Rasa does not do well in identifying intents that are close to each other

Hello I am working on a chatbot which is the Persian language. rasa does not do well in identifying intents that are close to each other. I defined between 5 and 10 examples for each intent in nlu file. Can you help?

Hi @a_vakily

Please share couple of intents and its associated examples.

Does the intents have different entities?

5 - 10 examples are too less. I’d add more examples (at least >20, ideally > 100). Unfortunately there aren’t that many existing language models for Persian in contrast to “mainstream” language like English. :grimacing: If there are good language models available, then you can e.g. use spacy in your model to make use of the already trained language model. In your case you’re training a model from scratch which means that you need way more examples as if you could use an existing language models :grimacing:

If you think that that examples for intents are indeed very similar, you could also try restructuring your intents and then distinguish based on extracted entities.