Is there a similar way to print out action classifications as interpreter's parse?

For example, result = interpreter.parse(“Hello!”) returns a list of intents with varying probabilities, and this is very good and intuitive to gauge the quality of the NLU classification. Is there a similar way to do this in Rasa Core?

If I use responses= agent.handle_text(“Hello!”), the object returned isn’t the list of action, but the current action selected by the dialog manager. How to achieve something similar to the NLU’s parse so that it is very convenient to debug on command line?

Hi @lingvisa,

The setup for actions is a little different from that for intents – yes, there is a list of actions with their probabilites, but it corresponds with the best policy chosen for that turn in the dialogue. So really, seeing this information wouldn’t be all that helpful because you wouldn’t see all the other policies’ confidence matrices. However, with multiple policies and lots of actions, storing everything instead of overwriting could get very big, so we don’t currently store it all together. If you want to see deeper into how the policies are predicting the actions, I would suggest throwing some print statements into the ensemble method that chooses the action here.

Thanks Ella. I will explore that direction.

1 Like