Skip to content

Natural Language Generation

Natural language generation (NLG) encompasses techniques and systems for automatically producing human-readable text from non-linguistic or intermediate representations. NLG systems span diverse applications:

  • Abstractive summarization: Condensing documents into shorter summaries while preserving key information
  • Dialogue systems: Generating conversational responses in task-oriented and open-domain settings
  • Question answering: Synthesizing answers from retrieved documents or knowledge bases
  • Data-to-text generation: Converting structured data (tables, databases) into natural language descriptions
  • Machine translation: Converting text from one language to another
  • Image captioning and vision-language generation: Describing images in natural language

Modern NLG systems are built on neural architectures—particularly transformer-based sequence-to-sequence models like BERT, GPT, and T5—trained end-to-end on large corpora. These models achieve fluent, contextually appropriate outputs but are prone to systematic failures including hallucination, factual inconsistency, and copying errors.

Evaluation challenges

A core research area in NLG is automated evaluation. Human evaluation is expensive and slow, but automatic metrics (BLEU, ROUGE, METEOR) often disagree with human judgments and miss semantic failures. Recent work proposes reference-free evaluation using neural entailment models or question answering models.

Key papers