Specialty Digest

Global English editionLearners’ Study Dashboard — in development
The Knowledge Library for All

Natural Language Generation

nounNatural Language Processingalso NLG, automated text generation, data-to-text generation

In one line

Natural language generation (NLG) is the area of artificial intelligence concerned with producing readable human language from data, meaning representations or prompts.

Natural language generation (NLG) is the field of artificial intelligence concerned with producing written or spoken language that people can read or hear naturally. The input may be a table of numbers, a database record, a logical form or a short prompt, and the output is a sentence, paragraph or full document. NLG is the counterpart of natural language understanding: understanding turns language into structure, while generation turns structure back into language. Both belong to the wider discipline of natural language processing.

Overview

A weather service that writes “Expect light rain in the afternoon, clearing by evening” from a grid of forecast values is performing natural language generation. So is a financial tool that drafts a summary of quarterly results, or a chatbot that composes a reply. The shared challenge is to decide what to say, how to organize it and which words to use, while keeping the result accurate and fluent.

The term covers a wide range of systems, from simple fill-in-the-blank templates to large neural networks. Today the most visible examples are large language models, which perform generation as their core function; see also text generation.

The classical pipeline

Before neural methods dominated, researchers described NLG as a pipeline of stages, a view set out by Reiter and Dale [1].

  1. Content determination. Choosing which information from the input should be mentioned.
  2. Document structuring. Ordering the information and grouping it into paragraphs or sections.
  3. Aggregation. Combining related facts into single sentences to avoid repetition.
  4. Lexicalization. Selecting the words and phrases that express each concept.
  5. Referring expression generation. Deciding how to refer to entities, for example using a name the first time and a pronoun afterward.
  6. Surface realization. Producing grammatical sentences with correct word order, agreement and punctuation.

This modular design gave developers tight control over what a system said, which suited domains such as weather reports and medical summaries where accuracy mattered.

Main approaches

### Template-based generation

Templates insert values into fixed sentence patterns, for example “Sales in {region} rose by {percent} percent.” They are simple, predictable and easy to check, but they sound repetitive and need manual work for each new case.

### Rule-based and grammar-based generation

These systems apply linguistic rules and grammars to build sentences from a meaning representation. They offer more variety than templates and good control, but they take expert effort to build and do not transfer easily between domains.

### Statistical and neural generation

Statistical methods learn word patterns from example text. In 2014, sequence-to-sequence neural networks showed that a model could learn to map one sequence to another end to end, which made learned generation practical for tasks such as translation and summarization [2]. These models rely on recurrent and later attention-based architectures.

### Large language models

Transformer-based models trained on large text collections can generate fluent text from a prompt on almost any subject. They are pretrained, often fine-tuned and then guided through prompt engineering. Generation proceeds one token at a time, and settings such as temperature, top-k sampling and top-p sampling control how varied the output is.

Applications

  • Automated reporting. Turning data into financial, sports, weather or business reports.
  • Conversational systems. Producing replies in chatbots and virtual assistants.
  • Summarization. Condensing long documents, as in text summarization.
  • Machine translation. Generating text in a target language; see machine translation.
  • Personalized content. Writing product descriptions, emails or recommendations tailored to individual users.
  • Accessibility. Describing images or charts in words, linked to image captioning.
  • Voice output. Text from NLG can be passed to speech synthesis to be spoken aloud.

Evaluation

Evaluating generated text is hard because many different texts can be correct. Researchers use three broad approaches.

  • Automatic metrics. Measures such as BLEU compare generated text with one or more human reference texts by counting overlapping word sequences [3]. They are cheap and repeatable but correlate imperfectly with human judgments, especially for open-ended text.
  • Human evaluation. People rate fluency, accuracy, coherence and usefulness. This is more reliable but costly and slow.
  • Task-based evaluation. The text is judged by how well it helps users complete a task.

Factual accuracy deserves separate attention, since a fluent text can still contain mistakes.

Challenges and risks

  • Hallucination. Neural generators can state false things confidently, a problem described under hallucination. Techniques such as retrieval-augmented generation and grounding aim to reduce it.
  • Bias and toxicity. Models can reproduce stereotypes or harmful language found in their training data.
  • Repetition and incoherence. Long outputs may drift, repeat themselves or contradict earlier statements.
  • Attribution and misuse. Generated text can be used for spam, misinformation or academic dishonesty, which has led to interest in detection and disclosure practices.
  • Control. It can be hard to guarantee that a model follows every constraint on style, length or content.

History

Automatic text generation began in the 1950s and 1960s with experiments that produced sentences from grammars, and with early systems that composed simple stories or answered questions in restricted domains. Through the 1980s and 1990s, NLG became a research area of its own, with work on text planning, sentence planning and realization that led to the pipeline described above. Commercial use grew in the 2000s with systems that wrote weather forecasts and financial summaries from data feeds. The shift to neural methods in the 2010s, and then to transformer-based language models, changed the field by removing much of the hand-built pipeline and allowing one model to handle many generation tasks.

Controlling generation

Because modern generators are flexible but sometimes unpredictable, much current work concerns control. Developers shape output through instructions in the prompt, through structured output formats that force a model to return valid fields, and through system prompts that set tone and rules. Constrained decoding can require a model to include certain words or follow a grammar. For high-stakes uses, such as medical or legal summaries, teams add checks that compare the generated text with the source data and route doubtful outputs to a person for review, as in human-in-the-loop AI.

Choosing an approach

The right method depends on the task. When the input is structured, the wording is repetitive and errors are costly, templates or rule-based systems remain a sound choice because every output can be predicted and audited. When the text must be varied, conversational or cover open subject matter, a neural model is more suitable, usually with safeguards. Many production systems combine the two, using templates for fixed legal or numeric statements and a language model for the surrounding explanation.

FAQs

What is natural language generation? It is the AI task of producing human-readable text or speech from data, meaning representations or prompts.

How is NLG different from NLU? Natural language understanding interprets language and extracts meaning. NLG creates language from structured information or prompts.

Is ChatGPT an example of NLG? Yes. Chat assistants built on large language models generate text, so they are a modern example of natural language generation.

What are examples of NLG in daily life? Automated weather and sports reports, email suggestions, product descriptions, chatbot replies and image descriptions are common examples.

How is generated text evaluated? Through automatic metrics such as BLEU, human ratings of fluency and accuracy, and tests of whether the text helps users do a task.

Key facts

  • Reiter and Dale's 1997 paper described a standard pipeline for rule-based natural language generation with stages such as content determination and surface realization1.
  • Sequence-to-sequence neural models, introduced in 2014, made end-to-end learned text generation practical2.
  • BLEU, proposed in 2002, is a widely used automatic metric that compares generated text with human reference texts3.

Related concepts

Quick checkWhat is natural language generation?Show answer

Natural language generation (NLG) is the part of artificial intelligence that produces human-readable text from structured data, a meaning representation or a prompt, such as turning a table of sales figures into a written report.

Sources

  1. Ehud Reiter and Robert Dale, Building Applied Natural Language Generation Systems, Natural Language Engineering, https://doi.org/10.1017/S1351324997001502
  2. Ilya Sutskever, Oriol Vinyals and Quoc Le, Sequence to Sequence Learning with Neural Networks, https://arxiv.org/abs/1409.3215
  3. Kishore Papineni et al., BLEU: a Method for Automatic Evaluation of Machine Translation, https://aclanthology.org/P02-1040/

Editorially reviewed by Specialty Digest Editorial TeamLast reviewed November 17, 2026Researched and drafted with AI assistanceReport an issue