AI Glossary / Natural language and speech / Intermediate
Adjacency pairs
Definition
Adjacency pairs are linked conversational actions in which a first contribution creates an expectation for a particular kind of response. A question invites an answer, a greeting invites a greeting, and a proposal invites acceptance or rejection. The two contributions are called the first pair part and second pair part, even when other talk occurs between them.
What it means
The word “adjacency” can suggest that the contributions must be immediately next to each other. In conversation analysis, their relationship is more important than simple physical proximity in a transcript. Someone may ask a clarifying question before answering the original question.
For conversational AI, the concept helps identify what remains unresolved. A system should recognize which question a short answer addresses, whether a proposal was accepted, and when an intervening exchange temporarily suspends an earlier conversational obligation. Merely producing a relevant-sounding next sentence is not always enough.
How it works
The first pair part establishes a response-relevant action. The second part responds to that action, with its interpretation depending on context. “Yes” after a factual question can affirm a proposition; “yes” after an offer may accept the offer. The same surface word therefore supports different dialogue functions.
An insertion sequence can occur inside the pair. The recipient asks for information needed to respond, the first speaker supplies it, and the recipient then completes the original exchange. A dialogue system can represent this with linked acts and an unresolved-question state rather than pairing every turn with its immediate neighbor.
These are patterns of interaction, not deterministic laws. Participants may decline to answer, change topics, repair misunderstandings, or leave the conversation. An expected response should not be fabricated when the evidence is missing.
Three practical examples
1. Scheduling: “Does Thursday afternoon work?” followed by “I am travelling then” can function as a rejection even without the word “no.” The system needs the conversational context.
2. Customer support: A user asks about delivery timing; the assistant requests a region; the user supplies it; the assistant answers the original delivery question. The middle exchange supports completion of the outer pair.
3. Voice interfaces: A greeting followed by a greeting establishes an opening exchange. The assistant should then attend to the user's next action instead of treating every greeting as a request for detailed information.
Think of it this way
Think of a question as opening a labeled envelope awaiting a response. A clarification can open a smaller envelope inside it, which is handled before the original one is closed. The analogy illustrates tracking pending actions, but conversation is more flexible than a strict stack: people can abandon, combine, or revisit topics, and a single turn can perform several actions.
Worked example
Consider this invented four-turn exchange. Turn 1, customer: “Can you deliver the chair on Friday?” Turn 2, assistant: “Which postcode is the delivery for?” Turn 3, customer: “AB12.” Turn 4, assistant: “Friday delivery is available for AB12.”
The outer question-answer pair links turns 1 and 4. The inner clarification question-answer pair links turns 2 and 3. Four turns contain two linked pairs, but the outer pair's members are not consecutive.
After turn 2, there are two unresolved questions in this simple representation: Friday delivery availability and the postcode. After turn 3, the postcode question is resolved while the availability question remains pending. Turn 4 resolves that question with an affirmative answer.
Crucially, this exchange does not contain an order to buy the chair or an instruction to book delivery. Confirming availability is different from accepting a purchase or authorizing a transaction. A system that marks a delivery as booked after turn 4 has inferred an action beyond these pair relationships. The example demonstrates dialogue bookkeeping, not a claim about any real delivery service.
Illustrative case study
Fictional case: Harbour Furniture develops a support assistant. During testing, it repeatedly loses the original question after asking for a postcode and responds with a generic acknowledgment instead of delivery information.
The team annotates the clarification pair and links it to the pending availability question. Its revised dialogue design resumes that question once the missing detail arrives. Reviewers test explicit answers, indirect answers, corrections, and abandoned requests. This fictional case illustrates the design problem without reporting invented improvements or treating one dialogue pattern as universal.
Common misconceptions and mistakes
Do not pair every two consecutive turns automatically. A long turn may contain several questions, while an answer may address an earlier one. Nor should “OK” always be labeled acceptance: it may acknowledge hearing a message without agreeing to a proposed action.
Another mistake is equating a conversationally expected response with permission. A request for information and an instruction to perform an operation have different consequences. Silence, topic change, and ambiguous acknowledgment should be represented honestly rather than converted into a positive answer.
Business use
Adjacency-pair analysis can improve the design and evaluation of support bots, scheduling assistants, and spoken interfaces. It provides a way to ask whether an assistant actually responded to the user's action and whether it returned to an interrupted task.
Useful evaluation examples include nested clarification, multiple outstanding questions, indirect refusal, and correction of a previously supplied detail. A high count of fluent replies says little about whether the correct conversational dependencies were resolved. Review the linked actions and resulting state, not just the wording of each isolated turn.
People also ask
Must the pair contain exactly two sentences?
No. Pair parts are conversational actions, and a contribution may contain several sentences or only one word.
Can another question appear before the answer?
Yes. A clarification can form an intervening pair before the original pair is completed.
Does an answer always use the expected wording?
No. Context can make an indirect statement function as an answer, acceptance, or rejection. Ambiguity may still require clarification.
Check your understanding
Someone asks “Shall we meet at three?” and the reply is “Which time zone?” Has the meeting proposal been accepted?
Answer: No. A clarification sequence has begun. The original proposal remains unresolved until a later contribution accepts, rejects, or otherwise addresses it.
Related terms
- Dialogue act
- Clarification
- Grounding
- Common ground
- turn-taking
- insertion sequence
- conversational initiative
- dialogue state
Learn next
Annotate a short conversation with first and second pair parts, then add clarification sequences and grounding acts. Check which requests remain pending after each turn and distinguish understanding from agreement.
More in Natural language and speech
- Add gate - An add gate is the learned control in an LSTM network that regulates how much newly proposed information enters its memory state.
- Add-k - Add-k is a probability-smoothing method that adds the same positive amount k to every possible outcome's observed count before normalizing.
- Adequacy - Adequacy is the degree to which a translation preserves the meaning of its source text.
- Adjective - An adjective is a word class whose members typically express properties of nouns or the things those nouns refer to. In “a quiet room,” quiet modifies room.
- Adverb - An adverb is a word class whose members commonly modify verbs, adjectives, other adverbs, or a wider proposition.
- Adversarial stylometry - Definition: Adversarial stylometry studies deliberate changes to writing intended to disrupt automated inferences about its author.
Sources
Source: Jurafsky and Martin, Speech and Language Processing, Chapter 26, Section 26.1.4, read for adjacency pairs, pair parts, and intervening subdialogues: https://web.stanford.edu/~jurafsky/slp3/26.pdf . The chapter is labelled a draft stub by its authors. The examples and operational analysis here are original; the supplied full-book retrieval was unavailable.
Previous: Adequacy Next: Adjacent values