I think there is a bad behavior in two_stage_fallback_policy.If a message is classified with an low confidence,then action_default_ask_affirmation will be excuted whick will usually show button message for user.
Users are asked to choose an intent from the button message.However, users can also type their own message regardless of button message.If the user’s new message is classified with low confidence,the fallback_nlu_action will be execued.
How can I identify this situation? In the rasa demo-robot,it give its implementation for “action_default_fallback”,which can’t deal with situaction above.It only revert one UserUtterance,which is the new message that users type. If user still type a message with low confidence next time, then two_stage_fallback_policy will excute action_default_fallback again and every time this function is executed ,the template “utter_default” will be executed,which like “I didn’t quite get that, could you rephrase your message?”
If user keep ask some question with low confidence,this action will be excuted again and agin and action_default_ask_affirmation will not be excuted utill user type some message with high cofidence.