Failing Gracefully with the TwoStageFallbackPolicy


(Tobias Wochinger) #1

Check out our blog post on the TwoStageFallbackPolicy which came with Rasa Core 0.13. It provides an intuitive and user-friendly conversation flow in case of NLU messages with low classification confidence. Have you tried it already? Share your experiences with us in the thread below :rocket:


(Romain Huber) #2

Hey @Tobias_Wochinger, we are going to replace the new fallback policy in our project, and looking at the code it seems that the default configuration is only saying something like “Do you mean [name_intent]”, but in your blog, you are proposing a choice between (I think) the two highest confident intent.

Could you share the code for this custom action ?

I would like to personalise it with something like propose two intent if the difference in confidence between then is lower than 10, or only propose the highest confident one.

But it’s a really nice feature, everyone was looking forward to have it in our bot :rocket:


(Tobias Wochinger) #3

@huberrom Great to your that you happy with the new policy :rocket: This is our current implementation which we are using inour bot on the documentation: rasa-demo/actions.py at 096568e64df4c25902f91fcf7b409fa2fcd05b16 · RasaHQ/rasa-demo · GitHub .

We are currently also working on an improved version of it, which is currently a pr include entities in button titles by wochinge · Pull Request #277 · RasaHQ/rasa-demo · GitHub (might be merged till you read this :smile:).

The basic idea of the implementation is that we read the button titles from a csv file. We saw that it is important to add potentially recognised entities to the payload of the button and to the button message title. We first put the entities in brackets after the intent description, but now we want to actually use different button titles depending on the intent and the extracted entities.

Let me know if you have any more questions :rocket: