Fact-to-Text

Fact-to-text is the NLP task of generating natural language sentences that express specific factual statements — converting structured facts (entity-attribute-value triples, factual records, or knowledge assertions) into grammatically correct, fluent sentences that accurately convey the stated facts.

What Is Fact-to-Text?

• Definition: Generating text that expresses given factual statements.
• Input: Structured facts (triples, key-value pairs, assertions).
• Output: Natural language sentence(s) stating those facts.
• Goal: Accurate, fluent verbalization of factual information.

How Fact-to-Text Differs from General Data-to-Text

• Scope: Facts are typically atomic statements (single assertions).
• Focus: Emphasis on factual accuracy over narrative flow.
• Input: Usually simpler structures (single or few triples).
• Output: Often single sentences or short passages.
• Evaluation: Factual correctness is the primary criterion.

Input Fact Types

Entity-Attribute-Value:
- (Barack Obama, birthDate, 1961-08-04).
- (Python, creator, Guido van Rossum).
- (Silicon, atomicNumber, 14).

RDF Triples:
- (dbr:Paris, dbo:country, dbr:France).
- (dbr:Einstein, dbo:field, dbr:Physics).

Simple Assertions:
- Company X was founded in 2020.
- Product Y costs $99.

Composite Facts:
- Multiple related facts about one entity.
- Example: Einstein was born in Ulm, Germany in 1879. He won the Nobel Prize in Physics in 1921.

Generation Approaches

Template-Based:
- Method: Predefined sentence patterns for each relation type.
- Example: "[Subject] was born on [Date] in [Place]."
- Benefit: Perfect factual accuracy guaranteed.
- Limitation: Repetitive, limited to known relation types.

Neural Generation:
- Method: Seq2Seq/Transformer maps facts to text.
- Training: Parallel corpus of (facts, sentences).
- Benefit: Natural, varied output.
- Challenge: Risk of hallucination (adding unstated facts).

LLM Prompting:
- Method: Provide facts in prompt, instruct to verbalize.
- Technique: "Express these facts in a natural sentence: [facts]."
- Benefit: Strong quality without fine-tuning.
- Challenge: May embellish beyond stated facts.

Constrained Generation:
- Method: Decode text while constraining to express given facts.
- Lexical Constraints: Required words/phrases in output.
- Semantic Constraints: Entailment checking during generation.
- Benefit: Balances fluency with factual accuracy.

Key Challenges

• Factual Faithfulness: Express ALL given facts, add NO extra facts.
• Natural Language: Output should read naturally, not robotically.
• Aggregation: Combine multiple facts into coherent sentences.
• Referring Expressions: Use appropriate pronouns and references.
• Numerical Precision: Preserve exact numbers, dates, quantities.
• Negation: Handle negative facts accurately.
• Rare Entities: Generalize to unseen entity names.

Evaluation Metrics

Factual Accuracy:
- Precision: % of stated facts actually in input.
- Recall: % of input facts expressed in text.
- F1: Harmonic mean of precision and recall.

Text Quality:
- BLEU, METEOR, BERTScore: Similarity to reference text.
- Fluency: Human-rated naturalness.
- Grammaticality: Error-free text.

Faithfulness:
- NLI-based: Does the generated text entail from the facts?
- Fact extraction: Extract facts from generated text, compare with input.

Applications

• Knowledge Base Completion: Verbalize new KB entries.
• Fact-Checking: Generate claims from facts for verification testing.
• Question Answering: Verbalize factual answers from KBs.
• Education: Generate factual quizzes and explanations.
• Data Journalism: Auto-generate factual news from data.
• Chatbots: Provide factual responses from structured backends.

Key Datasets

• WebNLG: RDF triples → text (standard benchmark).
• WikiBio: Wikipedia infobox facts → biography sentences.
• E2E NLG: Meaning representation facts → restaurant descriptions.
• KELM: Knowledge-enhanced language model corpus.
• GenWiki: Wikidata facts → Wikipedia sentences.

Tools & Models

• Models: T5, BART, GPT-4 for fact verbalization.
• Constrained Decoding: NeuroLogic, FUDGE for constrained generation.
• Evaluation: FactCC, DAE for faithfulness checking.
• KG Tools: RDFLib, SPARQLWrapper for fact extraction.

Fact-to-text is the atomic unit of data-to-text generation — getting individual facts right in natural language is the foundation for all more complex data narration tasks, from report generation to knowledge base verbalization to automated journalism.