Key Takeaways
- An utterance is the raw thing a user says or types to a chatbot in one turn - the input before the bot interprets it.
- Utterance, intent, and entity are three layers of one message: the raw phrase, the goal behind it, and the key data extracted from it.
- Chatbots learn to recognize intents from example training utterances, and the best examples come from real chat logs rather than imagined phrasings.
- The most reliable way to improve a bot is a tight loop of collecting real utterances, reviewing the ones it mishandled, and folding them back into training.
What Is an Utterance?
An utterance is a single thing a user says or types to a chatbot or voice assistant in one turn of a conversation. It is the raw, unedited input - exactly as the person expressed it. "Where is my order?", "track package 12345", and "i never got my stuff" are three different utterances that a bot might receive.
The term comes from linguistics, where an utterance is simply a unit of speech. In chatbots it means the same thing: whatever the user submitted before the software tries to interpret it. That interpretation - figuring out what the utterance means and what to do about it - is a separate step handled by natural language processing.
Why Utterances Matter
Utterances are the ground truth of every conversation. They are what your users actually say, in their own words, with typos, slang, and shorthand. A chatbot is only as good as its ability to handle the real range of utterances people throw at it. Studying real utterances - not the tidy phrases a team imagines - is the fastest way to find where a bot is failing and how to improve it.
Utterance vs Intent vs Entity
These three terms are the core vocabulary of chatbot understanding, and they are easy to mix up. They actually describe three different layers of a single message.
- Utterance - the raw text or speech the user submitted. Example: "Can I return the blue shoes I bought last week?"
- Intent - the goal behind the utterance, which the bot classifies. Here the intent is return_item. Recognizing it is the job of intent recognition.
- Entity - the specific data points inside the utterance. Here "blue shoes" is a product entity and "last week" is a date entity, pulled out by entity extraction.
An Analogy
Think of the utterance as a full sentence someone hands you, the intent as the reason they wrote it, and the entities as the underlined key words within it. One utterance carries exactly one raw form, usually one primary intent, and zero or more entities.
| Layer | What It Is | Example |
|---|---|---|
| Utterance | The raw input | "return the blue shoes from last week" |
| Intent | The user's goal | return_item |
| Entity | Key data extracted | product: blue shoes, date: last week |
How Utterances Train a Chatbot
To recognize an intent, a chatbot needs to see many example utterances that express it. These are called training utterances or training phrases, and they teach the model the variety of ways people ask for the same thing.
Building Variety
For a single track_order intent, you might supply utterances like:
- "where is my order"
- "track my package"
- "has my order shipped yet"
- "i want to check delivery status"
- "when will my stuff arrive"
The wider and more realistic the range, the better the bot generalizes to phrasings it has never seen before. Weak coverage is the most common cause of a bot that keeps misunderstanding users.
Where to Get Real Utterances
The best training examples come from real conversations, not guesses. Support tickets, live chat logs, search queries, and chatbot analytics transcripts all reveal how customers genuinely phrase their needs. Feeding these real utterances back into training is a core loop of conversation design.
How Utterances Are Handled on a Chatbot Platform
On a chatbot platform, every utterance flows through a short pipeline before the bot decides how to respond. Understanding it helps you see where to intervene when things go wrong.
1. Capture
The user's raw utterance arrives from whatever channel they are on - a website widget, WhatsApp, or Messenger. It is logged exactly as received.
2. Classify
The natural language understanding model compares the utterance against trained intents and assigns the best match along with a confidence score.
3. Extract
Any entities in the utterance are pulled out so the bot has the specific details it needs, such as an order number or a date.
4. Respond or Fall Back
If confidence is high, the flow proceeds; if it is low, the bot triggers a fallback and asks the user to rephrase. In a platform like Conferbot, you add example utterances to each intent in a visual builder, and the tool surfaces unmatched utterances so you can fold them back into training. Over time this steadily widens the range of phrasings your bot handles.
What Makes a Good Set of Training Utterances
Not all training data is equally useful. A thoughtful set of utterances covers the real spread of how people communicate.
Cover the Full Range
- Formal and casual - "I would like to check my balance" and "balance?" both need to work.
- Different lengths - single keywords, full sentences, and rambling multi-part messages.
- Typos and slang - real users misspell words and use shorthand.
- Synonyms - "refund", "money back", and "return my payment" may share an intent.
Avoid Overlap Between Intents
If two intents share very similar utterances, the model gets confused and misclassifies both. Keep each intent's examples distinct, and split or merge intents when their training phrases start to blur together.
Keep It Balanced
Giving one intent hundreds of utterances and another only a handful biases the model. Aim for reasonably even coverage across the intents you care about, and keep adding real examples as new phrasings appear in your logs.
Utterances in Real Conversations
Seeing how utterances behave across different situations makes the concept concrete.
A Simple, Clear Utterance
"Reset my password" is short, unambiguous, and maps cleanly to a single intent with no entities. Most bots handle these well.
A Rich Utterance
"I ordered a large red jacket on Monday and want to change it to a medium" is one utterance carrying an intent (modify_order) plus several entities - size, color, product, and date. Good slot filling pulls all of these out at once so the bot does not have to ask again.
An Ambiguous Utterance
"It's not working" has no clear intent or entities on its own. A well-designed bot responds to this vague utterance by asking a clarifying question rather than guessing.
A Multi-Intent Utterance
"Cancel my order and update my email" packs two goals into one message. Advanced bots detect both; simpler ones handle the first and prompt for the second. Reviewing these tricky utterances in your logs is where the biggest accuracy gains usually hide.
Common Mistakes With Utterances
Teams new to chatbots tend to trip over the same issues when working with utterances. Knowing them in advance saves a lot of rework.
| Mistake | Consequence | Fix |
|---|---|---|
| Too few training utterances | Bot fails to understand common phrasings | Add real examples from chat logs |
| Only tidy, formal phrasings | Bot breaks on casual real-world input | Include slang, typos, and short forms |
| Overlapping intents | Frequent misclassification | Make each intent's examples distinct |
| Ignoring low-confidence utterances | Recurring failures go unfixed | Review unmatched utterances regularly |
| Never updating training data | Accuracy drifts as language changes | Fold new utterances in continuously |
The Underlying Lesson
Almost every utterance problem traces back to a gap between the phrases a team imagined and the phrases users actually send. The remedy is the same in every case: look at real utterances often, and keep feeding them into training. Building a chatbot in a no-code platform makes this loop fast because unmatched utterances are surfaced automatically.
Why Utterances Are the Foundation of Chatbot Quality
Everything a chatbot does begins with an utterance. The model cannot classify an intent, extract an entity, or fill a slot until a user has said something. That makes the range and quality of utterances your bot understands the single biggest driver of how competent it feels.
The Feedback Loop
The most reliable path to a better bot is a tight loop: collect real utterances, review the ones the bot mishandled, add them to training, and repeat. This is less about clever algorithms and more about disciplined attention to what users actually type. Teams that review their unmatched utterances weekly tend to see steady improvement; teams that set training data once and forget it tend to see accuracy quietly decay.
Utterances and the Bigger Picture
Utterances feed directly into intent recognition and entity extraction, which in turn drive the dialog flow and any slot filling the bot performs. Get the utterance layer right and everything above it becomes easier. Neglect it, and no amount of downstream design can compensate.