How does Multi-Intent classification work?

Hi - I was hoping to get some more details on how multi-intent classification works ‘under the hood’.

I’ve had a look at the docs, and the suggested tutorial as well as the GitHub repo for that tutorial. Unfortunately, I haven’t found any of those sources particularly enlightening when it comes to how the multi-intent classification itself works.

In the source code I’ve see that the intent names seem to be split into tokens here but I am unclear as to what the implications of this are. I think my query is quite similar to this previous question here

I would be really grateful if anyone was able to shed some more light on the topic for me. Thanks in advance!

Bump! Hoping that someone see this and gives me some advice!

The Tensorflow Pipeline Tutorial is deprecated as multi-intent classification is done now by DIETClassifier (I guess?). But I’m also interested in more details like:

  • training data for intent “ask_question”
  • training data for intent “topic_book_hotel”
  • do we need more training data for the multi-intent “ask_question+topic_book_hotel” - how much?

Other case:

  • training data for “inform”
  • training data for “thanks”
  • do we need for each story the turn “thanks+inform” (for example) for
    • utter_welcome
    • utter_process_information
  • or is it possible to have one story only for thanks and utter_welcome that is merged with other stories if we have multiple intents like the one above?

Thanks :slight_smile: