Currently, each utterance can only be classified with one intent. However, there can be multiple intents associated with each utterance. (E.g., Greeting, Inquire).
Annotation assumes each utterance is de-contextualized (i.e., self-contained). However, this is not always the case. In many cases, it might be useful for NLU to have access to the dialogue state.
We should consider refactoring NLU.