I am using a lot of intents and many of them contains non meaningful words like:
- “Je voudrais que tu [action]…” (I want you to [action]…)
- “J’aimerais que tu [action]…” (I’d like you to [action]…)
- “Tu peux [action]…” (Can you [action]…)
- “Pourrais-tu [action]…” (Could you [action]…)
The [action] part is what is really important
This adds a lot of possible variations to the intents for no real value for the NLU.
What is the acceptable upper limit as to the number of training sentences for an intent (if I add all possible variations, I may very well get to multiple 10,000s)?
Is there a way to simplify the training sentences to ignore these variations?
What may be the best way to define training sentences to maximize NLU recognition across a wide range of requests without creating all possible variations?