Tytan: Interactive Neurosymbolic Construction of Analytic Semantic Schemas from Relational Data

TL;DR

Tytan combines symbolic analysis and LLM inference to automatically construct semantic schemas from relational databases, enhancing data understanding and querying.

cs.DB 🔴 Advanced 2026-08-07 79 views
Donna Hooshmand Shubham Shahi Cameron Barrie Abhratanu Dutta Marko Sterbentz Harper Pack Kristian J. Hammond
Knowledge Automation Semantic Modeling Neurosymbolic System Relational Databases Natural Language Interfaces

Key Findings

Methodology

Tytan employs a hybrid approach integrating symbolic structural analysis with large language model (LLM) based semantic inference. The process begins with deterministic symbolic profiling—detecting primary keys, foreign keys, sample values, and data types. Next, the LLM proposes candidate entities, attributes, and roles grounded in database descriptions and structural clues. These proposals are validated through rule-based checks such as value overlap, key uniqueness, and plausibility assessments. When ambiguity remains, Tytan engages in human-in-the-loop interaction, asking targeted natural language questions to resolve uncertainties. The final output is a structured JSON schema called 'ring', encapsulating entities, attributes with roles, verified relationships, and human-readable surface forms. This schema supports downstream analytical tasks, natural language queries, and report generation, with all components validated against the actual data to ensure correctness and robustness.

Key Results

  • Tytan achieved 100% coverage of entities, attributes, and features in seven benchmark domains, matching expert-annotated schemas. All 1,678 self-generated retrieval instructions executed correctly, demonstrating high functional accuracy. Semantic role agreement ranged from 92% to 100%, indicating reliable semantic inference. In a blind test with a live, ten-table database lacking declared keys, Tytan successfully reconstructed the full entity structure, verified keys, and satisfied all five independent annotators’ expectations, confirming its robustness in complex, real-world scenarios.
  • Across diverse datasets, Tytan outperformed existing static schema recovery tools and pure neural approaches by combining structural verification with semantic inference. Its ability to detect missing relationships, validate entity keys, and generate human-readable labels significantly enhances usability. The system’s interactive component effectively resolves ambiguities, reducing errors and increasing trustworthiness of the generated schemas.
  • The experimental results validate Tytan's capacity to produce accurate, comprehensive, and verifiable semantic schemas, enabling more reliable natural language interfaces and automated data analysis pipelines. Its performance on both benchmark and real-world databases demonstrates its potential for scalable deployment in enterprise data environments.

Significance

This work addresses a critical bottleneck in data management—the manual creation of semantic layers that interpret raw relational schemas. By automating the construction of high-quality, verifiable semantic schemas, Tytan reduces dependency on domain experts, accelerates data onboarding, and enhances the accuracy of natural language query systems. Its neurosymbolic approach combines the interpretability and reliability of symbolic analysis with the flexibility and inference power of LLMs, setting a new standard for knowledge automation in relational data management. The ability to generate schemas that are both human-readable and machine-verifiable paves the way for more intelligent, accessible, and scalable data ecosystems, impacting fields from business intelligence to AI-powered data integration.

Technical Contribution

Tytan introduces a novel neurosymbolic pipeline that tightly integrates symbolic structural analysis with LLM-based semantic inference, validated through deterministic rules and human interaction. Its core innovations include a structured 'ring' schema format, multi-source evidence validation, and an interactive question-answer loop to resolve ambiguities. The system leverages specific algorithms such as value overlap checks for join validation, key verification via sample uniqueness, and LLM confidence scoring for semantic plausibility. This comprehensive framework ensures that the generated schemas are both structurally sound and semantically meaningful, enabling downstream systems to perform accurate queries and analyses. The approach advances the state-of-the-art by providing a fully automated, verifiable, and human-in-the-loop capable solution for semantic schema construction.

Novelty

Unlike prior work that either relied solely on symbolic schema recovery or on textual descriptions for conceptual modeling, Tytan uniquely combines symbolic structural analysis with LLM-driven semantic inference in an interactive framework. It is the first system to perform end-to-end automated construction of a verifiable semantic schema (ring) directly from relational data, incorporating rigorous validation steps and human feedback. This dual-method fusion addresses the limitations of existing approaches, such as hallucination in neural models and structural ambiguity in symbolic methods, establishing a new paradigm for knowledge-driven data understanding.

Limitations

  • Tytan's accuracy diminishes when faced with highly noisy or incomplete data, especially in cases with insufficient samples for reliable type inference or relationship validation.
  • The current implementation handles primarily one-to-many relationships and may struggle with complex many-to-many or cyclic relations, requiring further algorithmic refinement.
  • The human-in-the-loop component, while effective for accuracy, may limit scalability in fully automated, large-scale deployments where minimal manual intervention is desired.

Future Work

Future directions include enhancing the system’s robustness against noisy data, extending relationship discovery algorithms to better handle complex schemas, and integrating semi-supervised learning to reduce human interaction. Additionally, expanding support for semi-structured and unstructured data types, improving real-time schema updates in dynamic databases, and developing more sophisticated user interaction strategies will further advance the practical deployment of Tytan in enterprise environments.

AI Executive Summary

In the era of big data, relational databases serve as the backbone of information storage across industries. However, their structural complexity and lack of explicit semantic descriptions pose significant barriers to effective data utilization. Traditional approaches rely heavily on manual schema annotation, which is labor-intensive, error-prone, and quickly outdated as data evolves. This creates a critical bottleneck, especially for non-technical users who depend on semantic layers to interpret and query data effectively.

Addressing this challenge, the paper introduces Tytan, a pioneering neurosymbolic system designed to automatically construct high-quality semantic schemas—referred to as 'rings'—from relational databases. Tytan’s architecture integrates symbolic structural analysis with the inference capabilities of large language models (LLMs). The process begins with deterministic analysis of database metadata—detecting primary keys, foreign keys, and data samples—to establish a symbolic foundation. Subsequently, the LLM proposes candidate entities, attributes, and relationships, grounded in the structural clues and optional user descriptions.

The core innovation lies in the validation mechanisms embedded within Tytan. Every inference step is verified through rule-based checks—such as value overlap for join plausibility, key uniqueness, and semantic consistency—ensuring that the constructed schema accurately reflects the underlying data. When ambiguity persists, the system engages in human-in-the-loop interaction, asking targeted natural language questions to clarify uncertainties. This iterative process results in a structured, verifiable JSON schema that captures entities, their attributes with roles, relationships, and natural language labels, making the data accessible and interpretable.

Experimental evaluations across eight diverse databases demonstrate Tytan’s effectiveness. It achieves 100% coverage of entities and features in benchmark domains, with all retrieval instructions executing correctly and high semantic role agreement. Notably, in a blind test database lacking declared keys, Tytan successfully reconstructed the entire entity structure, verified key fields, and satisfied all annotator expectations. These results highlight its robustness and practical applicability in complex, real-world scenarios.

The significance of this work extends beyond technical novelty. By automating the creation of reliable semantic layers, Tytan reduces dependence on domain experts, accelerates data onboarding, and enhances natural language query systems. Its approach addresses longstanding issues in knowledge engineering, offering a scalable, transparent, and verifiable solution that bridges the gap between raw data and intelligent analysis. As data ecosystems grow in size and complexity, such automated semantic modeling tools will become indispensable, enabling more accessible, accurate, and automated data-driven decision-making.

Looking ahead, future work aims to improve robustness against noisy data, extend support for complex relationships, and reduce human intervention through semi-supervised learning. The integration of multi-modal data understanding and real-time schema updates will further enhance its utility. Overall, Tytan marks a significant step toward autonomous, trustworthy, and scalable knowledge construction in relational data management, promising transformative impacts across industries and research domains.

Deep Dive

Abstract

From natural-language query interfaces to automated report generation, data analysis tools need a description of the data: the real-world entities it contains, which columns function as measures or identifiers, and how tables connect into units of analysis. Today, this semantic layer is usually written by hand. This is a knowledge-acquisition bottleneck that limits the scalability of analytic systems, keeps non-technical users dependent on experts, and is itself error-prone. We present TYTAN, a system for automatically constructing an analytic semantic schema from a relational database and, when available, a short user-provided description. TYTAN combines symbolic analysis of the database with LLM-based semantic inference for entity proposal, role assignment, and naming. When the evidence leaves a decision ambiguous, TYTAN asks the user a targeted natural-language question. We evaluate TYTAN on eight databases spanning real-world and benchmark domains along the three axes that define a schema's functional utility: (i) coverage, are all important entities and features captured?; (ii) retrieval correctness, do the schema's instructions actually reach the data; and (iii) characterization accuracy, are semantic types correct? Across the seven reference domains, TYTAN reaches every entity, attribute, and aggregable feature of the expert-corrected reference schemas (100% coverage). Additionally, 100% of its retrieval instructions execute correctly (1,678 of 1,678 self-generated claims), and semantic roles agree with the reference on 92-100% of matched attributes. Checking the underlying data showed the small disagreement is in the reference, not in TYTAN. On a held-out blind test (a live, ten-table database with no declared keys), TYTAN recovers the full entity structure with verified keys and satisfies 100% of the satisfiable expectations of five independent blind annotators.

cs.DB cs.AI

References (20)

Lightweight Knowledge Representations for Automating Data Analysis

Marko Sterbentz, Cameron Barrie, Donna Hooshmand et al.

2023 2 citations ⭐ Influential View Analysis →

Satyrn: A Platform for Analytics Augmented Generation

Marko Sterbentz, Cameron Barrie, Shubham Shahi et al.

2024 3 citations ⭐ Influential View Analysis →

Natural language to SQL: Where are we today?

Hyeonji Kim, Byeong-Hoon So, Wook-Shin Han et al.

2020 162 citations

Business Intelligence and Analytics: From Big Data to Big Impact

Hsinchun Chen, Roger H. L. Chiang, V. Storey

2012 6057 citations

CatSQL: Towards Real World Natural Language to SQL Applications

Han Fu, Chang Liu, Bin Wu et al.

2023 85 citations

Holistic primary key and foreign key detection

Lan Jiang, Felix Naumann

2019 43 citations

From data to information: automating data science to explore the U.S. court system

Andrew R. Paley, Andong Luis Li Zhao, Harper Pack et al.

2021 10 citations

Divide & Conquer-based Inclusion Dependency Discovery

Thorsten Papenbrock, Sebastian Kruse, Jorge-Arnulfo Quiané-Ruiz et al.

2015 77 citations

Recovering Semantics of Tables on the Web

Petros Venetis, A. Halevy, Jayant Madhavan et al.

2011 384 citations

Sherlock: A Deep Learning Approach to Semantic Data Type Detection

Madelon Hulsebos, K. Hu, Michiel A. Bakker et al.

2019 229 citations View Analysis →

Schema summarization

Cong Yu, H. V. Jagadish

2006 134 citations

U.S. Bureau of Labor Statistics

William F. Alterman

2380 citations

Judging LLM-as-a-judge with MT-Bench and Chatbot Arena

Lianmin Zheng, Wei-Lin Chiang, Ying Sheng et al.

2023 10321 citations View Analysis →

CHORUS: Foundation Models for Unified Data Discovery and Exploration

Moe Kayali, A. Lykov, Ilias Fountalis et al.

2023 58 citations View Analysis →

Towards an LLM-based Tool for Automated Database Design

Predrag Divljan, Drazen Brdjanin

2025 1 citations

Can LLM Already Serve as A Database Interface? A BIg Bench for Large-Scale Database Grounded Text-to-SQLs

Jinyang Li, Binyuan Hui, Ge Qu et al.

2023 1075 citations View Analysis →

On multi-column foreign key discovery

Meihui Zhang, Marios Hadjieleftheriou, B. Ooi et al.

2010 116 citations

Spider: A Large-Scale Human-Labeled Dataset for Complex and Cross-Domain Semantic Parsing and Text-to-SQL Task

Tao Yu, Rui Zhang, Kai-Chou Yang et al.

2018 2018 citations View Analysis →

The entity-relationship model: toward a unified view of data

Peter P. Chen

1975 6523 citations

How to Develop a Drug Target Ontology – KNowledge Acquisition and Representation Methodology (KNARM)

Hande Küçük-McGinty, U. Visser, S. Schürer

2019 5 citations