Tiered models in NLU

With a large domain area to cover, many people resort to a multiple model to classify both intents and entities. For instance if I had a Chatbot for Reservations, my first model might decide if it was a hotel reservation or a restaurant reservation. From there I may have 2 sub models that are called based on the result of the first model (only one gets called at any one time, both they both need to exist). Is there a way to support this ‘branching’ behavior in the NLU pipeline? For the chatbot itself, it would be no different than one monolithic model that does everything, but if I had a wide range of topics, I may want to split up my models into a ‘Gross’ several ‘Fine Tuned’ (both intent and entity) models that are area specific.

How would one attempt to handle this branching behavior?

Hi Robert,

thanks for posting - what you describe is a pretty common thing for larger projects. Do you currently have a dataset where NLU is not performing well? Have you considered using multi intents like book+hotel and book+restaurant? there’s a tutorial here