Can stories start with bot utterance trained with FullDialogueTrackerFeaturizer?

Rasa Open Source
1

Hello everyone,

I have 2 questions related to 1 issue,

A. Does anyone know if stories start with bot utterance can be trained with FullDialogueTrackerFeaturizer?

e.g.,

  ## bot_asking_feeling  
    - utter_ask_feeling  
  * inform_good  
    - utter_ask_to_listen_music  
  * inform_yes
    - custom_action_1

When I train KerasPolicy with MaxHistoryTrackerFeaturizer for above story, the training finishes without issue, but using EmbeddingPolicy with FullDialogueTrackerFeaturizer, it faces error saying “Found two unpredictable actions in one story. Check your story files.”

Have anyone successfully finished training stories start with bot utterance with EmbeddingPolicy + FullDialogueTrackerFeaturizer ?

B. By looking at rasa_core/featurizers.py , the above error is to detect
delete_first_state is true when event.unpredictable happens. Does anyone know what is this for? If I hack this and let actions.append(event.action_name) even when delete_first_state is TRUE, finish training and make things work. I was wondering if I am setting things for failure since I do not know why delete_first_state should only happen once. Would anyone help me understanding this?

Thank you very much!

2

Hi @yongwoo,

EmbeddingPolicy will freak out a bit with these stories because of the way we stitch them together before training. Try passing the --augmentation 0 flag

As for the reason this happens…actually it’s a bit of a bug, thanks for bringing it up! Essentially we are not expecting two unpredictable actions in a dialogue (usually it’s just the first action_listen) but as you can see, with augmentation it can happen. Would you mind adding this as an issue on our Github board?

3

Ho would you implement this for training under Rasa-X?