Does this transformer architecture for intent classification introduces any kind of uncertainty in the usage?