Rasa Algorithm Whiteboard - Diet Architecture 2: Design Decisions

[@eturkoz]

I am really impressed by your way of explaining and handling the subject.

[@jayhu6075]

Great explanation. Hopely more in the future. Thanks.

[@igen8603]

That's very interesting ! Is it feasible to modify the DIET architecture in order to extract several intents (without knowing the number of instructions present in a natural sentence)?

[@RasaHQ]

Technically the output of the model is an array of intent and entitiy probabilities. Technically you could hack around with this array to get multiple intents out but this would involve a lot of hacking. DIET is designed for use-cases where we need to classify towards a single intent.

[@RasaHQ]

It deserves mentioning that Rasa does support retreival models that are in line with what you suggest and we call these ResponseSelectors. They're designed to handle simple responses like FAQ and chitchat. This may be of interest to you in the meantime.

[@TheZiolite]

Is it possible to train the entity extraction part?

[@ilyaovchinnikov6256]

Hello. Could you explain how do you encode input? How "intent: play_game" is represented?

[@venkateshdas5422]

Hi. I do have a very silly doubt. How can we have a shared weight on all blocks of Feed Forward Neural Network when we have three different loss functions? How does the backprop work here? (It is also not clearly mentioned in the paper).

[@venkateshdas5422]

Does this mean that the weights are updated only based on the total loss? So only the total loss is optimised?