Conversational grounding is the collaborative process through which participants establish that they have understood one another sufficiently for their current purposes. In Clark and Brennan’s account (1991) 1, the criterion is adequate mutual understanding for the activity at hand, not perfect agreement about every possible implication.
Evidence of understanding
A contribution involves presenting an utterance and obtaining evidence that it has been understood. That evidence can include an acknowledgment, an appropriate next response, or continued attention. Repair and clarification become useful when the evidence indicates a problem. An explicit confirmation after every sentence is not required.
In a hypothetical scheduling exchange, “Friday at three?” followed by “Yes, Canberra time” helps establish a shared interpretation of the proposed time. A generic “OK” may provide weaker evidence about the time zone. What is sufficient depends on the task and the communication medium.
Grounding updates Common ground and is studied within Pragmatics. It should be distinguished from factual grounding in some AI discussions: attaching a source to an answer, or using Retrieval-augmented generation, does not establish that a conversational partner understood the answer. Conversely, participants can understand an assertion while disagreeing about its truth.
The original account concerns human communication. Applying its criterion to a conversational agent is a design interpretation; a fluent acknowledgment alone does not demonstrate reliable internal understanding.
Footnotes
-
Clark, Herbert H., and Susan E. Brennan. “Grounding in Communication.” 1991. In Perspectives on Socially Shared Cognition, 127–149, especially pp. 128–133 on the grounding criterion, presentation and acceptance, and evidence of understanding. Author-hosted chapter. ↩