Intent classification for negative sentences

I’ve trained my model with several utterances and it has great performance.

one thing that I am still unable to solve for is high confidence % of the negative statements. I am looking for help on the same.


I want to know what is surfing? - works well and is supposed to work as well

however, I don’t want to know what is surfing… - this also has high confidence % obviously but it should be going to fallback instead…

can anyone help

we can’t add this in out-of-scope intent since our training corpus is huge and adding everything is not possible,

We actually just covered this on the office hours: Rasa Office Hours: Henry Carse, PreSales Solution Engineer - YouTube

Possible options:

  • adding a shallow parsing component to handle modelling negation and scope (likely to lead to slightly more latency)
  • a negation intent that triggers a clarifying question (intent is likely to be fairly greedy, might need to do some data engineering to get around that)