Pith. sign in

REVIEW 4 major objections 4 minor 48 references

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

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read TYTAN automatically constructs an analytic semantic schema—a structured 'ring' of entities, attribute roles, join paths, and surface names—from a relational database, verifying every LLM proposal against the stored data and asking the…

desk verdict TYTAN's neurosymbolic grounding is the right design, but its self-generated retrieval tests don't prove semantic correctness; the paper overclaims 100%. read the letter →

arxiv 2608.06331 v1 pith:BMM4PEZ4 submitted 2026-08-06 cs.DB cs.AI

classification cs.DBcs.AI
keywords analyticsemanticschemalayerdiscoveryLLMgroundingforeignkeyverificationrelationaldatabasesneurosymbolic
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

TYTAN is a system that builds a semantic layer for a relational database automatically, turning tables, columns, and joins into named entities, analytic roles, and validated relationships. The paper claims that on seven reference databases the generated schemas cover every entity, attribute, and aggregable feature of expert-corrected references, and that all 3,758 self-generated retrieval instructions execute correctly against the source data. On a blind ten-table database with no declared keys, TYTAN recovers the full entity structure and satisfies every satisfiable expectation written by independent annotators. The central idea is that an LLM can propose a schema, but every proposal should be checked against deterministic facts extracted from the database itself.

What carries the argument

The central artifact is the 'ring': a structured JSON schema that names entities, assigns each attribute an analytic role (identifier, categorical, numeric/metric, or datetime), records verified join paths and their direction, and supplies human-readable surface forms. The load-bearing process is 'verify-after-propose': symbolic profiling extracts keys, samples, cardinalities, and null rates; the LLM proposes entities, attributes, roles, and joins; and deterministic checks—value overlap, uniqueness, type evidence, and key-role exclusions—validate or reject each proposal before it enters the ring. An interactive clarification loop resolves decisions the data alone cannot settle.

What would settle it

On the metadata-stripped hospital database, four coincidental-overlap joins leaked through the guards, and the first blind FIFA run accepted three spurious joins; a decisive test would run TYTAN on a corpus of tables with many overlapping small-integer identifier spaces and check whether every accepted join corresponds to a genuine foreign-key relationship, or verify whether the revised guards actually reject the three spurious FIFA joins.

Watch

Extended reading notes

Core claim

The central claim is that a high-quality analytic semantic schema can be constructed semiautomatically by combining symbolic database analysis with LLM-based inference, provided every LLM proposal is grounded in the data. Entities are proposed by an LLM and then matched to tables; joins are proposed and then filtered through value-overlap checks, key-role guards, and an LLM plausibility threshold; attribute roles are inferred semantically but corrected by sample values and declared types; and entity keys are accepted only after uniqueness is verified against the rows. When the evidence is ambiguous, the system asks the user a focused natural-language question and applies the answer through the same validation pipeline, so the final schema is both complete and executable by a downstream analytic system.

Load-bearing premise

The deterministic guards and the LLM confidence threshold together keep spurious joins—overlaps that look like relationships but are not—out of the final schema.

Editorial extensions

If this is right

  • Semantic layers can be generated automatically for databases that lack declared primary keys and foreign keys, including raw CSV exports, without requiring manual annotation.
  • The self-generated retrieval test suite—ID uniqueness, attribute mappings, join-chain execution, and reference resolution—provides a regression harness for schema correctness that can catch key-selection and join-direction errors.
  • A generated ring can be loaded into a schema-aware analytics platform to answer natural-language analytical questions over the live database, as demonstrated for the school-shooting and wildfire domains.
  • Because structure and roles survive metadata stripping in the ablation, the method is not dependent on well-curated database constraints, though join recovery degrades when many identifier columns share overlapping value ranges.
  • The clarification loop lets a non-expert resolve ambiguity with targeted questions, and every question can be skipped with a default answer, so the pipeline works fully autonomously when needed.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the retrieval tests are generated from TYTAN's own output, a 100% pass rate does not by itself certify semantic correctness; the coverage evaluation against independent expectations is the complementary check, and the two should be used together in future evaluations.
  • The paper's own stress cases—three spurious joins on the first blind FIFA run and four coincidental-overlap joins leaking past the guards in the hospital ablation—suggest that value-overlap-based join filters are the weakest link; a stricter criterion based on functional dependency or join cardinality could be tested as a drop-in guard.
  • The ring's demonstrated compatibility with the SATYRN platform does not yet establish portability to other semantic-layer consumers, which the paper explicitly lists as future work.
  • A testable extension would be to measure how often the clarification loop's default answers (used in the fully autonomous runs) diverge from what a human would answer, since the paper does not evaluate the quality of each default separately.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper presents TYTAN, a pipeline that combines deterministic database profiling, LLM-based proposal of entities, attributes, and joins, deterministic verification against the database, and a clarification loop to produce "rings," JSON semantic schemas for analytical use. It evaluates TYTAN on eight databases — seven with author-written "expert-corrected" reference schemas and one held-out blind FIFA World Cup 2026 dataset with no declared keys — along three axes: coverage, retrieval correctness, and characterization accuracy. The paper reports 100% entity/feature coverage on all reference domains, 100% execution of 3,758 self-generated retrieval tests, 92–100% role agreement, and 100% recall on the blind expectation suite after human adjudication. It also presents a metadata ablation and an artifact-grounded audit as supporting evidence.

Significance. If the central claims were fully supported, TYTAN would be a valuable contribution to automated semantic schema construction, a real bottleneck for analytic systems. The paper's strengths include deterministic grounding of keys and joins, live-database execution of retrieval claims, a metadata-ablation control, a frozen expectation suite for the blind test, and a defect taxonomy for auditing. These are concrete, reproducible mechanisms and the failure analysis is unusually candid. However, as detailed below, the evaluation does not establish that the generated schemas are semantically correct: the self-generated retrieval tests only demonstrate executability, and spurious joins can pass both the guards and the tests (as the paper's own ablation shows). In addition, the reference ground truth is authored and arbitrated by the same group, creating circularity in the coverage and characterization claims. The system and the evaluation framework are nonetheless promising; a careful revision that tightens the claims and adds an independent precision oracle for joins could make this a strong paper.

major comments (4)
  1. [§5.7, Table 5] The metadata ablation on the stripped hospital database shows four "coincidental-overlap joins" leaking past the §4.3 guards, while the RQ2 suite still passes 589/590 claims, with the only failure being a composite-key fallback. This directly demonstrates that the self-generated retrieval tests validate executability, not semantic correctness: a spurious join between unrelated tables executes successfully because both columns exist, their values overlap, and the path returns rows. The abstract's and §7's claim that "successfully executed all 3,758 self-generated retrieval tests" therefore cannot be read as evidence that the generated schemas are semantically correct; at most it shows that the rings are internally executable. Please either add an oracle-based precision metric for inferred joins (for example, all inferred joins independently judged on a held-out set, or precision against declared foreign keys in the Spider databases), or restrict the claims to executability and report join precision explicitly.
  2. [§5.1, §5.4, abstract] The ground-truth references for the seven reference domains were authored by the authors themselves ("we annotated the schema ourselves, learning the data through queries"), and disagreements are resolved in the system's favor: the abstract states that "Checking the underlying data showed the small disagreement is in the reference, not in TYTAN." Since the reference is the only oracle, declaring it wrong based on the same data and the same system's output is circular; RQ1 coverage and RQ3 role-agreement numbers are therefore not measured against an independent standard. Please re-run the evaluation against externally available ground truth (for example, Spider's original annotations or independent domain experts who did not see TYTAN output), or explicitly relabel the references as "author-defined target schemas" and avoid external-correctness wording.
  3. [§5.5] The blind FIFA result is reported only for the pipeline after post-hoc modifications: the first run produced three spurious joins that passed LLM verification, and a deterministic guard was subsequently added ("Each resolution is deterministic and dataset independent, and can be regression-tested"). Because the fixes were derived after seeing the held-out test, the final 100% recall and retrieval numbers are not a fully blind evaluation. Please report first-run and final-run metrics separately, and, if possible, validate the new guard on a second held-out database or via cross-validation over the existing domains.
  4. [§5.2, Table 4] The blind expectation suite includes three LLM annotators, two of which are from the same model family as TYTAN's generator (Claude Opus/Sonnet and gpt-5.5); the paper acknowledges this potential bias. The human annotators' expectation lists are very small (Human #1: 4 entities, 14 features, 2 relationships; Human #2: 3 entities, 12 features, 3 relationships), so the 100% recall is based on a thin sample. Please report confidence intervals or the full list of human expectations, and consider an independent human annotation with more annotators.
minor comments (4)
  1. [Table 2 / §7] Please clarify that the 3,758 self-generated retrieval tests aggregate only the final runs of the seven reference domains plus the fixed FIFA run, and exclude the metadata-stripped arms in Table 5, which add 590 claims and one failure.
  2. [§5.4, Table 3] The text says two misclassified attributes were "corrected after the checks and verified at 100%," but Table 3 reports college role agreement of 93.9% and hospital role agreement of 94.7%; please reconcile the text with the table.
  3. [§4.3] The "strict confidence threshold" used for LLM join verification is never quantified; please report the exact threshold and, ideally, a sensitivity analysis, since join acceptance depends on it.
  4. [Appendix E] The phrase "schema-contrained outputs" contains a typo; it should be "schema-constrained outputs."

Circularity Check

3 steps flagged · score 6.0 of 10

Headline coverage and retrieval numbers rest on self-authored references and self-generated tests; the paper's own ablation shows spurious joins pass the RQ2 suite.

  1. other [§5.1 (Datasets and ground truth) and Abstract]
    "For these, we annotated the schema ourselves, learning the data through queries and deciding what the schema should look like. ... we resolve any structural disagreements we find between the dataset and the generated schema. ... Checking the underlying data showed the small disagreement is in the reference, not in TYTAN."

    The RQ1/RQ3 ground truth is written by the same group that built TYTAN, and the evaluation protocol lets the authors resolve structural disagreements and later declare the disagreement to be in the reference, not in TYTAN. A reference that can be revised after seeing the system's output is not a fixed external standard, so the reported 100% coverage and 92-100% role agreement measure agreement with the authors' own adjustable annotations rather than with an independent benchmark.

  2. self definitional [§5.3 (RQ2: Self-generated retrieval tests) and §7 (Conclusion)]
    "Because these tests are generated from TYTAN’s own output, they cannot catch omissions, that's why we do a coverage evaluation on entities. Retrieval Correctness checks that everything the ring claims is executable and truthful. ... successfully executed all 3,758 self-generated retrieval tests against the source data."

    The retrieval tests are synthesized from the ring itself, so the 100% pass rate checks only that the ring's self-declared table, column, and join mappings execute against the database. The paper's own §5.7 metadata ablation shows four coincidental-overlap joins 'leak past the guards' while the RQ2 suite still passes 589/590, so a ring with spurious joins can report near-100% retrieval. The headline number is therefore a self-consistency check, not an independent validation of semantic correctness.

1 more flagged steps
  1. other [§5.2 (Independent expectation suite) and Abstract]
    "Tytan’s own generator is an LLM reading the same description, and one of the panel LLMs belong to the same model family. So LLM annotated lists partly measure the inter-model convergence. We therefore treat the human list as ground truth. ... satisfies 100% of the satisfiable expectations of five independent blind annotators."

    The abstract's 'five independent blind annotators' includes three LLM annotators, and the paper concedes that one of them belongs to the same model family as TYTAN's generator, so those lists 'partly measure inter-model convergence.' The 100% blind-suite figure is thus partly auto-correlated with the generator itself; only the two human lists are fully independent, though they also pass.

full rationale

TYTAN's pipeline has genuine independent content: deterministic verification runs against live databases, the frozen pre-generation human expectation lists, and the artifact-grounded audit all provide real signals, and the paper honestly reports failures (three spurious joins on the first FIFA run, four leaking joins in the hospital ablation). However, the headline evaluation claims that support the strongest conclusions are partly self-referential. The seven reference schemas used for RQ1/RQ3 coverage and characterization are authored by the same group and are treated as revisable when the system disagrees. The RQ2 '100% retrieval' result is computed on tests generated from TYTAN's own output, and the paper's own ablation demonstrates that such tests pass even when spurious joins are present. Finally, the blind expectation panel includes LLMs from the generator's model family, weakening the word 'independent' in the abstract. These do not make the system's engineering contribution vacuous, but the central claim of semantic completeness and correctness is supported less independently than the abstract suggests.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a small number of engineering thresholds and evaluation assumptions rather than mathematical axioms. The key free parameters are the LLM confidence threshold, the overlap-guard multiplicity cutoff, the entity-count range, and the clarification defaults. No invented physical or semantic entities are introduced.

free parameters (4)
  • LLM join acceptance confidence threshold
    Section 4.3 requires candidates to be accepted only when the model assigns confidence above a fixed threshold; the threshold value is not reported and its calibration is not evaluated.
  • Value-overlap guard multiplicity cutoff = 3 or more
    Section 4.3 drops a column's join candidates if its values overlap keys of three or more tables; this tuning choice is ad hoc and is the exact point where the hospital ablation leaks four spurious joins.
  • Entity proposal count range = 2-6
    Appendix E instructs the LLM to propose between two and six entities; this hand-chosen range bounds the schema and is not ablated.
  • Clarification default answers
    Section 4.6 and the Limitations state each clarification has an AI-suggested default so the pipeline can run unattended; the quality of those defaults is not measured, yet the no-answer blind runs rely on them.
assumptions (4)
  • domain assumption Value-based validation is sufficient to establish semantic correctness of LLM proposals.
    The design in §4 grounds every LLM output with keys, value overlap, and sample types, but does not use external semantic ground truth during construction; the correctness of "verified" joins depends on this assumption.
  • domain assumption Declared primary and foreign key metadata is trustworthy.
    Section 4.3 accepts declared constraints directly without data validation, and §4.4 trusts declared types; the paper acknowledges dynamically typed databases can contradict declarations but still uses declarations as strong evidence.
  • ad hoc to paper Expert-corrected reference schemas written by the authors are an unbiased ground truth.
    Section 5.1 says "we annotated the schema ourselves" and the ground-truth rings are "expert-corrected" versions; if the references encode the authors' expectations, the 100% coverage result is less informative.
  • domain assumption Self-generated retrieval claims test correctness rather than internal consistency.
    Section 5.3 says the tests are generated from TYTAN's own output and "cannot catch omissions", so passing them only proves internal consistency with the database, not semantic accuracy of joins and types.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Tytan: Interactive Neurosymbolic Construction of Analytic Semantic Schemas from Relational Data." pith.science (2026). https://pith.science/paper/BMM4PEZ4

@misc{pith2026260806331,
  author       = {Pith},
  title        = {Pith review of: Tytan: Interactive Neurosymbolic Construction of Analytic Semantic Schemas from Relational Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BMM4PEZ4}},
  note         = {Machine review of arXiv:2608.06331}
}
read the original 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.

Figures

Figures reproduced from arXiv: 2608.06331 by the authors.

Figure 1
Figure 1. TYTAN overview. A database instance and an optional one-sentence description enter the pipeline. Symbolic profiling, LLM proposals, and deterministic verification interleave to produce a ring: a self-contained semantic catalog (an example of a ring is given in [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Anatomy of a generated ring (housing-rent [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Example end-to-end system interaction 6 Discussion Deterministic grounding. Most errors we en￾countered during development occurred when an LLM output was accepted without enough support from the database. Some example failures include: the model proposing joins because two columns had similar names, including example values that were not present in the data, inferring types from column names alone, and selecting ke… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: A snippet ring generated by TYTAN for the housing-rent domain, shown in the three-section presentation style used for the expert-annotated rings of Sterbentz et al. (2024): ring metadata (purple), data source and validated joins (pink), and the data abstraction layer (…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 45 canonical work pages

  1. [1]

    ACM Transactions on Database Systems (TODS) , volume=

    The entity-relationship model---toward a unified view of data , author=. ACM Transactions on Database Systems (TODS) , volume=. 1976 , publisher=

  2. [2]

    MIS Quarterly , volume=

    Business intelligence and analytics: From big data to big impact , author=. MIS Quarterly , volume=

  3. [3]

    ACM Computing Surveys (CSUR) , volume=

    A comparative analysis of methodologies for database schema integration , author=. ACM Computing Surveys (CSUR) , volume=. 1986 , publisher=

  4. [4]

    How to develop a drug target ontology:

    K. How to develop a drug target ontology:. Bioinformatics and Drug Discovery , series=. 2019 , publisher=

  5. [5]

    Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pages=

    Sherlock: A deep learning approach to semantic data type detection , author=. Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pages=

  6. [6]

    Proceedings of the VLDB Endowment , volume=

    Sato: Contextual semantic type detection in tables , author=. Proceedings of the VLDB Endowment , volume=

  7. [7]

    Deng, Xiang and Sun, Huan and Lees, Alyssa and Wu, You and Yu, Cong , journal=

  8. [8]

    Proceedings of the 2022 International Conference on Management of Data (SIGMOD) , pages=

    Annotating columns with pre-trained language models , author=. Proceedings of the 2022 International Conference on Management of Data (SIGMOD) , pages=

Show all 48 references
  1. [9]

    The VLDB Journal , volume=

    A survey of approaches to automatic schema matching , author=. The VLDB Journal , volume=. 2001 , publisher=

  2. [10]

    12th International Workshop on the Web and Databases (WebDB) , year=

    A machine learning approach to foreign key discovery , author=. 12th International Workshop on the Web and Databases (WebDB) , year=

  3. [11]

    Proceedings of the VLDB Endowment , volume=

    Divide & conquer-based inclusion dependency discovery , author=. Proceedings of the VLDB Endowment , volume=

  4. [12]

    Spider: A large-scale human-labeled dataset for complex and cross-domain semantic parsing and text-to-

    Yu, Tao and Zhang, Rui and Yang, Kai and Yasunaga, Michihiro and Wang, Dongxu and Li, Zifan and Ma, James and Li, Irene and Yao, Qingning and Roman, Shanelle and Zhang, Zilin and Radev, Dragomir , booktitle=. Spider: A large-scale human-labeled dataset for complex and cross-do...

  5. [13]

    Floratou, Avrilia and Psallidas, Fotis and Zhao, Fuheng and Deep, Shaleen and Hagleither, Gunther and Tan, Wangda and Cahoon, Joyce and Alotaibi, Rana and Henkel, Jordan and Singla, Abhik and others , booktitle=

  6. [14]

    Natural language to

    Kim, Hyeonji and So, Byeong-Hoon and Han, Wook-Shin and Lee, Hongrae , journal=. Natural language to

  7. [15]

    Fu, Han and Liu, Chang and Wu, Bin and Li, Feifei and Tan, Jian and Sun, Jianling , journal=

  8. [16]

    The dawn of natural language to

    Li, Boyan and Luo, Yuyu and Chai, Chengliang and Li, Guoliang and Tang, Nan , journal=. The dawn of natural language to

  9. [17]

    Towards an

    Divljan, Predrag and Brdjanin, Drazen , booktitle=. Towards an

  10. [18]

    arXiv preprint arXiv:2311.12848 , year=

    Lightweight knowledge representations for automating data analysis , author=. arXiv preprint arXiv:2311.12848 , year=

  11. [19]

    Sterbentz, Marko and Barrie, Cameron and Shahi, Shubham and Dutta, Abhratanu and Hooshmand, Donna and Pack, Harper and Hammond, Kristian J , booktitle=

  12. [20]

    arXiv preprint arXiv:2509.04632 , year=

    Conceptual schema inference for tabular datasets using large language models , author=. arXiv preprint arXiv:2509.04632 , year=

  13. [21]

    Information Sciences , volume=

    English sentence structure and entity-relationship diagrams , author=. Information Sciences , volume=. 1983 , publisher=

  14. [22]

    Proceedings of the VLDB Endowment , volume=

    On multi-column foreign key discovery , author=. Proceedings of the VLDB Endowment , volume=

  15. [23]

    Journal of Intelligent Information Systems , volume=

    Holistic primary key and foreign key detection , author=. Journal of Intelligent Information Systems , volume=. 2020 , publisher=

  16. [24]

    Proceedings of the ACM on Management of Data , volume=

    D. Proceedings of the ACM on Management of Data , volume=

  17. [25]

    Column type annotation using

    Korini, Keti and Bizer, Christian , booktitle=. Column type annotation using

  18. [26]

    Feuer, Benjamin and Liu, Yurong and Hegde, Chinmay and Freire, Juliana , journal=

  19. [27]

    Proceedings of the VLDB Endowment , volume=

    Can foundation models wrangle your data? , author=. Proceedings of the VLDB Endowment , volume=

  20. [28]

    Proceedings of the VLDB Endowment , volume=

    How large language models will disrupt data management , author=. Proceedings of the VLDB Endowment , volume=

  21. [29]

    Li, Jinyang and Hui, Binyuan and Qu, Ge and Yang, Jiaxi and Li, Binhua and Li, Bowen and Wang, Bailin and Qin, Bowen and Geng, Ruiying and Huo, Nan and Zhou, Xuanhe and Ma, Chenhao and Li, Guoliang and Chang, Kevin and Huang, Fei and Cheng, Reynold and Li, Yongbin , booktitle=. Can

  22. [30]

    Zheng, Lianmin and Chiang, Wei-Lin and Sheng, Ying and Zhuang, Siyuan and Wu, Zhanghao and Zhuang, Yonghao and Lin, Zi and Li, Zhuohan and Li, Dacheng and Xing, Eric P and Zhang, Hao and Gonzalez, Joseph E and Stoica, Ion , booktitle=. Judging

  23. [31]

    Panickssery, Arjun and Bowman, Samuel R and Feng, Shi , booktitle=

  24. [32]

    Preregistering

    van Miltenburg, Emiel and van der Lee, Chris and Krahmer, Emiel , booktitle=. Preregistering

  25. [33]

    Proceedings of the National Academy of Sciences , volume=

    The preregistration revolution , author=. Proceedings of the National Academy of Sciences , volume=

  26. [34]

    The VLDB Journal , volume=

    Profiling relational data: a survey , author=. The VLDB Journal , volume=. 2015 , publisher=

  27. [35]

    The VLDB Journal , volume=

    A survey on semantic schema discovery , author=. The VLDB Journal , volume=. 2022 , publisher=

  28. [36]

    Proceedings of the 32nd International Conference on Very Large Data Bases (VLDB) , pages=

    Schema summarization , author=. Proceedings of the 32nd International Conference on Very Large Data Bases (VLDB) , pages=

  29. [37]

    Proceedings of the VLDB Endowment , volume=

    Recovering semantics of tables on the web , author=. Proceedings of the VLDB Endowment , volume=

  30. [38]

    Proceedings of the 43rd International Conference on Conceptual Modeling (ER 2024) , series=

    Enhancing domain modeling with pre-trained large language models: An automated assistant for domain modelers , author=. Proceedings of the 43rd International Conference on Conceptual Modeling (ER 2024) , series=. 2024 , publisher=

  31. [39]

    Kayali, Moe and Lykov, Anton and Fountalis, Ilias and Vasiloglou, Nikolaos and Olteanu, Dan and Suciu, Dan , journal=

  32. [40]

    , title=

    Short, Karen C. , title=. 2022 , howpublished=

  33. [41]

    Housing data - Zillow Research , year =

  34. [42]

    Unemployed persons in Cook County, IL , year =

  35. [43]

    Estimate of people age 0-17 in poverty in Cook County, IL , year =

  36. [44]

    Estimate of people of all ages in poverty in Cook County, IL , year =

  37. [45]

    Personal income in Cook County, IL , year =

  38. [46]

    Estimate of median household income for Cook County, IL , year =

  39. [47]

    arXiv preprint arXiv:2601.05451 , year=

    RingSQL: Generating Synthetic Data with Schema-Independent Templates for Text-to-SQL Reasoning Models , author=. arXiv preprint arXiv:2601.05451 , year=

  40. [48]

    Proceedings of the Eighteenth International Conference on Artificial Intelligence and Law , pages=

    From data to information: automating data science to explore the US court system , author=. Proceedings of the Eighteenth International Conference on Artificial Intelligence and Law , pages=

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.