Is a dialogue act the same as a topic?

No. Topic concerns what is discussed; the act concerns the contribution's function.

Can one utterance have several dialogue acts?

Yes, depending on its content and the annotation scheme.

Can dialogue acts be identified from text alone?

Sometimes, but dialogue history, prosody, and other context may be needed to resolve ambiguity.

AI Glossary / Natural language and speech / Beginner

Dialogue act

Definition

A dialogue act is the communicative function performed by an utterance or part of an utterance in a conversation. Examples include asking for information, answering, requesting an action, accepting a proposal, correcting a misunderstanding, and acknowledging what someone said. Dialogue-act labels describe what a contribution does in the interaction, not only its topic or grammatical form.

What it means

The same words can play different conversational roles. Okay might accept a proposal, acknowledge hearing information, or help close an exchange. Its interpretation depends on the preceding conversation and sometimes on tone, timing, or other signals.
A dialogue-act annotation scheme gives names to these functions so that conversations can be studied or processed systematically. Some schemes are designed for a particular task; others describe more general conversational behavior.
A contribution can have more than one function, and a single turn can contain several acts. For example, a speaker may acknowledge an answer and immediately ask a new question. Treating every turn as exactly one intent can hide that structure.

How it works

Choose a labeling scheme and decide which units will be annotated, such as utterances or smaller meaningful segments. Examine the contribution together with its conversational context and assign the relevant functional labels.
A computational system may predict those labels from text, speech features, dialogue history, or combinations of evidence. A task-oriented representation can also attach content, such as the requested field or the value being confirmed.
The labels can inform response selection, conversation analysis, or error review. Their usefulness depends on consistent definitions and realistic examples. A predicted label is evidence about communicative function, not certainty about a person's private intentions or automatic permission to perform an external action.

Three practical examples

A support assistant can distinguish a customer's answer from a correction to an earlier account detail.
A scheduling interface can separate a proposed meeting time from acceptance of that proposal.
A conversation analyst can study where users ask for clarification, helping identify prompts that repeatedly cause misunderstanding.

Think of it this way

Think of the moves in a cooperative board game: proposing, agreeing, asking, and revising are different moves even when they concern the same piece. Dialogue-act labels describe those interaction moves. The analogy is limited because human conversation is flexible, multifunctional, and context-dependent rather than governed by one fixed, universally agreed move list.

Worked example

Consider this invented exchange about preparing a report. Assistant: Which month should the report cover? User: September. Assistant: I will prepare the September draft. User: Actually, October, please. Assistant: October, understood. User: Thanks.
Under a simple task-specific scheme, the first assistant turn is an information request about the month. September is an answer supplying that field. I will prepare the September draft communicates the planned action and repeats the understood month. Actually, October, please corrects the earlier value. October, understood acknowledges the correction and displays the revised interpretation. Thanks performs a thanking function.
The exact labels depend on the scheme. Some schemes would represent the correction through multiple functions, such as rejecting the previous value and supplying a replacement. Others have a dedicated correction label.
Now replace the user's final Thanks with Okay after a proposal. That word might indicate acceptance, but an annotator should consider what was proposed and whether the context supports commitment. After a long explanation, the same word may merely acknowledge understanding.
The example demonstrates why topic extraction alone is insufficient. Both September and October concern the same month field, yet one supplies an initial answer and the other revises the shared state. A dialogue system needs to account for that relationship to avoid preserving the wrong value.

Illustrative case study

In a fictional service chat, an assistant asks whether the customer wants more information or wants to proceed with a change. The customer responds with a clarification question, but a simplistic classifier treats any positive-sounding reply as acceptance.
The team revises its examples and response policy to distinguish clarification from agreement. It tests turns containing acknowledgment plus a question, rather than assuming every message has one uncomplicated function.
The case illustrates that dialogue-act recognition is useful only when its distinctions connect to appropriate conversational behavior. A label should not replace checking whether the actual request and context support the next step.

Common misconceptions and mistakes

Do not equate dialogue act with grammatical sentence type. A question-shaped sentence can function as a request, and a statement can indirectly communicate a constraint.
Do not assume an acknowledgment is always agreement or commitment. Its role depends on context.
Do not assume there is one universal list of labels. Schemes differ in granularity and purpose.
Do not force every turn into exactly one act when it contains several functions. Also, topic and dialogue act are different: two utterances can discuss the same subject while one asks, another rejects, and another corrects.

Business use

Dialogue-act analysis can help teams understand how conversations succeed or break down. Useful questions include whether an assistant recognizes corrections, asks appropriate clarification questions, and distinguishes supplying information from making a commitment.
Evaluate labels in their conversational context and examine the resulting system behavior. High label accuracy on isolated sentences may not establish good interaction. Annotation guidance should explain ambiguous cases and how multiple functions are represented, so training data and quality reviews use compatible interpretations.

Check your understanding

A user says Actually, October after previously supplying September. Why is simply extracting both month names insufficient?

Answer: The later contribution corrects the earlier value. Its dialogue function tells the system how the information should update the conversation.

Related terms

Speech act
Grounding
Clarification
Adjacency pairs
intent recognition
acknowledgment
dialogue state
adjacency pair
and conversational context

Learn next

Annotate a short invented conversation with a small, clearly defined act set. Include one correction and one ambiguous acknowledgment, then explain how preceding turns affect the labels and the appropriate response.

Sources

Sources: Jurafsky and Martin, Speech and Language Processing, Chapter 26, speech acts, grounding, and dialogue-act annotation schemes: https://web.stanford.edu/~jurafsky/slp3/26.pdf

Previous: Diagonal linear discriminant analysis Next: Diameter