Beyond natural language
Whereas natural language derives intent from individual bits of language, natural conversation derives intent from sequences of bits of language and action. Saying is also always doing. Verbal actions are not only accomplished through the meaning of the words but through the social context in which they’re produced. Natural conversation consists of recognizable types of action sequences.
The building blocks of most natural conversation are pairs of actions. Conversation analysts call these “adjacency pairs,” and they include common pairs such as question-answer, greeting-greeting, request-accept/deny, offer-accept/reject, complaint-remedy/excuse, and more.
These action pairs are like an accordion: they can occur in a compact two turns, or they can be expanded to arbitrary lengths with additional dependent pairs, or “expansions.” For example:
In the example, the base action pair is marked in bold, but the additional pairs are used to manage the conversation itself in a variety of ways.
As we can see, expansions of base action pairs are critical resources because they enable speakers to manage dependencies, hearing troubles, understanding troubles and closings.
With adjacency pair expansions, speakers can tailor their talk to their particular recipients by adapting to what the recipient knows and doesn’t know, as well as how the conversation has progressed.
With a robust expansion and repair system, your conversational agent does not need to achieve mutual understanding with the user on the first try. Instead, the agent and user can try an efficient utterance first and then expand as needed based on the recipient’s response. This is what humans do.