A dialogue act is a label for what a single utterance does in a conversation. “Hello” greets, “Can you move my booking?” requests, “Yes, that’s right” confirms. Chatbot designers use these labels to plan which moves a bot must understand and which it must make, instead of treating every message as a bag of keywords.
In linguistics and natural language understanding, a dialog act is described as a unit of conversation that is more fine-grained than an utterance, with a specific communicative function, and questions, statements and requests for action are standard examples. That definition, summarised on Wikipedia’s dialog act page, is a useful working one.
Where the idea comes from
The notion grows out of speech act theory. The philosopher J. L. Austin, in How to Do Things with Words (published posthumously in 1962), argued that speaking is a way of acting as well as describing. His example: asking “Is there any salt?” can carry the implied request “please pass the salt”. John Searle developed the theory further in Speech Acts (1969).
Austin separated three levels of action in an utterance:
- the locutionary act, the words and their literal meaning
- the illocutionary act, what the speaker is trying to do with them
- the perlocutionary act, the effect on the listener
Dialogue act labels are mainly concerned with the second level: what the speaker is trying to do.
Searle’s broad families
Searle proposed in 1975 that illocutionary acts fall into five broad classes. As summarised on Wikipedia’s illocutionary act page, they are assertives (committing to the truth of a statement), directives (trying to get someone to act, such as requests), commissives (committing to a future action, such as promises), expressives (conveying an attitude, such as thanks or apology) and declarations (changing a state of affairs by saying so).
Common acts in a chatbot conversation
Real taxonomies are more detailed than five classes. One study described on Wikipedia’s dialog act page used 42 tags, including statement, opinion, agreement and yes-no question. A bot rarely needs that many. A small working set is easier to manage:
| Act | What the user is doing | Example | Typical bot response |
|---|---|---|---|
| Greeting | Opens the exchange | “Hi there” | Greeting plus a prompt for what the bot can do |
| Request | Asks for an action or information | “I need to change my delivery date” | Gather details, then act or answer |
| Inform | Supplies a detail | “It’s the fourteenth” | Store it, move on |
| Confirm | Agrees | “Yes, go ahead” | Proceed with the action |
| Deny | Rejects or corrects | “No, the other address” | Repair and re-ask |
| Thanks | Closes politely | “Thanks, that’s all” | Acknowledge and end |
Adapt the list to your domain; the aim is a shared vocabulary for writers, designers and engineers.
The bot has acts too
A well-designed bot has its own repertoire: greet, ask, offer options, confirm, apologise, hand over. Planning these helps avoid common faults.
- Stacking acts. A reply that confirms, asks two new questions and offers a menu is hard to answer. Aim for one main act per turn.
- Missing a confirmation. Before an irreversible step, such as cancelling a service, the bot should repeat the key details and ask the user to confirm. The principles of conversation design apply here.
Indirect acts
Form and function can differ. “Could you tell me the opening time?” is grammatically a question about ability, but nearly everyone treats it as a request. Test your bot on indirect phrasings such as polite requests and hints, and include many forms of each act in training data.
How acts are recognised
The task of labelling an utterance with its act is called dialogue act recognition. According to Wikipedia, it is part of spoken language understanding and uses statistical or AI models to classify utterances. In an intent-based chatbot, much of the same work happens under another name: “intents” are close cousins of acts, often combined with a topic, such as “request: change delivery date”. The page on rule-based and AI chatbots explains how different systems carry out the recognition step.
Frequently asked questions
Is a dialogue act the same as an intent?
Not exactly. A dialogue act describes the communicative function, such as requesting or confirming. An intent usually also names the subject, such as “check order status”. Many systems combine the two.
Do language-model chatbots still need dialogue acts?
They may not need them to function, but acts are a practical tool for designing, reviewing and testing conversations.
A small exercise
Take ten real transcripts and label every user message with one act from the table above. Count which acts are most common and where the bot’s reply used the wrong act. The gaps point to the next improvements.