0

I'm working on a supervised multi-label ML model which for now works by predicting a "tag" (label) for an input from the user, based on the dataset it was trained with. The training dataset had two columns - posts & tags.

Whenever I give anything as an input, the prediction is a relevant tag for it. But when the input is something which is not in the dataset, the output is again some random tag. I'm looking for a solution in the context of creating a chatbot, which would return something like "Sorry, I couldn't understand that" if the input is something different from what the machine has been trained for. How do I do that? Is there any way to get an "empty label" in such a case so that I can simply map my responses accordingly for that condition?

Thanks in advance!

Gizmosoft
  • 133
  • 1
  • 5

1 Answers1

0

You can use a softmax in your last layer and create a treshold for your prediction.

If the score of your maximum prediction is under the treshold, the chatbot can answer "Sorry, I couldn't understand that".

As you didn't give any code, I can't show you how yo add a treshold but it is quite easy to find it on Stack Overflow.

Jonor
  • 1,102
  • 2
  • 15
  • 31