Detecting Knowledge Inconsistencies Across Text, Tables, and Knowledge Graphs

Researchers introduced Kontrast, a framework for detecting and categorizing knowledge conflicts across text, tables, and knowledge graphs.

The paper defines a taxonomy of cross-modal knowledge inconsistencies and checks table answers against KG evidence using Text-to-SPARQL and LLM reasoning.

Experiments on Table-QA datasets show these inconsistencies are common and can come from real-world conflicts, incomplete graph structure, or temporal mismatches.

The approach could help audit and reconcile facts across Wikipedia, Wikidata, and other sources, improving search, RAG, and LLM pretraining.