Tytan: Interactive Neurosymbolic Construction of Analytic Semantic Schemas from Relational Data
TYTAN builds analytic semantic schemas from relational databases, using LLM inference only where symbolic database evidence needs semantic judgment.
The system identifies entities, attributes, measures, identifiers, and table connections, then asks targeted natural-language questions when a decision is ambiguous. In seven reference domains, the authors report 100% coverage of expert-corrected schema elements and 1,678 of 1,678 generated retrieval claims executing correctly. Semantic role agreement landed between 92% and 100%, with the paper saying the remaining disagreement came from reference errors. On a blind ten-table database with no declared keys, TYTAN recovered the full entity structure and met all satisfiable expectations from five independent annotators. ArXiv · AI/CL/LG's note
The system identifies entities, attributes, measures, identifiers, and table connections, then asks targeted natural-language questions when a decision is ambiguous. In seven reference domains, the authors report 100% coverage of expert-corrected schema elements and 1,678 of 1,678 generated retrieval claims executing correctly. Semantic role agreement landed between 92% and 100%, with the paper saying the remaining disagreement came from reference errors. On a blind ten-table database with no declared keys, TYTAN recovered the full entity structure and met all satisfiable expectations from five independent annotators. ArXiv · AI/CL/LG's note
score 4