A customer says, “I can’t log in,” but the chatbot only recognizes requests that say “reset my password.” The customer gets a generic reply instead of help. To avoid that failure, a chatbot needs to recognize the different ways people express the same need. section 1.1.3 section 7.2
An utterance is the input a user gives a chatbot: what the user says, not necessarily what they mean. An intent represents what the user wants. In an intent-based chatbot, a classifier groups utterances with similar meanings under an intent, which can then lead to a response. section 1.1.3 section 2.1.2
Different words, same request
“What time are you open?” and “What are your hours?” are two utterances that can share a store-hours intent and receive the same answer. “What cakes do you sell?” calls for a different intent and response. The goal is to distinguish needs, not every variation in wording. section 2.1.2
The utterance has several practical roles:
- Training: supply varied example utterances for each intent, rather than relying on one expected phrase. Historical chats, call transcripts, and emails can reveal how users actually ask. section 1.1.3 section 2.1.2
- Routing: recognize a request from example wording, but retain the whole utterance when its details matter. A recipe request can be passed into an LLM prompt without first enumerating every possible dish. section 2.3.2
- Evaluation: compare annotated utterances in logs with training examples and inspect high-volume intents or requests outside the current scope. A mismatch suggests where to investigate; it does not prove poor accuracy by itself. section 4.5.1
When an intent misses common wording, add representative variations and test them against real user utterances. An LLM can suggest synonyms or grammatical variations and offer a quick check for confusing intent groups, but its suggested labels need human judgment. Keep a representative test set so changes can be measured rather than guessed. section 4.6 section 7.2 section 7.3.2
