How should bot phrases be grouped for the intent recognition algorithm to work effectively?
We should try to put (group) all sentences with the same ACTION in the same INTENT, e.g. all phrases denoting resignation should be with the intent of RESIGNATION.
If we do not have such a possibility, because the process is so divided or the language is ambiguous, make sure that:
There was a clear distinction between these phrases in the form of a different type of OBJECT that we will define as an entity in the bot, e.g.
- Give up a complaint - COMPLAINTS
- Give up a loan/product - CANCELLATION
So all sentences with a verb about cancellation and different products will be in the intent of CANCELLATION, and only the sentence about cancellation (and its paraphrases) from the complaint will be in the intent of COMPLAINT.
However, it should not be possible for some of the sentences containing the cancellation from the product (e.g. a deposit, loan) to be in the intent of a COMPLAINT, and some in the intent of CANCELLATION.
Otherwise, we have similar sentences that neither humans nor machines can clearly distinguish. Consistency in the collection gives higher chances for better effectiveness :)
Updated 3 months ago