From the perspective of a dialog system I see it as a bunch of assumptions that are made prior to engaging in the dialog. And given that hypothesis I have looked at people interactions to see how people deal with this sort of 'communication mismatch' and how they detect it.
Certainly there are repeated asks of the same question slightly rephrased as an indicator of a mismatch, but there is also a general resistance to interacting. So with people if someone gets frustrated talking with you and stops, I've seen people use that as a signal to seek out an understanding of where the mismatch is, but in computer dialoging systems that check, and subsequent re-framing is completely missing.
The other thing I've observed is that often a dialog system seems to try to be 'human' in its interaction but because it is a computer the user communicates to it like a 'user' not like a human. Adding what might be relevant search terms to the utterance as an example. For example, I listened as a person pitched to Siri "I want pizza" and was frustrated at the response ("This is what I've found on the web about Pizza") And added "I want Dominos Pizza" (additional search term, vendor name). But Siri appeared stuck on being unable to parse an acceptable language target for 'I want'.
Changing that to "Call dominos" or "Where is the nearest Dominos" works well because it as a pre-built in answer action (Telephone call, map directions).
All of this the "level 4" version of autonomous conversations where the computer can navigate what it is you are saying and what you expect as a response.