In TwoStageFallbackPolices,if the nlu cofidence is below the threshold, then a button list will be show to the user and let them select the right intent. What if I type the message to the robot rather than click the buttons? I try this and get some results:
conversation above like this
Me: can you give the some information about LOL?
robot: Do you mean…?
1.I want to get information about LOL
2.something else
Type out your message…
Me: type my qusetion …
robot: I don’t know what are you saying ,can you rephrase it …?(action_default_fallback trigger by the TwoStageFallbackPolices because cofidence below nlu threshold)
Me: type my qusetion again…
robot: I don’t know what are you saying ,can you rephrase it …?(action_default_fallback trigger by the TwoStageFallbackPolices because cofidence below nlu threshold)
Me: type other qusetion …
robot: I don’t know what are you saying ,can you rephrase it …?(action_default_fallback trigger by the TwoStageFallbackPolices because the intent confidence of this question still below nlu threshold)
Me: type other qusetion …
robot: return the right message (the intent confidence of this question over nlu threshold and KerasPolicy take over control!)
After I type my message, the action_default_fallback was triggered by the TwoStageFallbackPolices . I keep asking another three question to my robots. However, because these thress message’s intent confidence is below the nlu threshod , the action_default_fallback will be triggered again and agin by the TwoStageFallbackPolices until I ask onther question whose intent is over the nlu confidence.
If user keep asking robot some questions whose intent is below the nlu threshold,the TwoStageFallbackPolices will response like the FallbackPolices, which I think is not a good behavior.