Skip to content

Data-to-Text Generation

Data-to-text generation systems automatically produce natural language descriptions from structured data sources—tables, databases, knowledge bases, and logical forms. Applications include weather reports, financial summaries, sports scores, medical report generation, and knowledge base documentation.

Hallucination in data-to-text

Data-to-text generation is particularly vulnerable to numeric hallucinations—errors in numbers, dates, quantities, and other scalar values. These hallucinations are problematic because:

  1. Small numeric errors (wrong date by one day, wrong quantity by 1%) are fluent but factually incorrect
  2. Users trust numeric information and may not notice or catch errors
  3. Incorrect numeric claims can have serious consequences (financial reports, medical data)

Other hallucination types include: - Attribute hallucination: Generating descriptions of data not present in the input table - Relational hallucinations: Claiming relationships (e.g., "X is larger than Y") unsupported by the data

Architectures

Sequence-to-sequence models with table encoding (flattened tables, graph neural networks over table structure, or retrieval-augmented approaches) are standard. Key models include TGEN, Struct2Seq, and recent pretrained models (BART, T5) fine-tuned on data-to-text datasets.

Evaluation

  • Automatic metrics: BLEU, METEOR, PARENT (penalizes hallucinations outside the table)
  • Human evaluation: Fluency, informativeness, factuality
  • Data correctness metrics: Check whether each numeric claim in the generated text is supported by the input table

Datasets

  • E2E (restaurant reviews from data)
  • WebNLG (describing knowledge base records)
  • Table-to-Text (Wikipedia tables)

Key papers