User Utterance
Definition
In conversational AI, an utterance is any single message produced by the user — a question, command, statement, or fragment. The same intent can generate vastly different utterances: 'refund my purchase', 'I want my money back', 'how do I return this?', and 'I'm not happy with this product' might all map to the same 'refund_request' intent. Collecting a diverse range of utterances per intent and using them as training data is fundamental to building an NLU model that generalizes well to real user input. Utterance variety reflects the natural diversity of human language.
Why It Matters
The quality and diversity of training utterances directly determines how well a chatbot handles real user input. Users never phrase things exactly as developers expect. A bot trained on a narrow set of utterances will fail on paraphrases, typos, slang, or indirect expressions. Regularly expanding the utterance set with examples from real conversations is one of the highest-leverage improvements a chatbot team can make.
How It Works
Each intent in the NLU model is associated with a set of training utterances — example inputs labeled with the correct intent. The model learns to map new, unseen utterances to intents by generalizing from these examples. Modern transformer-based NLU models can generalize effectively from 10-20 utterances per intent. LLM-based systems can handle zero-shot intent matching, where no labeled utterances are needed.
Real-World Example
A chatbot for a software product is trained with these utterances for 'password_reset': 'I forgot my password', 'how do I reset my login?', 'can't log in', 'locked out of my account', 'reset password please'. When a user types 'I can't get in to my account' — a phrase not in the training set — the model correctly maps it to 'password_reset' by generalizing from the examples.
Common Mistakes
- ✕Training with too few utterances per intent (fewer than 5-10), leading to poor generalization.
- ✕Using only formal, 'perfect' utterances in training and missing the typos, slang, and fragments of real user input.
- ✕Not reviewing actual conversation logs to harvest new utterances that users produce in production.
Related Terms
Intent Recognition
Intent recognition is the process by which a chatbot identifies the goal or purpose behind a user's message. It classifies free-form user input into predefined categories (intents) — such as 'check order status', 'request refund', or 'get pricing' — enabling the bot to route the conversation appropriately.
Entity Extraction
Entity extraction is the process of identifying and pulling specific pieces of information from a user's message — such as names, dates, order numbers, or locations. These extracted values (entities) fill in the details the chatbot needs to complete a task, working alongside intent recognition to fully understand the user's request.
Natural Language Understanding (NLU)
Natural Language Understanding (NLU) is the AI capability that interprets the meaning behind human text or speech — identifying what the user wants (intent) and extracting key details (entities). NLU is the 'comprehension' layer of a chatbot, translating raw input into structured information the system can act on.
Chatbot Training
Chatbot training is the process of teaching a chatbot to understand user intent, recognize entities, and respond appropriately — using labeled conversation data, example utterances, and feedback loops to improve accuracy over time. It encompasses both initial model training and ongoing improvement based on production data.
Fallback Response
A fallback response is what a chatbot says when it cannot understand the user's message or find an appropriate answer. Instead of returning an error or going silent, the bot delivers a graceful fallback — acknowledging the limitation and offering alternatives like rephrasing, browsing the FAQ, or speaking to a human agent.
Ready to build your AI chatbot?
Put these concepts into practice with 99helpers — no code required.
Start free trial →