Predicted intent is always 'greet' for every message that doesn't exists in the training data

Rasa Open Source version

2.8.12

Rasa SDK version

No response

Rasa X version

No response

Python version

3.8

What operating system are you using?

Linux

What happened?

‘greet’ intent always return as the selected intent with confidence 1 when user types a message that doesn’t exists in the training data.

nlu.yml

version: "2.0"
nlu:
  - intent: greet
    examples: |
      - γεια
      - γεια σας
      - καλησπέρα
      - καλημερα
      - kalispera
      - geia
      - geia sas
      - kalimera
      - hey
      - hello
      - hi
      - hello there
      - good morning
      - good evening
      - hey dude
      - goodmorning
      - goodevening
      - good afternoon
  - intent: affirm
    examples: |
      - yes
      - sure

domain.yml is the default config.yml is the default + added FallbackClassifier with threshold 0.3

User input: “I am writting something”

predicted intent is always “greet”

Command / Request

No response

Relevant log output

No response

[koaning]

My gut says that adding more examples will go a long way. I’d try having at least 12 examples per intent. At the moment, the model may be overfitting on the few examples that exist and is thus learning only to output high confidence values.

[koaning]

I’ll close this issue since it doesn’t seem like Rasa is directly breaking here.

If you’d like to discuss this issue further, please start a thread on our forum at https://forum.rasa.com. That’s a preferable place to talk about specific bot issues. You can ping me via my @koaning user name there as well.