Automated Construction of Relational Data Schemas
Listen to the summary
Uses a voice available on your device
Audio options
On this page
Key Takeaways
- Tytan achieved 100 percent coverage of entities and attributes across seven diverse reference domains.
- The system guarantees execution correctness, with 100 percent of generated retrieval instructions executing successfully against test databases.
- Tytan automates the schema construction process to solve the knowledge acquisition bottleneck, particularly for non-technical users.
- The process uses a human-in-the-loop interactive clarification loop to resolve ambiguous database structures.
Summary & Methodology Analysis
Tytan utilizes a neurosymbolic approach, which is a hybrid architecture combining the high-level reasoning capabilities of LLMs with the rigid, rule-based verification of deterministic code. The pipeline begins with structural profiling to gather metadata, cardinalities, and null rates. This data informs an LLM-driven inference engine, which proposes potential entities, join paths, and attribute roles. To ensure reliability, the system subjects these proposals to deterministic verification, such as uniqueness checks and value-overlap probes, which confirm that the inferences are grounded in actual database state evidence. An interactive loop prompts the user for clarification when evidence is ambiguous, though it provides AI-suggested defaults to maintain progress without blocking. The final output is a semantic ring artifact containing verified join paths and metadata for downstream analytical use.
The system was evaluated across various domains, including wildfire incidents, school shootings, housing rents, regional income, and FIFA World Cup 2026 data, as well as the Spider text-to-SQL benchmark. Results show high performance, with semantic role agreement ranging from 92 percent to 100 percent against reference schemas. In a blind test on a ten-table database, the system successfully recovered the full entity structure and met 100 percent of the expectations defined by five independent annotators. Tytan is designed to integrate into environments using models like Claude Opus, Claude Sonnet, gpt-5.5, and the Satyrn system.
Despite these results, the system has clear architectural constraints. It is strictly limited to relational and tabular data formats and cannot process graph-based structures. Its join discovery mechanism is intentionally conservative, which may lead to missing relationships where identifier representations diverge across tables. Furthermore, value-overlap checks suffer from decreased reliability when only a data sample is available. Finally, the system relies heavily on database metadata; removing this information increases construction costs by approximately a factor of five.
Interactive System Flowchart
Cross-Examination & FAQs
A deeper dive clarifying mechanics, constraints, and baseline evaluations.
Q1. What problem does Tytan solve?
It addresses the knowledge acquisition bottleneck that makes it difficult to create semantic schemas for relational data, especially when database structures are ambiguous or lack context for non-technical users.
Q2. What is the final output of the Tytan process?
It produces a ring artifact, which includes entities, attributes, verified join paths, and surface forms ready for use by analytical systems.
Q3. Does the system require human input?
It uses an interactive clarification loop to ask the user questions when evidence is ambiguous, but it will use default AI-suggested answers if the user does not provide input.
Q4. How does Tytan ensure its LLM inferences are accurate?
It uses deterministic verification, which involves checking database evidence like uniqueness constraints and performing value-overlap probes to validate join paths.
Q5. Can Tytan work with graph databases?
No, the system is strictly limited to relational and tabular data and cannot process graph-shaped sources.
Q6. What is the impact of removing database metadata on performance?
The paper specifies that removing database metadata increases the construction cost by approximately a factor of five.
Q7. How did Tytan perform in blind testing?
On a held-out test of a ten-table database, it recovered the full entity structure and satisfied 100 percent of expectations from five independent annotators.
Q8. What are the limitations regarding join discovery?
The discovery is conservative and may miss relationships where identifier representations differ between tables, and its reliability is reduced when only a data sample is available.
Q9. What specific models were used or referenced in the paper?
The paper references Claude Opus, Claude Sonnet, gpt-5.5, and the Satyrn system.