Multiintents with TensorFlow

Hi RASAs :wink:

How can I parse intents like “Can I have a glass of water. Can I have a cup of tea” “Can I have a cup of tea and water”


AskForTea “Can I have a cup of tea”

AskForWater “Can I have a glass of water”

I mean it should parse to intent AskForTea+AskForWater

I can generate / design content of AskForTea+AskForWater intent by glueing AskForTea and AskForWater intent content’s together (give much more examples in RASA and longer training time)

or look in tracker’s responses with lower confidence: "Can I have a cup of tea and water" { “confidence”: 0.8798392415046692, “name”: “AskForWater” }, { “confidence”: 0.3667337894439697, “name”: “AskForTea” }

Any ideas? @

Check out this blog post for more information about multi-intents.


I would rather like not to augument data.

Having 3 intents A, B, C:

A: What is A? B: What is B? C: What is C?

with question: “What is A and what is B?

(Currently with no additional learning i have got, well, almost correct but insufficient score: 0.70 on A , 0.40 on B)

I would like these 2 intents which are contained in this message (A and B) to show up in the ranking like: 0.99 on A , 0.99 on B

@pax How to use multi-class classifier (?any architecture with or without tensorflow?). Important for me is having a tagging possibillity.