How to handle Multi intents

Rasa Open Source
22

hey! you’re absolutely right that these are just new classes. The difference is that the way DIET works is that the labels are also embedded in the same vector space as the utterances. By splitting the label into tokens, the label is then represented as the sum of multiple vectors. So if you have a token that shows up in multiple label names, the model already treats these labels as related by construction.

Moreover, if you have two phrases making up the multi-intent, e.g. “yes, I want a pizza” with label affirm+order_pizza, the model can just see this as vector addition of the phrases and the labels respectively

1 Like
23

Thank you for the fast and detailed response! This would seem to suggest that there is indeed a benefit to using multi-intents (provided you have sufficient samples for each), namely, to take advantage of this “relatedness” factor.