.yml and let it infer actions only based on stories and it works. From now on I will use triggers as little as possible. Thanks Vladimir

1 Like