You're touching upon a very nuanced point of identifying the true "intent" behind these words. Do you think the way these models are trained should be different to correctly map the potential intent vs the true meaning?

like your example, "the meeting is at 3pm", _we got enough time_ intends something else with "the meeting is at 3pm" _where the hell are you?_ intends something else. It is not so obvious to get that intent without a lot of context (like, time, environment, emotion etc.)