What is Intent Classification?
Intent classification is a natural language processing task that assigns an incoming message to one of a predefined set of purposes, such as booking, canceling, or requesting a refund. It converts free text into a routing decision that downstream logic can act on. The label set is designed for a specific application, and unfamiliar requests must be handled explicitly.
It is a text classification problem, usually single label but sometimes multi label when a message contains more than one request. Intents are typically paired with slots, meaning the parameters a request needs, so a booking intent is accompanied by extracted dates, quantities, and locations. Together the intent and its slots form a structured command that software can execute.
Designing the label set is the hard part. Intents that overlap in meaning produce inconsistent training data and unstable predictions, while an overly fine set becomes unmaintainable. A fallback or unknown class is essential, because a classifier restricted to known labels will otherwise assign every out of scope message to whichever known intent happens to be nearest.
Implementations range from keyword rules, through classifiers trained on labeled utterances or on sentence embeddings with nearest neighbor lookup, to prompting a general purpose language model with intent descriptions. The last approach removes the need for many labeled examples and makes adding an intent cheap, at the cost of less predictable behavior on edge cases.
Confidence handling determines the user experience. A high confidence prediction can trigger an automated action, a middling one can prompt a clarifying question, and a low one should escalate to a person or to a general assistant. Monitoring the fallback rate over time reveals both new user needs and gradual degradation as language usage drifts.
Key points
- Maps a message to a predefined purpose label.
- Usually paired with slot extraction for parameters.
- Needs a fallback class for out of scope requests.
- Overlapping intent definitions cause unstable predictions.
- Confidence thresholds decide automate, clarify, or escalate.
In practice
A messaging assistant receives the request to move an appointment to Thursday. The classifier returns the reschedule intent with high confidence, and slot extraction returns the new day. The system looks up the existing booking and proposes times. Had the message asked why a card was charged twice, the same classifier would return a billing intent and route it elsewhere.