Suddenly my test stories fail because many of the test stories predict action_default_fallback. Can anyone explain why this can happen?
action_default_fallback
- intent: fines_penalties entities: - offence-mask: kutambara agapfukamunwa - action: action_check_fines_penalties # predicted: action_default_fallback
This happens for many more test cases
The following applies to DIET, ResponseSelector, and TED:
I suggest using Tensorboard with Rasa so that you can look at the accuracies.
Maybe you need more or less epochs for example.
You can also decrease the fallback threshold in the pipeline/policies. A threshold of 0.1 for example means that fallback will be triggered if confidence goes below 0.9, a threshold of 0.2 is for a confidence of 0.8, etc.