We are training a rasa model in the intelligent home area. We define 619 intents offering user nlp for airconditioner/tv/refrige etc., the accuracy of the model is quite low. When we seperate those intents into two models, (310 for the tv, 309 for other application), the accuracy is acceptable. We wonder is there a limit for the count of intents in one model or what can we do to improve the accuracy?
Modelling your intent is not an exact science, rather it is about the sentences you have encoded as examples for each intent. some of them might make more sense merging them together for example
account_balance and card_balance usually have the same intent with two different product class. This is important when you are creating your intents and providing examples. It is a little obvious that reducing intents would definitely make the model better since there is less variance in examples now. It is important to ensure what intents would make sense from a language perspective. Models don’t understand the subtlety of english language as we do.
Please revise your intents using evaluation techniques to see confusion matrices and iron out intents that would make more sense merged together instead. that should improve your accuracy
2 Likes