datatotext

Data-to-text, also known as data-to-text generation (D2T), is a branch of natural language generation that automatically converts structured data into natural language text. It enables computers to describe data-driven insights in readable prose, reports, or narratives. Data can come from relational databases, spreadsheets, JSON, XML, or semantic graphs.

Typical architectures range from template-based systems that map data fields to fixed sentences, to rule-based or statistical approaches, to end-to-end neural models that generate text directly from raw data. Most systems use a pipeline with components for data input, content selection, discourse planning, and linguistic realization. Some also include post-editing or quality control stages.

Common applications include weather and financial reports, business intelligence summaries, sports recaps, product descriptions, and medical or weather risk summaries. In accessibility contexts, data-to-text can generate summaries for screen readers.

Evaluation usually combines automatic metrics such as BLEU or ROUGE with human judgments focusing on accuracy, completeness, and readability. Challenges include ensuring factual fidelity to the data, avoiding hallucinations, controlling style and tone, and adapting to multilingual generation.

The field emerged from rule-based natural language generation in the late 20th century and has evolved toward neural and hybrid methods since the 2010s. Researchers study content selection, factuality, data integrity, and controllable generation to improve reliability.