Confuse/Meaningless Intent Recognition Problem

Hi everyone. When I develop my Rasa Nlu based on some projects, here is one thing that I made almost no progress during this month: confuse/meaningless intent recognition. Users in my project will say something meaningless/confuse words, i.e no need to recognition. The first thing I tried create a new intent “meaningless” and added 100 examples under it. It works but in a very bad accuracy. I thought it is just a simple problem and add more examples which may solve it. So I added it to 1500 examples, around 1/5 total examples of the whole rasa nlu model. But sadly in my test set, it only achieved like, 66% accuracy with 170 querys whose true label are all “meaningless”. Another problem is that there is a certain probability that it will affect other intent recognition, if I continuely add more “meaningless” examples, which will make the total accuracy drop massively. Any good ideas on how to solve this problem? Or any experience on similar problems? Thanks a lot. It really confuses me as it recogizes “confuse”.

It’s common to create an out_of_scope intent. You’ll find an example in the rasa-demo bot here and in this blog post.

Oh great. I will check it out later. Thanks!