Intent recognition and dialogue acts.
A spoken line goes in. What the person wants, the details they named, and the kind of turn it is come out. Here is the field that reads intent, in plain language.
Picture a live sales call. The buyer says, "Can you push our renewal to early March and loop in our finance lead?" A person hears three things at once: what is being asked, the specifics that fill it in, and the kind of move the line makes. Intent recognition is the part of language technology that teaches a system to read the same line and write it down as a small structured form. We walk it the way the field is built: what it is, why it is hard, how a modern system does it, what it makes possible, and where it is heading. The papers behind each idea are linked, so you can read the source yourself.
People speak in prose. Systems need a form.
Intent recognition turns a line of language into a small structured form: what they want, the details that fill it in, and what kind of turn it is.
Three pieces come off the same line, named in the order a reader can follow. First, what they want done. That is the intent: the goal behind the line, like reschedule_renewal or check_status. One label per request. Second, the details that fill it in. Those are the slots: the specific values the intent needs, a date, a name, an amount. Slot filling tags the words that carry them, so "early March" becomes a date and "finance lead" becomes a contact. Third, what kind of move the turn is making. That is the dialogue act: a statement, a question, a request, an agreement, a backchannel like "uh-huh." The same words can be a genuine question or a polite command depending on the act.
The result is not a paragraph. It is a fixed shape the next system can act on: update a field, draft a follow up, or surface the right next step while the call is still live. The rest of this page is about how a machine reads that, and why doing it well is harder than it sounds.
“Can you move our renewal review to Friday?”
- Intent
- Reschedule a meeting
- Meeting
- Renewal review
- Day
- Friday
- Dialogue act
- Request
People do not speak in clean commands.
A fluent spoken line is easy for a person and ambiguous for a program. The gap between "a person can tell what they meant" and "a program can tell" is the whole problem.
Intent and slots lean on each other.
Knowing the intent helps you read the slots, and the slots hint at the intent. A model that decides them separately can disagree with itself: the right intent with a slot that does not belong to it.
The move is not the words.
"Do you think you could send that over?" looks like a question and works as a request. "Sure" can be agreement or a flat non-answer. The dialogue act lives in context, not just the wording.
Meaning carries from earlier.
"Make it the same as last time" only resolves against what was said before. "Early next week" only resolves against when the call happened. A turn read in isolation loses meaning.
Speech arrives imperfect.
Disfluencies, restarts, filler, and recognition errors from the speech-to-text step upstream all land on the input. The intent system has to stay steady on messy audio, not clean typed text.
Speed is part of correctness.
On a live call, structure that arrives a minute late arrived too late to help the turn. Reading intent for a live setting means reading it while the conversation is still moving.
A wrong intent is worse than none.
Confidently filing a request as the wrong intent, canceling when the buyer asked to reschedule, is more dangerous than flagging "not sure," because a person downstream may act on it. A system reading an intent is not the same as the intent being right.
From a line of speech to a structured form.
Tag the turn for intent and slots, name the move, read the two together, then measure it all on shared data. The field grew from tagging words one at a time to reading a whole turn at once.
1One label for the request, a tag on each word
Tag the turn for intent and slots.
The classic framing reads the line two ways at once. The model reads the sentence once and asks two questions. Overall, what is being requested? That is one intent label for the whole turn. And for each word, is this a date, a city, a name, or nothing? That is a slot tag, read left to right with the surrounding words in view.
Spoken language understanding has used this intent-plus-slots frame for decades, going back to the air-travel query systems that the ATIS corpus (Hemphill et al., 1990) was built to measure. It is an early shared benchmark and still in use.
Book a flight to Boston on Friday morning.
- Intent
- book_flight
- Destination slot
- Boston
- Date / time slot
- Friday morning
2Name the move, not just the words
Read the dialogue act.
Dialogue act recognition labels each turn with its function: statement, question, request, agreement, backchannel, and so on. Stolcke et al. (2000) modeled the conversation as a sequence where each act depends on the ones before it, combining the words in a turn with the flow of the dialogue and even prosody, the rhythm and pitch of speech.
The idea that matters for a reader: the type of a turn is partly predictable from the turns around it. The same words, "Do you think you could send that over?", can be a real question or a polite request, and only the context decides which.
Information question
Can you send encrypted files?
Request
Can you send the revised quote?
3Decide what they want and the details together
Read intent and slots in one pass.
Because intent and slots inform each other, the field moved to joint models that decide both at once and let each improve the other. Goo et al. (2018) added a slot gate so the intent reading directly shapes the slot reading, reducing the case where the two disagree.
The everyday version: decide what they want and the details that go with it in one pass, not two. That keeps the structured form internally consistent, instead of pairing the right intent with a slot that belongs to a different one.
Move our review to Friday.
Intent prediction
reschedule_meetingSlot prediction
meeting: our review
date: FridayIdentify the meeting and the intended Friday before making a change.
4Read the whole turn at once
Label off a model that already knows language.
Newer methods hand the whole line to a model that already learned a lot of language in advance, then read intent and slots off that shared understanding. Chen, Zhuo, and Wang (2019) applied a pretrained language model jointly to intent classification and slot filling, a widely cited anchor for reading the whole turn at once, then labeling it. (BERT is named here as the subject of that paper, not as a description of any particular product.)
The honest note: these models are strong and flexible, and they can still confidently mislabel an unfamiliar request. That is exactly why a confidence check or a fallback to a clarifying question matters, so a best guess is not mistaken for a sure thing.
Move the finance review to the second week of April.
| Argument | Status |
|---|---|
| Meeting | Finance review |
| Date range | Second week of April |
| Exact day / time | Not specified |
Which day and time in that week should I use?
5Where the systems get measured
Measured on public, repeatable data.
Progress here is tied to shared datasets that everyone tests on. ATIS (Hemphill et al., 1990) is air-travel requests. SNIPS (Coucke et al., 2018) is everyday voice-assistant commands across several domains, released alongside a system the authors describe as built to run on small devices. Schema-Guided Dialogue (Rastogi et al., 2020) pushes toward many services and new intents.
Naming these matters because it shows intent and slots are measured against public, repeatable benchmarks, not vibes. Any accuracy or slot score you read belongs to the paper that reports it, an industry finding, never a number attributed to us.
| Metric | Question |
|---|---|
| Intent accuracy | Was the correct task label selected? |
| Slot F1 | Were the right detail spans and labels extracted? |
| Joint correctness | Were the task and all required details correct together? |
What reading intent makes possible.
At the field level: structure off a live turn, fewer dropped details, turns a system can route, and more kinds of request without a rebuild.
This family of techniques is how the industry turns a spoken request into an action a system can take. For a sales co-pilot, the bridge is at the outcome level. When a request becomes intent and slots as it is spoken, the next step is something a system can act on while the call is still live. The harder cases above, the vague date, two requests in one breath, a question that is really a request, are exactly the ones that decide whether a live call turns into a usable next step.
A spoken request becomes a next step.
A request made out loud on a call can become a next step without waiting for someone to type it up. See Live deal assist.
Fresh detail reaches the app fast.
A date, a name, a next step named out loud can land in the right field while the conversation is still live, instead of being retyped after the call.
Turns you can route.
Knowing a turn is a question versus a request versus a backchannel lets a system decide whether to answer, act, or simply listen. That is the difference between a useful assistant and a noisy one.
More kinds of ask, no rebuild.
Shared benchmarks and pretrained models mean adding a new intent does not start from zero, so the service can keep up as the kinds of request grow. See After-meeting notes.
This explainer covers published research. To see how AI REV Voice helps with sales conversations, explore Prism.
The direction the field is heading.
We describe the direction, not a roadmap of our own.
Scaling to many domains.
The frontier is handling an ever-growing list of services, including intents with little or no training data. The Schema-Guided Dialogue work (Rastogi et al., 2020) describes each service with a schema so a model can handle new intents it was not trained on, rather than retraining for every new ask.
Intent, slots, and act together.
Large pretrained models increasingly read intent, slots, and dialogue act together, and connect them to the response, rather than a pipeline of separate parts. Flexible, and it raises the stakes on knowing when the model is unsure.
Knowing when not to act.
The hard part is not labeling more, it is knowing when to hold back. Confidence, fallback to a clarifying question, and grounding the read back to the words that justify it are active work. A trustworthy read can point at the part of the turn it came from.
Steady on messy input.
Staying accurate on disfluent, error-prone, real-world speech rather than clean text remains a live research area, especially for live settings where there is no second pass.
The papers behind this page.
- Hemphill, Godfrey, and Doddington (1990). The ATIS Spoken Language Systems Pilot Corpus. ACL Anthology H90-1021.
- Stolcke et al. (2000). Dialogue Act Modeling for Automatic Tagging and Recognition of Conversational Speech. Computational Linguistics, ACL Anthology J00-3003.
- Goo et al. (2018). Slot-Gated Modeling for Joint Slot Filling and Intent Prediction. ACL Anthology N18-2118.
- Coucke et al. (2018). Snips Voice Platform: an Embedded Spoken Language Understanding System for Private-by-Design Voice Interfaces. arXiv:1805.10190.
- Chen, Zhuo, and Wang (2019). BERT for Joint Intent Classification and Slot Filling. arXiv:1902.10909.
- Rastogi, Zang, Sunkara, Gupta, and Khaitan (2020). Towards Scalable Multi-Domain Conversational Agents: The Schema-Guided Dialogue Dataset. arXiv:1909.05855.
Keep reading.
Sibling explainers in the Research section.
Talk. Done.
That is the whole field in one promise: you talk, the request becomes a next step, and the app is already up to date. See it on a real call.
No app required. Just your phone number.