Detecting Knowledge Inconsistencies Across Text, Tables, and Knowledge Graphs
A recent publication on arXiv presents Kontrast, an automated system designed to identify inconsistencies at the modality level within Wikipedia and Wikidata. This research, documented under arXiv:2607.25959, focuses on recognizing and elucidating discrepancies among text, tables, and knowledge graphs. The authors introduce a classification system for cross-modal inconsistencies, which includes variations in information granularity, direct contradictions, temporal shifts, and gaps in knowledge graphs. By employing Text-to-SPARQL and LLM reasoning, Kontrast assesses table-based responses against KG data and organizes inconsistencies. Experiments conducted on multiple Table-QA datasets demonstrate that such inconsistencies are prevalent and revealing, underscoring genuine knowledge disputes and the lack of structure in knowledge graphs. This study holds significance for LLM pre-training and retrieval-augmented generation, where coherent knowledge across different modalities is essential.
Key facts
- arXiv:2607.25959 introduces Kontrast framework
- Detects modality-level inconsistencies across Wikipedia and Wikidata
- Taxonomy covers granularity differences, direct conflicts, temporal changes, KG incompleteness
- Kontrast uses Text-to-SPARQL and LLM reasoning
- Experiments on Table-QA datasets show inconsistencies are common
- Reveals true knowledge conflicts and missing KG structure
- Relevant for LLM pre-training and retrieval-augmented generation
- Published as a cross-type announcement on arXiv
Entities
Institutions
- arXiv
- Wikipedia
- Wikidata