Discussion: One-level vs multi-level intent classification

Hi, I’ve been experimenting with rasa intent classification for a while now and one question comes to mind: when projects get really big, the number of intents will follows and thus making it a bit harder for the classifier to predict intents especially when some of them are trained on similar vocabulary. One idea to handle this (never tested it, correct me if I am wrong) can be to create multi level intents, where for example for each use case we create an intent, and each intent has a sub-intent and so on. (like the intent retrieval with ResponseSelector but not limited to one sub-level). So the question is why is it limited only to one sub-level ? If I have for example 1000 intents, with 2 first-level main intents, 5 second-level intents for each and 100 third-level intents for each, wouldn’t it be better (performance wise) to have 3 steps where at each step we predict the correct intent at that level and depending on that predicted intent reduce the prediction space for the next step and thus making it a bit easier for the classifier?