Natural Language Generation — Data-to-Text
Natural Language Generation (NLG) · Also known as: NLG, data-to-text, text generation, Doğal Dil Üretimi (NLG)
Natural Language Generation (NLG) is the branch of natural language processing that automatically produces fluent, human-readable text from structured data, knowledge graphs, or semantic representations. Formalised in the classical pipeline by Reiter and Dale (2000) and surveyed comprehensively by Gatt and Krahmer (2018), NLG powers applications ranging from automated financial reporting and weather bulletins to data storytelling and conversational agents.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
NLG is appropriate when you have structured data or a knowledge representation and need to communicate it as readable text at scale — for example, generating thousands of personalised reports, narrating dashboard data, or producing descriptions from a database. The input data structure must be well-defined. Quality evaluation requires either reference texts (for BLEU/BERTScore) or human raters. A minimum of around ten representative input-output pairs is needed to verify correctness; neural approaches require substantially larger training corpora. NLG is not suitable when the input is already unstructured text — that is a summarisation or paraphrasing task.
Strengths & limitations
- Converts large volumes of structured data into readable narrative automatically, enabling scalable reporting.
- Classical pipeline approaches offer full interpretability and precise control over what is said and how.
- Neural approaches learn complex input-to-text mappings from examples without hand-crafted grammar rules.
- Applicable across domains — finance, weather, healthcare, sports — wherever structured data needs narration.
- Neural models can hallucinate — generating plausible-sounding text that contradicts the source data.
- High-quality output requires either carefully engineered templates (classical) or large paired training corpora (neural).
- Automatic metrics such as BLEU correlate imperfectly with human judgements of quality; human evaluation is costly.
- Multilingual NLG remains demanding: surface realisation rules and morphological complexity vary substantially across languages.
Frequently asked
What is the difference between NLG and text summarisation?
Text summarisation takes existing text as input and produces a shorter text. NLG takes structured data — tables, databases, knowledge graphs — as input and produces text from scratch. The two overlap when the source is a structured representation of a document, but the canonical distinction is input type: unstructured text versus structured data.
How is the quality of generated text measured?
Automatic metrics such as BLEU (n-gram precision against reference texts) and BERTScore (contextual embedding similarity) are standard, but both correlate imperfectly with human judgements. For consequential applications, human evaluation of fluency, correctness, and relevance is recommended alongside automatic metrics.
When should I use a template-based system rather than a neural model?
Template-based systems are preferable when outputs must be strictly factually correct, when training data is scarce, or when full control over phrasing is required (e.g., regulatory documents). Neural models excel when large paired datasets are available and some variation in phrasing is acceptable.
What is hallucination in NLG and how can I detect it?
Hallucination occurs when a model generates text that is fluent but factually inconsistent with the source data. Detection approaches include comparing generated claims against the structured input using information-extraction tools, computing factual consistency scores, and human review. Hallucination is most common in neural models trained without explicit faithfulness constraints.
Sources
- Gatt, A. & Krahmer, E. (2018). Survey of the State of the Art in Natural Language Generation: Core Tasks, Applications and Evaluation. Journal of Artificial Intelligence Research, 61, 65-170. link ↗
- Reiter, E. & Dale, R. (2000). Building Natural Language Generation Systems. Cambridge University Press. ISBN: 9780521620369
How to cite this page
ScholarGate. (2026, June 1). Natural Language Generation (NLG). ScholarGate. https://scholargate.app/en/text-mining/natural-language-generation
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Automatic Text EvaluationText mining↔ compare
- GPT Fine-TuningDeep learning↔ compare
- Machine TranslationText mining↔ compare
- Retrieval-Augmented GenerationText mining↔ compare
- Sequence-to-Sequence ModelDeep learning↔ compare
- Text SummarizationText mining↔ compare
- TransformerDeep learning↔ compare