How to use Rasa X to improve trained model for better intent classification?

I run Rasa X web interface with the ‘rasa x’ command and then test sample phrases. Some of these conversations are not correctly classified for their intent, although I use the exact phrase as in the training data set in the nlu.yml file. Another intent is chosen which has higher probability.

Let’s say I manually picked the correct intent for this phrase. I manually fix a bunch of intent/response pairs where the FAQ classifier was failing. How can I extract this information and feed it to the next training of the model?


What do you mean? Didn’t you already do that in your previous phrase? Or did you mean using interactive learning?

When correcting the bot in Rasa X, in the Talk to Your Bot screen, on the right, you can save the story as new training data (in stories.yml).

Also, in the NLU Inbox screen, you can see some sentences with their detected intent and entities (if any) and the confidence.

  • If they’re correct, that’s great! You can click on the checkmark to reinforce that behavior - especially if the confidence is low.
  • If they’re incorrect, that’s okay! You can choose the correct intent and entities then click on the checkmark to correct that behavior.

After clicking on the checkmark, the message will be added to the training data under its corresponding intent (in nlu.yml).