Trainings Phrases / Utterances

May 29, 2019
|By Daniel Rösch

–-> Go to BOTwiki

 
Utterances are training phrases used to AI Agent understands what users mean by comparing the examples stored in the system with the live query. This concept stems from traditional, intent-based dialogue architecture and has been a central element in the development of chatbots and virtual assistants for many years.

In modern agent-based systems based on large language models, the manual maintenance of utterances is becoming less and less important. These systems understand free-form text in context and can handle linguistic variation without having been trained in advance on specific example phrases. However, for those who work with intent-based NLU models or maintain existing systems of this type, the following section covers the essential basics.

Jede Utterance ist eine konkrete Beispielformulierung, die einem bestimmten Intent zugeordnet wird. Aus vielen solcher Beispiele lernt das NLU-Modell, sprachliche Variationen einer Anfrage zu erkennen. Je breiter und realistischer die Utterances, desto stabiler die Klassifikation in der Produktion.

 

What an utterance is and how it works

An utterance is a short example input formulated in natural language. For the "Book_appointment" “Book_Appointment,” these might include: “I need an appointment,” “Can you schedule an appointment for me?” or “I’d like to come in next week.”

The NLU model derives linguistic patterns from these examples. It then recognizes phrases that were not included verbatim in the training data. Utterances are therefore the training material that determines the quality of recognition.

 

Requirements for Good Utterances

For each intent, several utterances should be stored to cover different phrasings. The following are important: 

  • Linguistic variations: short and long phrases, with and without polite expressions.
  • Synonyms: “appointment,” “meeting,” “consultation.”
  • Realistic input from the target audience's perspective, not artificially constructed sentences.
  • Coverage of common typos, colloquial language, and dialects, as they occur in practice.

 It is equally important to ensure clear boundaries. Utterances with different intents must not overlap too much semantically. Otherwise, the confidence score will drop because the model can no longer make a clear decision.

 

Frequently Asked Questions (FAQ)

Utterances are training phrases that an AI agent uses to learn to recognize an intent. Each utterance is a specific example input, such as “I need an appointment” for the “Book_Appointment” intent. The NLU model derives linguistic patterns from many such examples, enabling it to recognize variations that were not trained on verbatim.

For Pro Intent, ten to thirty utterances are generally a good starting point. What matters is not the sheer quantity, but the linguistic variety. Short and long phrases, synonyms, and realistic typos improve recognition stability more effectively than additional similar sentences.

The Confidence Score indicates how confident the model is in assigning an intent. Well-formulated utterances increase the score because the model learns clear patterns. If utterances from different intents overlap, the score decreases and the system falls back on a fallback more frequently.

> Back to BOTwiki