Pith. sign in

REVIEW 2 major objections 6 minor 3 cited by

Conceptual Schema Inference for Tabular Datasets using Large Language Models

T0 review · 2 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read LLM prompts can infer a full conceptual schema from raw tables

desk verdict A useful integration of three schema-inference tasks, undermined by a circular relationship evaluation and a 'no-external-ontology' claim that doesn't hold up. read the letter →

arxiv 2509.04632 v2 pith:6NWRSG7L submitted 2025-09-04 cs.DB cs.AI

Zhenyu Wu , Jiaoyan Chen , Norman W. Paton This is my paper · ORCID
classification cs.DBcs.AI
keywords schemainferencetabulardatalargelanguagemodelslakesconceptualtypehierarchyattributerelationshipdiscovery
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

The paper claims that a prompt-only LLM pipeline, SI-LLM, can take a collection of heterogeneous tabular datasets and produce a concise conceptual schema — hierarchical entity types, their attributes, and relationships between types — using nothing but column headers and cell values. No bespoke training, no curated ontologies, no external metadata. Across two annotated repositories, the reported quality is comparable to or better than embedding-based methods on each inference step, and the end-to-end case study shows a coherent schema emerging from hundreds of tables. If true, this would let data-lake users get an organized semantic overview of an unfamiliar repository almost for free.

What carries the argument

The mechanism is a three-step prompt-based pipeline. Step 1 asks the LLM to output a top-down type path (Thing → … → leaf) for each table, then merges all paths into an edge-weighted global hierarchy and prunes self-loops, inverse edges, and is-a relations that a second LLM pass judges incorrect. Step 2 gives the LLM each column's header and up to five values to propose a canonical attribute name, then resolves and groups these names, and promotes an attribute to a parent type when it appears in at least a 0.9 fraction of child types. Step 3 samples up to five values from named-entity attributes, asks the LLM to pick the most specific matching type from the hierarchy, and generates a predica

What would settle it

Run SI-LLM on the WDC benchmark and inspect the merged hierarchy for the hallucination pattern the paper itself cites (e.g., Thing → Event → CreativeWork → MusicAlbum). If such an erroneous path survives the LLM-based pruning step in a non-trivial fraction of runs, the merging step cannot control LLM noise and the claimed reliability of the inferred schema is in doubt.

Watch

Extended reading notes

Core claim

The central discovery is that a large language model, prompted with just column headers and a handful of sampled cell values, can perform the three acts of conceptual schema inference that previously required separate tuned models: it can assign each table a path in a type hierarchy rooted at Thing, group equivalent columns across tables into canonical conceptual attributes, and detect that a named-entity attribute of one type refers to instances of another type, yielding a typed relationship with a predicate label. The authors show that these three steps can be chained end-to-end, and that merging per-table hierarchy paths with edge weighting and an LLM-based consistency filter produces a g

Load-bearing premise

The framework depends on LLMs producing mostly consistent and correct type paths across many tables, so that the merged hierarchy stays coherent; the paper's own filter for hallucinated is-a edges must catch enough noise to keep global structure intact.

Editorial extensions

If this is right

  • Data-lake catalogs could be auto-generated from raw tables alone, with no manual annotation or ontology engineering.
  • The pipeline can be rerun incrementally as new tables arrive, keeping the conceptual view of a repository current.
  • Because only headers and sampled values are used, the approach should transfer to new domains without retraining.
  • The edge-weighted merge plus LLM-based is-a filter may generalize to other LLM-inferred graph structures, such as ontology construction.

Reading between the lines

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

  • The promotion threshold θ=0.9 is a free parameter; making it adaptive to the distribution of child types could reduce both false promotions and missed inherited attributes, and is directly testable on the published benchmarks.
  • Relationship discovery as formulated only detects attributes whose values are instances of another type; the same mechanism could be extended to part-whole or temporal relationships by constraining the predicate-generation prompt, which the paper does not explore.
  • The reliance on a second LLM to validate is-a edges implies a cost-quality trade-off: swapping the verifier for a smaller or cheaper model may degrade the merged hierarchy, an implicit assumption that could be probed by substitution.
  • Removing column headers entirely would isolate whether the LLM reads semantics from headers or from values; the paper always feeds both, leaving this dependency untested.
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

2 major / 6 minor

Summary. The paper introduces SI-LLM, a three-step LLM-prompt-based framework for inferring a conceptual schema (hierarchical entity types, attributes, and inter-type relationships) from collections of tabular datasets using only column headers and cell values. S1 infers per-table type hierarchy paths via few-shot prompting, then merges and prunes them into a global hierarchy; S2 infers conceptual attributes by naming and resolving column-level attributes and promoting attributes shared by child types; S3 discovers relationships by matching sampled attribute values to candidate descendant types. The method is evaluated on WDC and GDS benchmarks against embedding-based baselines, with ablations of prompt constraints and an end-to-end case study. The paper reports high purity and PTCS for type hierarchies, competitive attribute inference, and improved relationship discovery F1, concluding that SI-LLM works without labeled data or external ontologies.

Significance. If the claims hold, SI-LLM is a practically useful contribution: it addresses the under-studied task of conceptual schema inference over heterogeneous tabular repositories and shows that general-purpose LLMs can produce compact semantic schemas without task-specific training. The paper has notable strengths: code and full prompts are released, multiple LLMs are tested, the prompt-constraint ablation is informative, error patterns are discussed candidly, and an end-to-end case study is included. However, two load-bearing evaluation issues undermine the central claims as stated. First, the type-hierarchy step is conditioned on schema.org ground-truth few-shot paths and a Wikidata-derived blacklist, which contradicts the 'only column headers and cell values' / 'no external ontologies' claims and may constitute in-context test-set leakage if the exemplars are not disjoint from the benchmarks. Second, the relationship ground truth is constructed with the same 'values predominantly refer to instances of another type' rule that the proposed method implements, making the relationship evaluation circular. These issues affect the interpretation of the quantitative results, though the algorit

major comments (2)
  1. [III.B, Fig. 2-3; Abstract; VI] The paper repeatedly claims that SI-LLM infers schemas 'using only column headers and cell values' and 'without labeled data or external ontologies.' Yet Section III-B and Figures 2-3 describe 3-shot type-hierarchy prompts whose shots include 'the ground-truth output path(s) from the schema.org type hierarchy,' and the ABS constraint passes a Wikidata-derived blacklist of abstract types. The FET variant used for the main results (Table II) first asks for the leaf type and then constrains the output path to terminate at that leaf, but the second prompt still inherits the schema.org examples. Since S2 and S3 consume the S1 hierarchy, this external supervision propagates through the entire pipeline. The paper does not state whether the few-shot exemplars are disjoint from the WDC/GDS test tables; if they are not, the type-hierarchy results are partly in-context test-set leakage. Even if the
  2. [IV.A.1 vs III.D] Ground-truth relationships are constructed by the rule 'If the values of an attribute ai from type ti predominantly refer to entities of another type tj, we annotate a relationship from ti to tj' (Section IV.A.1). This is exactly the operational criterion used by the proposed relationship discovery method (Section III.D: 'when the values of an attribute ai of type ti predominantly refer to instances of another type tj, ai indicates a relationship from ti to tj'), and Algorithm 2 implements this rule with LLM-based type matching. Evaluating against a ground truth generated by the same rule measures how faithfully the prompt implements the rule, not whether the discovered relationships are semantically correct. This circularity likely inflates the relationship precision/recall/F1 in Table III. The authors should provide an independent relationship ground truth (e.g., manually annotated or
minor comments (6)
  1. [IV.B.3 / Table III] The text states that SI-LLM attribute Purity reaches 0.65-0.71 on GDS, but Table III shows SI-LLM (GPT-3.5) with GDS S2 Purity 0.478. Please reconcile this discrepancy and discuss the variance across LLMs.
  2. [Table II] The formatting for EmTT (SBERT) and EmTT (Starmie) rows appears anomalous (e.g., '1 158' and '1 117' in the PTCS/T# columns). Please clarify whether these are PTCS=1.000 or missing-value markers.
  3. [IV.A.2] The PTCS definition says L(p') is 'the length of the longest subsequence of p' that forms a path in the GT hierarchy.' A subsequence is not necessarily contiguous; if a contiguous subpath is intended, say so, and if not, explain how non-contiguous matches are counted.
  4. [III.C / Algorithm 1] Several hyperparameters (θ, N, k, K, τ1, τ2) are fixed without sensitivity analysis. The ablation covers prompt constraints only. Please add at least a short sensitivity discussion for θ and k, which directly control attribute promotion and relationship sampling.
  5. [Figures 2 and 4] The prompts are abbreviated with placeholders (<ORIGINAL INSTRUCTION>, <EXAMPLES>, [criteria]). Since the repository is promised, this is acceptable, but the paper should explicitly state that full prompts are in the repository or an appendix.
  6. [V, Case Study] The statement that the schema 'recovers 87% of annotated types overall' is undefined. Please specify whether this is exact type-name matching, path-based matching, or another criterion.

Circularity Check

2 steps flagged · score 6.0 of 10

Relationship discovery is evaluated against a ground truth built from the same 'predominantly refer' rule the method implements, and type hierarchy inference feeds schema.org ground-truth paths into the prompt as few-shot examples, compromising the claimed ontology-free/minimal-metadata setting.

  1. fitted input called prediction [Section III.B (Individual Table Type Hierarchy Inference); Section IV.A.1 (Datasets)]
    "We use 3-shot prompting, each shot providing: (i) a five-row table sample (as in [33]); and (ii) the ground-truth output path(s) from the schema.org type hierarchy. ... Each set is annotated with a ground-truth (GT) hierarchy derived from Schema.org, and each table is annotated with both the most specific and top-level entity types."

    The per-table type hierarchy is produced with 3-shot prompts whose examples include the ground-truth schema.org paths for those example tables, and the same schema.org-derived hierarchy is then used as the evaluation ground truth. The paper does not state that the few-shot tables are disjoint from the test tables. Even if they are disjoint, the model is being fed the target ontology's labels as supervision, so the result is not derived 'using only column headers and cell values' as claimed. The reported PTCS/RI/Purity therefore partly measure reproduction of the schema.org labels already present in the prompt, rather than inference from raw tables alone.

  2. self definitional [Section III.A (Problem Definition); Section III.D (Relationship Discovery); Section IV.A.1 (Datasets)]
    "Each r ∈ R is a triple ⟨ti, ai, tj⟩, where ti, tj ∈ T and ai ∈ Atts(ti) is a named entity attribute whose values predominantly refer to instances of type tj. ... This observation provides the insight of SI-LLM: when the values of an attribute ai of type ti predominantly refer to instances of another type tj, ai indicates a relationship from ti to tj. ... To construct GT relationships between types, we utilize named entity attributes with reference to Schema.org. If the values of an attribute ai from type ti predominantly refer to entities of another type tj, we annotate a relationship from ti"

    In the problem definition a relationship is defined as a named-entity attribute whose values 'predominantly refer to instances of type tj.' The S3 method explicitly adopts this same observation as its inference insight, and the ground-truth relationships are annotated with the same 'predominantly refer' criterion. Algorithm 2 is therefore an LLM implementation of the annotation rule itself, and the precision/recall/F1 evaluation compares the method against labels generated by that same rule. The evaluation cannot validate whether the discovered relationships are semantically correct; it only measures how faithfully the LLM applies the annotators' definitional test. The 'prediction' is equivalent to the input definition by construction.

full rationale

The clearest circularity is in relationship discovery (S3): the paper defines a relationship in terms of a named-entity attribute whose values 'predominantly refer to instances of another type,' then uses exactly that criterion to construct the ground-truth relationships against which S3 is scored. This is a self-definitional evaluation. The type hierarchy step (S1) is partially circular in a different way: the 3-shot prompts include the ground-truth schema.org output paths, and the same schema.org-derived hierarchy is the evaluation target; the paper does not establish disjointness, so the evaluation is at least ontology-conditioned and at worst in-context leakage. The ABS constraint additionally uses a Wikidata-derived blacklist, which contradicts the 'without external ontologies' claim, though that is a consistency issue rather than a formal circularity. Attribute inference and hierarchy merging/pruning do not exhibit an equivalent by-construction reduction. The self-citation [31] is used as a baseline and is not load-bearing. Overall, the central end-to-end claim is partially compromised because two of the three steps are evaluated against criteria that are built into the method's inputs or definitions.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method rests on LLM black-box semantics and a relationship criterion that is also used to build the ground truth. Free hyperparameters (theta, N, k, K, tau) are manually chosen or tuned; none are fitted to a held-out set beyond the baselines' tau thresholds, which are tuned on the test set.

free parameters (5)
  • theta (attribute inheritance threshold) = 0.9
    Threshold for promoting an attribute to a parent type; chosen by hand in Algorithm 1 (Line 14).
  • N (max path length) = 5
    Maximum path length in type hierarchy prompt; set to one above ground-truth depth (Section III.B).
  • k (sampled values) = 5
    Number of sampled cell values in attribute and relationship prompts (Algorithms 1-2).
  • K (TopKMatchTopLevelTypes)
    Number of top-level types selected in TopKMatchTopLevelTypes; value not specified in the paper (Algorithm 2, Line 6).
  • tau1, tau2 (baseline thresholds) = tuned per PLM
    Baseline relationship thresholds tuned per PLM to maximize F1 on the test set (Section IV.A.3).
assumptions (4)
  • domain assumption Large language models can infer accurate conceptual types, attributes, and relationships from column headers and sampled cell values.
    Central premise for all three pipeline steps; the paper relies on LLM pretrained knowledge to resolve heterogeneous naming and formatting.
  • domain assumption Schema.org provides a suitable ground-truth conceptual schema for tabular repositories.
    Evaluation and few-shot examples rely on schema.org hierarchy; if this taxonomy is not a valid target, the reported scores lose meaning.
  • ad hoc to paper An attribute whose values predominantly refer to instances of another type indicates an inter-type relationship.
    This rule defines both the method (Section III.D) and the ground truth (Section IV.A.1), making the relationship evaluation partially circular.
  • ad hoc to paper The PTCS metric, which scores inferred hierarchies by longest common subsequence against GT paths, is a valid measure of structural quality.
    Proposed in this paper, not externally validated; the paper acknowledges fine-grained inferred hierarchies reduce PTCS.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Conceptual Schema Inference for Tabular Datasets using Large Language Models." pith.science (2026). https://pith.science/paper/6NWRSG7L

@misc{pith2026250904632,
  author       = {Pith},
  title        = {Pith review of: Conceptual Schema Inference for Tabular Datasets using Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6NWRSG7L}},
  note         = {Machine review of arXiv:2509.04632}
}
read the original abstract

Large collections of tabular data from data lakes, web tables and open data portals often originate from heterogeneous sources, leading to representational inconsistencies. Understanding and organizing such repositories therefore remains a major challenge. While prior work has primarily focused on dataset discovery and exploration, this paper addresses the complementary problem of conceptual schema inference: automatically deriving a conceptual schema that captures entity types, attributes and inter-type relationships directly from raw tables. We propose two large language model (LLM)-based approaches that use only column headers and cell values: GeSI uses generative LLMs to infer hierarchical types and their attributes from table- and column-level semantics, and to integrate them into a global schema that also captures relationships across types; EmSI employs LLM-based table embeddings to group tables by column-level semantics, infer attributes within each group, and construct hierarchical structures from shared attribute patterns. Finally, we report an experimental analysis demonstrating the effectiveness of our approaches in terms of the conciseness and structural quality of the inferred schema components, their scalability to large repositories, and a case study illustrating end-to-end schema inference.

Figures

Figures reproduced from arXiv: 2509.04632 by the authors.

Figure 1
Figure 1. The SI-LLM schema inference pipeline. 3) An end-to-end case study with a visualized concep￾tual schema, illustrating how heterogeneous tables can be integrated into a coherent schema. II. RELATED WORK Schema inference has been employed for different purposes by a variety of communities. Most research on schema inference has been closely associated with a specific data model and its features. For example, schema infe… view at source ↗
Figure 2
Figure 2. Basic prompt for type hierarchy inference. ABS: <ORIGINAL INSTRUCTION> The top-level type directly under "Thing" must not be an abstract type, i.e. <ABSTRACT TYPE SAMPLES>. SLC: <ORIGINAL INSTRUCTION> Fixed second-level type if multiple paths: If the inferred hierarchy of the table contains multiple paths, the second-level type in all paths must be the same fixed type. FET: [Step 1] The following is a table. Your ta… view at source ↗
Figure 4
Figure 4. Prompts in Attribute Inference and Relationship Discovery [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: PTCS of the hier [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 7
Figure 7. Figure 7: End-To-End Example Top-level types. Figure 7a shows the inferred top￾level types with their attributes and relationships. The schema recovers 87% of annotated types overall; at the top level, 5 out of 6 ground-truth top level types are recovered, including Person, Orga…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

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

    cs.DB 2026-08 conditional novelty 6.0 of 10

    A neurosymbolic pipeline automatically constructs executable analytic semantic schemas ("rings") from relational databases, claiming 100% coverage and retrieval pass rates across eight domains.

  2. SINT-Flow: Schema Integration using Large Language Model Workflows

    cs.CL 2026-07 conditional novelty 6.0 of 10

    Five LLM operators arranged into workflows fully automate schema integration, including splitting denormalized multi-entity tables, reaching ≥83% mapping F1 on a new 93-table benchmark.

  3. SciSchema.org: A Multidisciplinary Collection of Schemas for Structured Scientific Process Descriptions

    cs.DL 2026-07 accept novelty 5.5 of 10

    Sixteen expert-annotated scientific-process schemas, built via LLM-assisted human-in-the-loop mining and released in JSON Schema and SHACL, form the first SciSchema.org collection.

Reference graph

Works this paper leans on

38 extracted references · 38 canonical work pages · cited by 3 Pith papers

  1. [1]

    Data lakes: A survey of functions and systems,

    R. Hai, C. Koutras, C. Quix, and M. Jarke, “Data lakes: A survey of functions and systems,” IEEE Trans. Knowl. Data Eng., vol. 35, no. 12, pp. 12 571–12 590, 2023

    work page 2023

  2. [2]

    Web table extraction, retrieval, and augmentation: A survey,

    S. Zhang and K. Balog, “Web table extraction, retrieval, and augmentation: A survey,” ACM Trans. Intell. Syst. Technol. , vol. 11, no. 2, pp. 13:1–13:35, 2020

    work page 2020

  3. [3]

    Governor: Turning open government data portals into interactive databases,

    C. Liu, A. Usta, J. Zhao, and S. Salihoglu, “Governor: Turning open government data portals into interactive databases,” in Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, CHI 2023 , 2023, pp. 415:1–415:16

    work page 2023

  4. [4]

    Dataset discovery and exploration: A survey,

    N. W. Paton, J. Chen, and Z. Wu, “Dataset discovery and exploration: A survey,” ACM Comput. Surv. , vol. 56, no. 4, pp. 102:1–102:37, 2024

    work page 2024

  5. [5]

    Data integra- tion for the relational web,

    M. J. Cafarella, A. Halevy, and N. Khoussainova, “Data integra- tion for the relational web,” Proc. VLDB Endow. , vol. 2, no. 1, p. 1090–1101, aug 2009

    work page 2009

  6. [6]

    Table union search on open data,

    F. Nargesian, E. Zhu, K. Q. Pu, and R. J. Miller, “Table union search on open data,” Proc. VLDB Endow. , vol. 11, no. 7, pp. 813–825, 2018

    work page 2018

  7. [7]

    Seeping semantics: Linking datasets using word embeddings for data discovery,

    R. C. Fernandez, E. Mansour, A. A. Qahtan, A. K. Elmagarmid, I. F. Ilyas, S. Madden, M. Ouzzani, M. Stonebraker, and N. Tang, “Seeping semantics: Linking datasets using word embeddings for data discovery,” in Proceedings of the 26th International Conference on Data Engineering, ICDE 2018 , 2018, pp. 989– 1000

    work page 2018

  8. [8]

    Organizing data lakes for navigation,

    F. Nargesian, K. Q. Pu, E. Zhu, B. G. Bashardoost, and R. J. Miller, “Organizing data lakes for navigation,” in SIGMOD 2020. ACM, 2020, pp. 1939–1950

    work page 2020

Show all 38 references
  1. [9]

    Learning semantic annotations for tabular data,

    J. Chen, E. Jim ´enez-Ruiz, I. Horrocks, and C. Sutton, “Learning semantic annotations for tabular data,” in IJCAI, S. Kraus, Ed. ijcai.org, 2019, pp. 2088–2094

    2019

  2. [10]

    TURL: table understanding through representation learning,

    X. Deng, H. Sun, A. Lees, Y . Wu, and C. Yu, “TURL: table understanding through representation learning,” Proc. VLDB En- dow., vol. 14, no. 3, pp. 307–319, 2020

    2020

  3. [11]

    Profiling relational data: a survey,

    Z. Abedjan, L. Golab, and F. Naumann, “Profiling relational data: a survey,” VLDB J., vol. 24, no. 4, pp. 557–581, 2015

    2015

  4. [12]

    Josie: Overlap set similarity search for finding joinable tables in data lakes,

    E. Zhu, D. Deng, F. Nargesian, and R. J. Miller, “Josie: Overlap set similarity search for finding joinable tables in data lakes,” in SIGMOD 2019, 2019, p. 847–864

    2019

  5. [13]

    Efficient join- able table discovery in data lakes: A high-dimensional similarity- based approach,

    Y . Dong, K. Takeoka, C. Xiao, and M. Oyamada, “Efficient join- able table discovery in data lakes: A high-dimensional similarity- based approach,” in Proceedings of the 26th International Con- ference on Data Engineering,ICDE 2021 , 2021, pp. 456–467

    2021

  6. [14]

    Summarizing semantic graphs: a survey,

    S. Cebiric, F. Goasdou ´e, H. Kondylakis, D. Kotzinos, I. Manolescu, G. Troullinou, and M. Zneika, “Summarizing semantic graphs: a survey,” VLDB J. , vol. 28, no. 3, pp. 295– 327, 2019

    2019

  7. [15]

    A survey on semantic schema discovery,

    K. Kellou-Menouer, N. Kardoulakis, G. Troullinou, Z. Kedad, D. Plexousakis, and H. Kondylakis, “A survey on semantic schema discovery,” The VLDB Journal , vol. 31, p. 675–710, 2022

    2022

  8. [16]

    Schema summarization,

    C. Yu and H. V . Jagadish, “Schema summarization,” in VLDB 2006, 2006, pp. 319–330

    2006

  9. [17]

    Struc- ture inference for linked data sources using clustering,

    K. Christodoulou, N. W. Paton, and A. A. A. Fernandes, “Struc- ture inference for linked data sources using clustering,” Trans. Large Scale Data Knowl. Centered Syst., vol. 19, pp. 1–25, 2015

    2015

  10. [18]

    RDF graph sum- marization for first-sight structure discovery,

    F. Goasdou ´e, P. Guzewicz, and I. Manolescu, “RDF graph sum- marization for first-sight structure discovery,” VLDB J., vol. 29, no. 5, pp. 1191–1218, 2020

    2020

  11. [19]

    Summarizing relational databases,

    X. Yang, C. M. Procopiuc, and D. Srivastava, “Summarizing relational databases,” Proc. VLDB Endow., vol. 2, no. 1, pp. 634– 645, 2009

    2009

  12. [20]

    Schema inference for property graphs,

    H. Lbath, A. Bonifati, and R. Harmer, “Schema inference for property graphs,” in Proceedings of the 24th International Con- ference on Extending Database Technology, EDBT 2021, Nicosia, Cyprus, March 23 - 26, 2021 , 2021, pp. 499–504

    2021

  13. [21]

    Inferring XML schema definitions from XML data,

    G. J. Bex, F. Neven, and S. Vansummeren, “Inferring XML schema definitions from XML data,” in VLDB. VLDB En- dowment, 2007, pp. 998–1009

    2007

  14. [22]

    Inference of concise regular expressions and dtds,

    G. J. Bex, F. Neven, T. Schwentick, and S. Vansummeren, “Inference of concise regular expressions and dtds,” ACM Trans. Database Syst., vol. 35, no. 2, pp. 11:1–11:47, 2010

    2010

  15. [23]

    Para- metric schema inference for massive JSON datasets,

    M. A. Baazizi, D. Colazzo, G. Ghelli, and C. Sartiani, “Para- metric schema inference for massive JSON datasets,” VLDB J., vol. 28, no. 4, pp. 497–521, 2019

    2019

  16. [24]

    An approach for schema extraction of JSON and extended JSON document collections,

    A. A. Frozza, R. dos Santos Mello, and F. de Souza da Costa, “An approach for schema extraction of JSON and extended JSON document collections,” in 2018 IEEE International Conference on Information Reuse and Integration, IRI 2018, 2018, pp. 356–363

    2018

  17. [25]

    Schema discovery in RDF data sources,

    K. Kellou-Menouer and Z. Kedad, “Schema discovery in RDF data sources,” in Conceptual Modeling - 34th International Conference, ER 2015 , ser. Lecture Notes in Computer Science, vol. 9381. Springer, 2015, pp. 481–495

    2015

  18. [26]

    Computing generic abstractions from application datasets,

    N. Barret, I. Manolescu, and P. Upadhyay, “Computing generic abstractions from application datasets,” in EDBT. OpenProceed- ings.org, 2024, pp. 94–107

    2024

  19. [27]

    Incremental schema integration for data wrangling via knowledge graphs,

    J. Flores, K. Rabbani, S. Nadal, C. G ´omez, O. Romero, E. Jamin, and S. Dasiopoulou, “Incremental schema integration for data wrangling via knowledge graphs,” Semantic Web, vol. 15, no. 3, pp. 793–830, 2024

    2024

  20. [28]

    Integrating data lake tables,

    A. Khatiwada, R. Shraga, W. Gatterbauer, and R. J. Miller, “Integrating data lake tables,” Proc. VLDB Endow., vol. 16, no. 4, pp. 932–945, 2022

    2022

  21. [29]

    Enhancing domain modeling with pre-trained large language models: An automated assistant for domain modelers,

    D. Prokop, S. Stenchl ´ak, P. Skoda, J. Kl ´ımek, and M. Necask ´y, “Enhancing domain modeling with pre-trained large language models: An automated assistant for domain modelers,” in Con- ceptual Modeling - 43rd International Conference, ER 2024 , 2024, pp. 235–253

    2024

  22. [30]

    Appli- cation of the tree-of-thoughts framework to llm-enabled domain modeling,

    J. Silva, Q. Ma, J. Cabot, P. Kelsen, and H. A. Proper, “Appli- cation of the tree-of-thoughts framework to llm-enabled domain modeling,” in Conceptual Modeling - 43rd International Confer- ence, ER 2024 , 2024, pp. 94–111

    2024

  23. [31]

    Taxonomy inference for tabular data using large language models,

    Z. Wu, J. Chen, and N. W. Paton, “Taxonomy inference for tabular data using large language models,” in Extended Semantic Web Conference (ESWC), 2025

    2025

  24. [32]

    End-to-end ontology learning with large language models,

    A. Lo, A. Q. Jiang, W. Li, and M. Jamnik, “End-to-end ontology learning with large language models,” in NeurIPS 2024, 2024

    2024

  25. [33]

    Chorus: Foundation models for unified data discovery and exploration,

    M. Kayali, A. Lykov, I. Fountalis, N. Vasiloglou, D. Olteanu, and D. Suciu, “Chorus: Foundation models for unified data discovery and exploration,” Proc. VLDB Endow., vol. 17, no. 8, p. 2104–2114, Apr. 2024

    2024

  26. [34]

    Quantifying lan- guage models’ sensitivity to spurious features in prompt design or: How i learned to start worrying about prompt formatting,

    M. Sclar, Y . Choi, Y . Tsvetkov, and A. Suhr, “Quantifying lan- guage models’ sensitivity to spurious features in prompt design or: How i learned to start worrying about prompt formatting,” in ICLR 2024, 2024

    2024

  27. [35]

    Self-refine: iterative refinement with self-feedback,

    A. Madaan, N. Tandon, P. Gupta, S. Hallinan, L. Gao, S. Wiegr- effe, U. Alon, N. Dziri, S. Prabhumoye, Y . Yang, S. Gupta, B. P. Majumder, K. Hermann, S. Welleck, A. Yazdanbakhsh, and P. Clark, “Self-refine: iterative refinement with self-feedback,” in NeurIPS 2023. Curran Ass...

    2023

  28. [36]

    Matching web tables to dbpedia - a feature utility study,

    D. Ritze and C. Bizer, “Matching web tables to dbpedia - a feature utility study,” in International Conference on Extending Database Technology, 2017

    2017

  29. [37]

    The webdatacommons microdata, rdfa and microformat dataset series,

    R. Meusel, P. Petrovski, and C. Bizer, “The webdatacommons microdata, rdfa and microformat dataset series,” in The Semantic Web – ISWC 2014 . Cham: Springer International Publishing, 2014, pp. 277–292

    2014

  30. [38]

    Google dataset search: Building a search engine for datasets in an open web ecosystem,

    D. Brickley, M. Burgess, and N. Noy, “Google dataset search: Building a search engine for datasets in an open web ecosystem,” in WWW. Association for Computing Machinery, 2019, p. 1365–1375

    2019

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.