Glossary
›
Natural language understanding (NLU)

Natural language understanding (NLU)

Glossary
Also known as:
NLU, language understanding, natural language interpretation

What is natural language understanding?

Natural language understanding (NLU) is the branch of AI that works out what a person means from what they say or write: the request behind the words, the details it contains, what earlier statements are being referred to and the tone. It is the comprehension side of natural language processing, the counterpart to generating language.

Natural language processing is the umbrella field. Within it, understanding handles what comes in and generation handles what goes out. For years, understanding was built from separate parts, an intent classifier, an entity extractor and hand-written rules for context. Large language models now perform much of that work in a single step, but the underlying tasks have not changed, and neither has the need to test each one.

How natural language understanding works

Understanding a customer message involves several tasks at once.

TaskQuestion it answersExample
IntentWhat does the person want?"Is it too late to change my order?" means change an order
EntitiesWhich details matter?The order number, the item and the new size
ReferenceWhat do words such as "it" and "that one" point to?"That one" is the jacket mentioned two turns earlier
ContextWhat has already been established?The customer's identity was verified earlier in the call
Sentiment and urgencyHow does the person feel, and how pressing is the matter?Frustration after a second failed delivery
AmbiguityIs more than one reading possible?"Cancel it" when two orders are open, so ask which

Good understanding includes knowing when it has not understood. A system that asks a short clarifying question at the right moment is more useful than one that guesses fluently and gets it wrong.

Why natural language understanding matters

Benchmarks of language understanding have been overtaken quickly. When the SuperGLUE benchmark was published in 2019, its authors estimated human performance at 89.8, with the best baseline model they tested nearly 20 points behind. Less than two years later, Microsoft researchers reported a model scoring 90.3, the first to pass the human baseline.

Benchmark scores and real conversations are different tests. Customers use product names, abbreviations and local expressions that general benchmarks do not contain; they correct themselves, ask two things at once and refer back to something said minutes earlier. On a phone line, understanding also has to cope with recognition errors from the audio. An organization's own conversations are the only reliable measure of how well a system understands its customers.

Where natural language understanding is used

  • Routing and self-service. Interpreting what a caller or chat user needs and starting the right process.
  • Search. Understanding the meaning of a query rather than matching its exact words.
  • Document understanding. Reading letters, emails and forms to find the request and the key details.
  • Analytics. Measuring contact reasons, sentiment and emerging issues across large volumes of conversations.
  • AI agents. Interpreting each customer turn before the agent decides its next step.

Understanding interprets; it does not decide or act. And where a choice is simple and structured, such as picking one of three plans, a button or a short list is clearer and less error-prone than free text that must be interpreted.

What good natural language understanding looks like

  • Domain vocabulary. Product names, internal terms and industry jargon are recognized correctly.
  • Language coverage. It handles the languages customers actually use, including mixing two languages in one message.
  • Clarification when unsure. It asks rather than guesses when two readings are plausible.
  • Use of conversation history. Earlier turns inform how each new message is read.
  • Robustness to errors. Typos and speech recognition mistakes do not derail interpretation.
  • Calibrated confidence. The system's certainty matches its actual accuracy.

How to measure natural language understanding

  • Intent accuracy. The share of messages assigned the correct intent in a labeled sample.
  • Entity accuracy. Precision and recall for each type of detail extracted.
  • Clarification rate. How often the system needs to ask, and whether those questions were necessary.
  • Misunderstanding rate. Conversations where an audit finds the system acted on a wrong reading.
  • Accuracy by language, dialect and channel. Gaps between groups show where understanding is weakest.

Language and disclosure rules

Language choices carry obligations once a business negotiates or discloses in a language other than English. The main US points follow; other countries add rules of their own.

  • Contracts negotiated in other languages. California Civil Code 1632 requires businesses that negotiate certain consumer contracts, including many loans and leases, primarily in Spanish, Chinese, Tagalog, Vietnamese or Korean, whether orally or in writing, to provide a translation in that language before signing.
  • Credit disclosures in other languages. Regulation Z allows disclosures in a language other than English, provided English versions are available on request.
  • Emotion recognition at work, outside the US. Since February 2, 2025, the EU AI Act has prohibited AI systems that infer the emotions of people in workplaces and education, except for medical or safety reasons. Sentiment analysis of employees in EU operations falls within that limit.

Frequently asked questions

What is the difference between NLU and NLP?

Natural language processing is the whole field of computing with human language. Natural language understanding is the part focused on interpreting meaning.

What is the difference between NLU and NLG?

Understanding interprets what comes in; natural language generation produces what goes out. A system that talks with customers needs both.

Have large language models made NLU obsolete?

They have absorbed many separate NLU components into one model. The tasks remain, and so does the need to measure intent accuracy, entity accuracy and context handling on real conversations.

Why does NLU still fail?

Mostly on what is specific to a business: unfamiliar product names, ambiguous references, two requests in one message and errors carried over from speech recognition.

How is NLU tested?

With a labeled sample of real customer messages, scored for intents, entities and context, and broken down by channel, language and customer group.

Understanding is where the conversation starts.

Encore AI's agents understand what each customer means and then do what your top performers would do next, following a playbook your compliance team approves.

Scale your best to be the best.

Book a Demo