Kontrast Audits Text, Tables, and Knowledge Graphs Against Each Other to Find Where Wikipedia and Wikidata Disagree

arXiv 2607.25959·medium signalXBlueskyLinkedInCopy link

Posted 2026-07-28 (arXiv 2607.25959), this work formalizes modality-level inconsistency detection for knowledge sources that feed LLM pre-training and RAG, with a taxonomy covering information-granularity differences, direct conflicts, temporal changes, and knowledge-graph incompleteness. Kontrast automates the comparison using Text-to-SPARQL plus LLM reasoning to check table-based answers against KG evidence and classify the disagreement. Experiments across Table-QA datasets show cross-modal inconsistencies are common and informative, exposing not only genuine knowledge conflicts but missing KG structure and temporal mismatches — though the pipeline is bounded by Text-to-SPARQL errors and noise. Code and data are released at github.com/ECLADATTA/KONTRAST.