Hi, I developed a rasa chatbot with 20 single intents, recently I want to handle multiple intents. So I wrote a python script to write every combination of the two most possible intents in a new YAML file. After I added multiple intents, I have 33 intents. This makes my chatbot confidence low, for example when I had single intents, “Hello” had confidence of 0.9 but now this word has a confidence of 0.6. This makes nlu_fallback intent trigger because I set the threshold of 0.7 for my fallback classifier. Also, this time chatbot lose its generalization, so that usually can’t predict unseen multi-intent examples as mentioned in rasa doc.
The model will not predict any combination of intents for which examples are not explicitly given in training data.
I Try chatbot with different thresholds for fallback classifier but it doesn’t help. Is there any way to segment user sentences based on intents before getting the message to the chatbot and send segmented sentences to the chatbot? Is there any multilingual pre-trained model for this work? because I train my chatbot for the Persian language.
I appreciate any help 