Natural Language Generation is the computational process of producing natural-language text or speech from structured, semi-structured, or otherwise non-linguistic information. It is commonly abbreviated NLG and is generally considered a subfield of natural language processing concerned with the generation side of language technology, in contrast with systems that primarily analyse or interpret language.
An NLG system may receive database records, numerical measurements, logical representations, extracted facts, application state, user preferences, or another machine-readable representation and transform that information into language intended for human consumption. The generated output can range from a short notification such as a weather alert to a multi-paragraph report, product description, financial summary, conversational response, or automatically composed document. The central challenge is not simply selecting grammatically valid words, but deciding what information should be communicated, organising it coherently, choosing suitable wording and level of detail, and producing an output appropriate for its audience and purpose.
Early Natural Language Generation systems were predominantly rule-based and template-driven. In data-to-text applications, developers explicitly specified templates, grammatical rules, lexical choices and content-selection conditions, allowing systems to generate predictable descriptions from structured data. Research subsequently introduced more sophisticated planning approaches involving content determination, document structuring, lexicalisation, referring-expression generation, aggregation and realisation. Statistical methods later enabled systems to learn generation patterns from corpora rather than relying exclusively on manually authored rules.
The development of neural language modelling substantially changed NLG by enabling models to learn complex relationships between linguistic context and likely continuations. Sequence-to-sequence architectures, attention mechanisms, transformer architectures and large-scale pretrained language models progressively increased the fluency, flexibility and breadth of generated language. Modern NLG therefore encompasses both specialised generation pipelines and general-purpose language models, although the underlying objective remains the transformation of information or context into useful natural-language output.
A traditional NLG pipeline can be described through several conceptual stages. Content determination decides which available facts are relevant; document planning establishes the structure and ordering of those facts; microplanning determines lexical choices, referring expressions and sentence aggregation; and surface realisation converts the planned representation into grammatical language.
For example, a system generating a football report might select the final score and significant events, determine that the score should be introduced before individual events, choose verbs appropriate to those events, combine related facts into sentences and finally realise the text. Neural systems can perform many of these functions implicitly rather than exposing each stage as a separate module. This can make development more flexible but can also make the generation process harder to inspect and control. In high-stakes or highly regulated environments, explicit planning, constrained decoding, templates or retrieval-based grounding may therefore remain useful even when a neural model is used for language production.
Modern NLG frequently uses large language models based on transformer architectures. Such models are trained on large collections of text to predict tokens or otherwise learn statistical representations of language, and they can subsequently generate text conditioned on prompts, structured inputs, retrieved information, examples or system instructions. Transformer self-attention enables a model to represent relationships between tokens across substantial contexts, while autoregressive decoding commonly generates output incrementally by selecting successive tokens according to a probability distribution.
Fine-tuning, instruction tuning, reinforcement-learning-based optimisation and other adaptation techniques can make a model more suitable for particular tasks or communication styles. Retrieval-augmented generation can supply externally retrieved information to the generation process, reducing reliance on information stored solely in model parameters. Tool use can similarly allow a generation system to obtain calculations, database results or current information before producing its response. These techniques demonstrate that contemporary NLG is often part of a larger information-processing architecture rather than an isolated text-generation component.
NLG has applications across many domains because structured information is abundant but is often difficult for people to interpret efficiently. Automated financial reporting can turn numerical performance data into narrative reports; weather systems can convert meteorological measurements into forecasts and warnings; business-intelligence platforms can describe trends and anomalies; healthcare applications can generate summaries from authorised clinical information; e-commerce systems can produce product descriptions; accessibility technologies can transform structured content into natural-language explanations; and software applications can generate notifications, instructions and personalised messages.
In customer-facing systems, NLG can produce responses that adapt wording to the user's request and conversational context. In data-to-text applications, the primary advantage can be the ability to communicate large quantities of structured information without requiring a human author to manually write every instance. However, generated language remains dependent on the quality, completeness and interpretation of its underlying information, so fluent wording does not guarantee factual correctness.
An important distinction exists between language generation and simple text templating. A template system normally selects predefined phrases and inserts values into predetermined positions, whereas NLG can involve dynamically determining content, structure and wording. Another distinction is between generation and summarisation: summarisation produces a shorter representation of existing information, while NLG is a broader concept that can produce language from structured data, instructions, internal representations or newly composed material.
Machine translation generates text in another language but has translation as its defining objective, while general NLG can operate entirely within one language. Dialogue systems use NLG to formulate responses, but a dialogue system also requires components for interpreting input, managing conversational state and selecting or executing actions. These distinctions matter because the term NLG describes the language-production component or capability, not necessarily an entire intelligent system.
Evaluation of NLG is difficult because multiple outputs can communicate the same underlying information while differing in wording and structure. Traditional automatic metrics such as BLEU, ROUGE and METEOR were developed for particular generation or comparison settings, but their correlation with human judgements can be limited when many valid outputs are possible. Evaluation can instead examine factual accuracy, relevance, completeness, grammaticality, coherence, fluency, diversity, consistency with source data, appropriateness of style and usefulness to the intended audience.
For data-to-text systems, factual fidelity is particularly important because an elegantly written sentence that misstates an underlying number is still defective. Neural and large-language-model systems introduce additional concerns including hallucination, unsupported claims, repetition, instruction-following failures, sensitivity to prompts and inconsistent outputs. Reliable evaluation therefore normally requires a combination of automated tests, structured datasets, human assessment and domain-specific verification rather than relying solely on linguistic similarity to a reference text.
The principal strengths of Natural Language Generation are scalability, speed, consistency, personalisation and the ability to make machine-readable information more accessible to people. It can generate thousands of individually tailored reports, notifications or descriptions that would be impractical to produce manually. Its limitations include factual errors, ambiguity, inappropriate wording, bias inherited from training data, difficulty guaranteeing deterministic behaviour and the possibility that fluent language will conceal incorrect reasoning or source information.
Systems can also produce unnecessary verbosity, omit important facts or select wording that is technically grammatical but unsuitable for a particular audience. Effective NLG therefore depends on careful input validation, appropriate grounding, output constraints, domain-specific terminology, evaluation procedures and human oversight where consequences are significant. Its long-term significance lies in providing a general interface between computational representations and human language, allowing software to communicate complex information in forms that people can read, hear and act upon.
