Datasets & Resources
ReTaT - Relation Extraction across Text and Table
A manually annotated benchmark for joint relation extraction across text and tables: each gold triple (subject, relation, object) draws its evidence from both a paragraph and a table, so it cannot be recovered from either one alone.
| Task | Cross-modal (text + table) relation extraction |
| Evaluation metrics | Precision, Recall, F1 at three levels: exact triple match, entity-level partial match, relation-level partial match |
| Languages | English, French |
| Size | 3 subsets - 387 Wikipedia pages - 1,064 tables - ~6,400 annotated triples |
| Relation schemas | CORE company relations (Business) and Wikidata properties (Female Celebrities, Telecommunications) |
| Format | Paired files per document: paragraphs (TXT), tables (CSV), and gold triples in a separate annotation file indexed by document ID |
| Sources | English and French Wikipedia dumps (2024); Business subset built on CORE |
| Licence | CC BY-SA 4.0 |
Subsets
| Subset | Language | Content | Relations | Triples |
|---|---|---|---|---|
| Business | English | 255 table–paragraph pairs | 9 CORE relations (closed list) | 2,997 |
| Female Celebrities | French | 100 full pages, 665 tables | 11 Wikidata properties (open) | 1,836 |
| Telecommunications | English | 32 full pages, 144 tables | 35 Wikidata properties (open) | 1,574 |
Access
- 📦 Dataset: https://github.com/ecladatta/ReTaT
- 📄 Paper: https://aclanthology.org/2026.lrec-1.104.pdf
How to cite
Mohamed Ettaleb, Thibault Ehrhart, Nathalie Aussenac-Gilles, Yoan Chabot, Mouna Kamel, Véronique Moriceau, Raphaël Troncy, Fanfu Wei. ReTaT: A Unified Benchmark for Relation Extraction across Text and Table. In Proceedings of the Fifteenth Language Resources and Evaluation Conference (LREC 2026), pages 1341–1351, May 2026. ELRA.
LATTE - Linking Across Table and Text for Relatedness Evaluation
A human-annotated benchmark for semantic relatedness classification between tables and text paragraphs, built from Wikipedia.
| Task | Table–text relatedness classification |
| Evaluation metrics | P@k, R@k, F1@k |
| Language | English |
| Size | 7,674 tables - 41,880 paragraphs - 3,826 Wikipedia pages |
| Annotations | 70,031 table–paragraph pairs (9,784 related / 60,247 unrelated) |
| Sources | Wikipedia, reconstructed from HybridQA and QTSumm |
| Licence | CC BY-SA 4.0 |
Access
- 📦 Dataset: github.com/ecladatta/LATTE
- 📄 Paper: [to be completed]
How to cite
Fanfu Wei, Thibault Ehrhart, Raphaël Troncy. From Rows to Narratives: Benchmarking Semantic Relatedness Across Tables and Paragraphs. TRIPLET 2026 Workshop @ ESWC 2026, Dubrovnik, Croatia, May 2026.
KONTRAST - Detecting Knowledge Inconsistencies Across Text, Tables, and Knowledge Graphs
A benchmark and framework for modality-level inconsistency detection: for questions grounded in Wikipedia tables and text, KONTRAST compares the table-based answer with the answer retrieved from Wikidata through Text-to-SPARQL, then classifies any disagreement.
| Task | Detect whether a table-based answer and a KG answer agree and, if not, categorize the inconsistency |
| Labels | Same - Higher accuracy in KG than in Table - Higher accuracy in Table than in KG - Different answer - Temporal changes |
| Evaluation | SPARQL execution validity, value-bearing answer rate, distribution of inconsistency labels |
| Language | English |
| Size | 2,870 questions (1,370 SimpleQA / 1,500 ComplexQA) from 9 Table-QA sources, with SPARQL outputs and inconsistency labels for 3 Qwen3 models |
| Sources | Wikipedia tables, infoboxes and text (via NQ-Table, CompMix, QAMPARI, MoNaCo, OTT-QA, SportReason); Wikidata dump of 2025-05-10 |
| Contents | Raw and filtered QA inputs, generated SPARQL queries and KG results (GRASP), LLM-as-a-judge pipeline, final labelled files and statistics |
| Licence | [to be completed] |
Question groups
| Group | Source datasets | Evidence | Questions |
|---|---|---|---|
| SimpleQA | NQ-Table, CompMix-simple, QAMPARI | Table, infobox, or table + text | 1,370 |
| ComplexQA | CompMix-infobox, CompMix-table, MoNaCo, MoNaCo-time, OTT-QA, SportReason | Table, infobox, text, multi-hop | 1,500 |
Access
- 📦 Dataset and code: github.com/ecladatta/KONTRAST
- 📄 Paper: arXiv:2607.25959
How to cite
Fanfu Wei, Thibault Ehrhart, Raphaël Troncy. Detecting Knowledge Inconsistencies Across Text, Tables, and Knowledge Graphs. In Proceedings of the 25th International Conference on Knowledge Engineering and Knowledge Management (EKAW 2026), Turin, Italy, 2026.