Multi-intent classification without multi-intent training samples?

I have training data consisting of single intents only.

I now have a requirement to handle multi-intents.

I have read the following:

Is it possible to handle multi-intents without having to create training set examples of multi-intents using the tokenized intent method suggested? (ie thanks+goodbye)

Perhaps using some confidence level threshold for any intent to classify something as multiple intent? That is, if “Thanks and goodbye” comes back .70 “goodbye” and .50 “thanks” could you say that it is both?


Would like to know this as well. So far I have found nothing.

Hey @zeroangel, to use multi-intents it is necessary for you to provide corresponding composite intent examples. Currently, it is not possible to do what you suggested using Rasa.