Table-to-text is the NLP task of generating natural language descriptions from tabular data — automatically converting rows, columns, and cells in tables into coherent, accurate text that describes the table contents, highlights key patterns, and presents data insights in readable form.

What Is Table-to-Text?

• Definition: Generating natural language from tabular data.
• Input: Table (rows, columns, cells, headers).
• Output: Fluent text describing table contents.
• Goal: Make tabular data accessible through natural language.

Why Table-to-Text?

• Accessibility: Tables are hard to read — text is easier to scan.
• Automation: Generate descriptions for thousands of tables.
• Insights: Highlight key patterns and trends in natural language.
• Assistive: Screen readers can convey table content through text.
• Summarization: Concise table summaries for reports and briefs.
• Search: Text descriptions make table content searchable.

Table-to-Text Approaches

Template-Based:

• Method: Pre-defined templates with slots filled from table data.
• Example: "[Player] scored [Points] points in [Minutes] minutes."
• Benefit: Guaranteed accuracy — only uses actual data values.
• Limitation: Repetitive, lacks variety and natural flow.

Neural Sequence-to-Sequence:

• Method: Encoder processes table, decoder generates text.
• Table Encoding: Linearize table or use structured encoding.
• Models: BART, T5, GPT with table-aware prompting.
• Benefit: Natural, varied output.
• Challenge: May hallucinate values not in table.

Pre-train + Fine-tune:

• Method: Pre-train on large table-text corpora, fine-tune on task.
• Models: TAPAS (Google), TaPEx, TableGPT.
• Benefit: Strong table understanding from pre-training.

LLM Prompting:

• Method: Provide table in prompt, ask for description.
• Technique: Markdown/CSV table format in prompt.
• Benefit: No training needed, leverages LLM capabilities.
• Challenge: Token limits for large tables.

Table Encoding Strategies

Linearization:

• Convert table to text: "Column1: value1 | Column2: value2."
• Row-by-row or column-by-column traversal.
• Simple but loses structural information.

Structured Encoding:

• Encode row/column positions explicitly.
• Cell-level embeddings with position encoding.
• Preserve table structure in representation.

Graph Encoding:

• Represent table as graph (cells as nodes, relationships as edges).
• Capture row, column, and cross-cell relationships.
• GNN-based encoding for structural reasoning.

Key Challenges

• Faithfulness: Generated text must be supported by table data.
• Numerical Reasoning: Correctly compute sums, averages, comparisons.
• Content Selection: Choose most important cells to describe.
• Large Tables: Handle tables with hundreds of rows/columns.
• Multi-Table: Reason across multiple related tables.
• Domain Specificity: Medical, financial, scientific tables need domain language.

Evaluation

• BLEU/ROUGE: Surface text similarity (limited for faithfulness).
• PARENT: Measures precision and recall against table content.
• Entailment: Does the text logically follow from the table?
• Human Evaluation: Fluency, faithfulness, coverage, informativeness.

Key Datasets

• ToTTo: Controlled table-to-text with highlighted cells.
• WikiTableText: Wikipedia infobox tables with descriptions.
• WikiBio: Wikipedia biography infoboxes → first paragraph.
• HiTab: Hierarchical tables with complex structure.
• SciGen: Scientific tables → descriptions.
• TabFact: Table fact verification (related task).

Applications

• Business Reports: Narrate financial tables and dashboards.
• Sports: Game statistics to game summaries.
• Healthcare: Lab results to patient-friendly descriptions.
• Accessibility: Table descriptions for visually impaired users.
• Data Cataloging: Automatic table documentation in data lakes.

Tools & Models

• Models: T5, BART, GPT-4 for generation; TAPAS, TaPEx for understanding.
• Libraries: Hugging Face, LangChain for table-aware LLM pipelines.
• Platforms: Narrative Science, Arria for enterprise NLG.

Table-to-text is a critical data communication capability — it enables automated, accurate narration of tabular data, making information trapped in spreadsheets and databases accessible to everyone through natural language.