A user says, “I can’t get into my profile,” but a chatbot responds as though an item was not received. The classifier has selected the wrong intent—the predefined goal that should guide the conversation. In a conversational assistant, a text classifier turns a user’s utterance into an intent such as login help or flight booking. That prediction determines which answer, dialogue flow, or next question the user sees. section 1.1.3 section 5.2.1 section 1.1.2
From request to next step
“I want to book a flight” can select a booking intent and prompt for a destination. “I need a one-way ticket to Costa Rica” can select the same intent while supplying Costa Rica as an entity, allowing the assistant to ask for departure details instead. An intent captures the goal; an entity refines the request. section 4.1.3
Table 4.1 Example utterances may be handled differently based on the presence or absence of entities.
|
Utterance
|
Intent
|
Entity
|
Possible next step
|
|---|---|---|---|
|
“How many bags can I check?”
|
Bag_Allowance
|
Display bag check policy
|
|
|
“I want to book a flight”
|
Book_Flight
|
Collect destination
|
|
|
“I need a one-way ticket to Costa Rica”
|
Book_Flight
|
Costa Rica
|
Collect departure details
|
|
“I’d like to upgrade my seat to first class”
|
Flight_Upgrade
|
first class
|
Initiate upgrade process
|
Compare the utterances, predicted intents, optional entities, and possible next steps. section 4.1.3
A traditional classifier learns from example requests grouped by intent. Its scope is deliberately limited: the team must decide which topics the assistant should recognize. Trying to cover every rare topic can blur distinctions between intents, while omitting requests users commonly make creates a different failure: the assistant seems not to understand them. Even for an untrained topic, a classifier may choose a familiar intent because some words resemble its training examples. section 4.1.3 section 4.5.1 section 5.3
Improve the classifier against real demand rather than adding near-duplicate phrases at random:
- Review chat logs, call transcripts, or emails for the ways users actually express each goal; favor varied, representative examples over many slight rewrites. section 1.1.3 section 7.1.3
- Test on utterances kept separate from training data. Inspect accuracy, recall, precision, and F1 score, then prioritize popular intents that perform poorly. section 7.1.3 section 5.5
- Use annotated logs to decide whether a missing topic warrants a new intent. For requests outside the classifier’s scope, plan an honest fallback, search, or agent escalation. section 5.3.2 section 4.5.1
Repairing missed login requests
A login-help intent trained on just four similar examples missed requests about password resets, locked accounts, and security codes. Adding representative phrases from user logs raised its recall on a blind set from about 44% to 89%, while precision remained at 100% in that test. The lesson is not that every added phrase helps: add meaningful variety, then retest for effects on other intents. section 5.2.1 section 5.3.2
A classifier need not answer every request itself. It can route a recognized recipe request to a large language model, or distinguish a service request the assistant can handle from one requiring a specialist. Purpose-built classification offers fast, controlled routing; an LLM can handle broader tasks but brings cost, speed, and output-control risks. Keep monitoring predictions and provide a useful path when recognition fails. section 2.3 section 7.1 section 11.4.2
