Conversation and its Structure

Conversation is an intricate joint activity with structure, whether between people or between people and language models; the concepts are useful tools for analyzing human–LLM conversations too.

A single contribution from one speaker, as in a game: I take a turn, then you. A turn can be one word (“OK.”), a sentence, or several sentences.

A system must know when to stop talking (the user may interrupt, perhaps with a correction) and when to start: people begin turns almost immediately after the other finishes, because they predict when the other is about to finish.

Detecting whether the user is done speaking so the system can respond; it is hard because of noise and because people often pause mid-turn.

Each utterance in a dialogue is an action performed by the speaker (Wittgenstein, worked out by Austin 1962).

Directive; directive (a question politely commands an answer); constative (stating a constraint); acknowledgment.

Common ground: what the participants both agree on (Stalnaker). Grounding: acknowledging that the hearer has understood the speaker (Clark), or signaling problems in understanding so mutual understanding can be reached.

An elevator button lights up when pressed, acknowledging that the elevator has been called (Norman 1988).

Presentation: the speaker presents an utterance, performing a speech act. Acceptance: the hearer grounds it, showing whether understanding was achieved.

Explicit acknowledgment (“OK”, “Right”); demonstration by repeating (“what day in May…”); next contribution starting with “And”, implying the answer was understood and this question is in addition.

A short optional utterance like uh-huh that acknowledges the other’s utterance and doesn’t itself require acknowledgment.

Paired dialogue acts from conversation analysis, a first pair part setting up an expectation for a second: QUESTION → ANSWER, PROPOSAL → ACCEPTANCE/REJECTION, COMPLIMENT (“Nice jacket!”) → DOWNPLAYER (“Oh, this old thing?”).

Material separating the two parts of an adjacency pair. A correction subdialogue: while the agent seeks a May 15 return flight, the client asks “what day of the week is the 15th?” and then decides to stay until Sunday, changing the plan.

A side sequence between a REQUEST and its RESPONSE, common after ASR errors: “What do you have going to UNKNOWN on the 5th?” — “Going where on the 5th?” — “Hong Kong.” — “OK, here are some flights…”

A preliminary exchange before a request, e.g. “Can you make train reservations?” — “Yes I can.” — “Great, I’d like to reserve a seat on the 4pm train.”

Control of the conversation. An interviewer has initiative over the interviewee. Mixed initiative, where control shifts back and forth, is the human norm. Primitive systems use system initiative (the user must answer the system’s question) or user initiative (a search engine passively answers queries). LLM systems come closer to mixed initiative but still don’t switch naturally.

Conversational implicature: the client doesn’t literally answer, but the agent infers travel dates (likely arriving on the 11th).

Grice’s term for a class of licensed inferences: the speaker communicates more than the words say, and the hearer infers it using maxims, general heuristics that guide interpretation.

Speakers try to be relevant, not utter random speech acts. Hearing about a meeting, the agent asks why it was mentioned: the meeting is probably the reason for the trip, so the flight should arrive before it.

A tag representing the interactive function of an utterance, combining speech acts and grounding; tagsets depend on the participants and on whether the dialogue is task-based.

Each act has slot–value arguments: “I’d like an Italian somewhere near the museum” = inform(food=Italian, near=museum); “Is it reasonably priced?” = confirm(pricerange=moderate); “What is the phone number?” = request(phone).

CONFIRM explicitly confirms values; CONFREQ (system only) implicitly confirms some values while requesting another; NEGATE (user only) negates and gives the corrected value; DENY (user only) just denies that a = x.

Forward-looking functions (speech-act functions like STATEMENT, INFO-REQUEST, ACTION-DIRECTIVE, OFFER, COMMIT) and backward-looking functions relating to the interlocutor’s previous utterance (agreement, answering, grounding).

STATEMENT; INFO-REQUEST; CHECK (confirming information); influence on addressee (Bach’s directives): OPEN-OPTION (weak suggestion), ACTION-DIRECTIVE (actual command); influence on speaker (commissives): OFFER, COMMIT; conventional: OPENING, CLOSING, THANKING.

Agreement (ACCEPT, ACCEPT-PART, MAYBE, REJECT-PART, REJECT, HOLD, often via subdialogue); ANSWER; understanding: SIGNAL-NON-UNDERSTANDING, SIGNAL-UNDERSTANDING via ACK (continuer or assessment), REPEAT-REPHRASE, or COMPLETION.

A	[info-req, ack]
C	[assert, answer]
A	[info-req, ack]
C	[assert, answer]

The “And” at the start of each question grounds the previous answer (ack).