Pith. sign in

REVIEW 3 major objections 6 minor 21 references

No single table embedding wins on retrieval, structure, and type at once; quality is multi-dimensional.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

No single table-level embedding model leads across retrieval, structural shuffling, and header-free type detection; quality is multi-dimensional.

T0 review reviewed 2026-07-31 challenge →

load-bearing objection Solid workshop benchmark: the value-shuffle negative is the real technical move, and the “no single model wins” result holds even after you discount the messier TTD axis. the 3 major comments →

arxiv 2607.24130 v1 pith:UMQXJ4CJ submitted 2026-07-27 cs.DB cs.LG

TEmBed-T: A Multi-Dimensional Benchmark for Table-Level Embeddings

Ayeen Poostforoushan , Liane Vogel , Carsten Binnig This is my paper
classification cs.DB cs.LG
keywords table embeddingstable-level representationsbenchmarktable retrievalstructural fidelitytable type detectiondata lakestabular representation learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Table-level embeddings power retrieval, data-lake discovery, and classification, yet most benchmarks treat quality as success on one downstream task, usually retrieval. This paper extends an existing tabular embedding testbed with three complementary tasks that each isolate a different property: cross-domain query-to-table retrieval, structural fidelity under row/column permutation versus value shuffling, and recovery of table type from header-stripped cells. Across five representative encoders the rankings reverse sharply from task to task. The practical claim is that table embedding quality cannot be reduced to retrieval performance alone, so model selection and model building must treat the capability profile as multi-dimensional.

Core claim

When the same model pool is scored on cross-domain retrieval, structural triplet discrimination, and header-independent type detection, no encoder leads on all three axes and strong performance on one property does not transfer to the others. Table-level embedding quality is therefore multi-dimensional rather than reducible to retrieval.

What carries the argument

TEmBed-T: a three-task diagnostic suite that operationalizes cross-domain robustness (heterogeneous query-to-table retrieval), structural fidelity (permutation-positive versus value-shuffled-negative triplets), and header-independent semantic preservation (frozen-probe type classification on header-stripped tables).

Load-bearing premise

The three chosen tasks cleanly isolate the intended properties, and the assumption that row and column order carry no intrinsic meaning does not break the structural test for the tables being evaluated.

What would settle it

Find or train one encoder that simultaneously ranks first (or statistically tied for first) on the retrieval MRR average, the canonical shuffling triplet accuracy, and the type-detection macro-F1 under the paper's own protocols and model pool.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Retrieval leaderboards alone are insufficient for choosing table embedders for data lakes or schema-sensitive workloads.
  • Structure-aware architectures can dominate structural probes while lagging on lexical type recovery, so training objectives must be balanced explicitly.
  • Serialization format (CSV vs markdown) and row budget become first-class evaluation axes rather than fixed preprocessing choices.
  • A genuinely general-purpose table encoder would need competitive scores on all three axes at once; none in the current pool does.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Value-shuffled negatives could become a standard structural unit test for any new table encoder, analogous to how inversion tests are used in sentence embeddings.
  • The same multi-axis idea likely applies at column and row granularity: single-task leaderboards there may also hide capability trade-offs.
  • Cost-quality plots in the paper suggest training distribution and architecture matter more than raw inference compute, which could guide smaller specialized table models.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The manuscript extends the TEmBed tabular-embedding testbed, whose table-level axis was previously limited to a single retrieval task, into a three-task diagnostic suite: (i) query-to-table retrieval across heterogeneous TARGET corpora, with row-count and serialization ablations; (ii) a table-shuffling triplet protocol that contrasts permutation positives with a novel value-shuffled negative that preserves each column's token multiset while destroying row integrity; and (iii) table type detection on header-stripped WDC Schema.org tables using frozen-embedding probes (XGBoost, MLP, KNN). Evaluating five approaches (three text-serialization transformers, the structure-aware HyTrel, and a HashingVectorizer lexical baseline), the authors find that per-task rankings diverge sharply — GritLM leads retrieval, HyTrel dominates shuffling, Hashing leads TTD under XGBoost — and conclude that no single embedding approach excels across all evaluated properties, so table-level embedding quality cannot be reduced to retrieval alone.

Significance. If the results hold, this is a useful diagnostic complement to TEmBed with several concrete strengths. The value-shuffled negative (T−) is a genuine methodological contribution over prior permutation probes (Observatory, Younes et al.): because the negative preserves the exact token multiset, a bag-of-words encoder cannot pass by token overlap, closing a real loophole in existing invariance benchmarks. The experimental apparatus is solid for a workshop paper: bootstrapped 95% CIs for retrieval and shuffling, a 15-variation shuffling grid with per-dataset breakdowns and silhouette scores that cleanly separate failure modes (Hashing at exactly 0 vs. transformers systematically inverted), and a public artifact. The cross-domain retrieval reversals (GritLM on Wikipedia corpora, Granite-R2 on Spider) are practically informative. The core comparative finding — that the structure-aware encoder and the text encoders diverge dramatically across axes — is well supported and robust to my concerns below.

major comments (3)
  1. [§3.4 / Table 6] §3.4 states: 'Classifier choice affects absolute scores but not approach rankings, which are consistent across probe types.' This is contradicted by the paper's own Table 6 and by Appendix A.3, which says Hashing 'shows a pronounced classifier dependence: it leads under XGBoost yet trails all transformer encoders under MLP and KNeighbors.' Under XGBoost, Hashing is 1st (macro-F1 0.9252); under MLP and KNN, GritLM (md) is 1st (0.9375, 0.8758) and Hashing falls behind all three transformers. The sentence in §3.4 is simply false as written and must be corrected; the appendix already has the right characterization.
  2. [Table 7 (Appendix A.4)] The caption claim 'No approach ranks first on more than one task' — which supports the headline 'no single model excels across all tasks' in the abstract and §3.5 — holds only because TTD is ranked using the XGBoost probe. Using the MLP or KNN columns of Table 6, GritLM ranks first on both Retrieval and TTD, so an approach does rank first on two tasks. Since the choice of probe is arbitrary (all three are trained in the pipeline), the headline pattern is probe-dependent. This does not undermine the deeper finding (GritLM vs. HyTrel diverge sharply, and HyTrel is last on retrieval and TTD while first on shuffling), but Table 7 should either report ranks per probe, aggregate over probes, or soften the claim to what is probe-invariant.
  3. [§2.3 / §3.4] TTD is motivated as isolating 'header-independent semantic preservation': stripping headers 'forces the embedding to recover type from cell values.' But header stripping removes only one shortcut; §3.4 itself concedes the result 'confirm[s] that table type is recoverable from surface token patterns alone,' and Hashing's XGBoost win demonstrates that a purely lexical bag-of-words representation tops this axis. So the third axis does not cleanly separate retained table semantics from class-typical token frequencies — the same confound the value-shuffled negative was designed to eliminate in §2.2. The suite's framing as measuring 'three complementary properties' should be tempered, or a control added (e.g., probing with class-frequency-matched token distributions, or reporting the Hashing-vs-encoder gap explicitly as the lexical-vs-semantic margin).
minor comments (6)
  1. [§2 opening / §2.1] The introduction of the task says retrieval is evaluated 'across seven heterogeneous corpora [8],' but §2.1 lists five TARGET corpora (FeTaQA, TabFact, OTT-QA, Spider2, BIRD) and Figure 1 shows five. Please reconcile the count.
  2. [Table 5 (Appendix A.2)] HyTrel's per-dataset shuffling accuracy ranges from 1.0000 (ECB, TabFact) to 0.3824 (CKAN). The ECB case is discussed as the clean diagnostic, but the CKAN collapse for the structure-aware encoder is unexplained and arguably the more interesting failure. A sentence on what distinguishes CKAN (e.g., table size distribution, numeric density) would help.
  3. [Figure 9 (Appendix A.4)] Figure 9 describes Hashing as having 'high embedding cost' attributed to its 32,768-dimensional sparse vectors. A HashingVectorizer is typically orders of magnitude cheaper than transformer encoding; please clarify what the measured time includes (I/O, probe fitting, serialization) so this point is interpretable.
  4. [§2.3] Typo: 'retained task-information information'.
  5. [§3.1 / Figure 6] §3.1 says shuffling CIs are computed by resampling per-triplet, but Figure 6's caption says silhouette CIs are 'pooled across datasets.' Please state the resampling unit consistently.
  6. [Table 1 (Appendix A.1)] Table 1 omits the Spider validation and test splits 'for space' while the main text (Figure 1) aggregates over all three; a footnote with the omitted means would make the serialization claim fully checkable.

Circularity Check

0 steps flagged

No circularity: empirical multi-task benchmark; rankings are measured outcomes, not forced by definition or self-citation chains.

full rationale

TEmBed-T is a diagnostic benchmark extension, not a first-principles derivation. Its central claim—that model rankings diverge across retrieval, shuffling, and type-detection and that no single encoder dominates—is obtained by running frozen encoders on external corpora (TARGET, WDC Schema.org, CKAN/ECB) and reporting MRR, triplet accuracy/silhouette, and probe macro-F1. Those scores are not algebraically identical to any fitted input, nor are they defined as the claim. Inheritance from TEmBed, TARGET, HyTrel, and Observatory is ordinary benchmark scaffolding (task formats, datasets, model pool) and does not load-bear via an unverified uniqueness theorem or self-defined identity. The value-shuffled negative is a designed control, not a circular reduction of the structural-fidelity claim. Construct-validity concerns about TTD rewarding lexical baselines are correctness/interpretation issues, not circularity. Score 0; steps empty.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

As a benchmark paper, load-bearing content is mostly domain assumptions about which properties matter and how to operationalize them, plus experimental knobs (perturbation magnitudes, row caps, subsample sizes, hashing dimension). No new physical entities; the ‘invented’ pieces are diagnostic task constructs. Central claim rests on those constructs faithfully measuring distinct capabilities and on the chosen model/dataset pool being informative.

free parameters (5)
  • positive/negative perturbation magnitudes (d+, f−, d−) = v0: 0.75 / 0.7 / 0.8
    Hand-chosen fractions controlling shuffling difficulty; canonical v0 uses d+=0.75, f−=0.7, d−=0.8. Rankings could shift under other regimes though the grid explores several.
  • retrieval row_limit = 0 and 100
    Headers-only vs first 100 rows is an experimental choice that changes absolute MRR and relative gains (e.g., MiniLM saturates more).
  • HashingVectorizer embedding dimension = 32768
    Fixed sparse dimension for the lexical baseline; affects cost and possibly probe behavior on TTD.
  • TTD subsample size per class = 800/100 × 10 classes
    800 train / 100 test per 10 Schema.org classes chosen for efficiency and balanced macro-F1; full corpus rankings unreported.
  • table size windows (default/BIG/SMALL) = ≤100×30, ≤500×80, ≤20×8
    Filters before triplet generation; SMALL excludes some corpora (CKAN/ECB), shaping size-ablation averages.
axioms (5)
  • domain assumption Row and column order carry no intrinsic semantics, so row/column permutation is a structure-preserving positive.
    Stated in §2.2 footnote; fails for temporally ordered or ranked tables, which would flip the intended positive.
  • ad hoc to paper Cross-domain robustness, structural fidelity, and header-independent semantic preservation are the complementary properties needed for table-level embedding quality.
    Motivating triad in §2; other properties (joinability, numeric distribution fidelity, hierarchical schema) are not included.
  • domain assumption Frozen probe classifier performance on header-stripped tables measures information retained in the embedding rather than probe skill when rankings are stable across XGBoost/MLP/KNN.
    §2.3 and §3.4; standard probing assumption, weakened by Hashing’s classifier-dependent lead under XGBoost only.
  • domain assumption Value-shuffled negatives that preserve per-column token multisets isolate structural encoding from lexical overlap.
    Core of §2.2 design; assumes encoders do not use other non-structural cues that still differ between T+ and T− (e.g., local n-gram order inside cells).
  • standard math Cosine distance on encoder outputs is an appropriate geometry for triplet structural tests and retrieval ranking.
    Used throughout §2–3; standard in embedding evaluation but not uniquely justified for HyTrel vs. dense text spaces.
invented entities (2)
  • TEmBed-T three-task diagnostic suite no independent evidence
    purpose: Package retrieval, shuffling, and type detection as a multi-dimensional table-level benchmark extension of TEmBed.
    Named contribution; existence is definitional to the paper rather than an external natural kind.
  • Value-shuffled structural negative (T−) no independent evidence
    purpose: Provide a hard negative that matches token multisets per column while breaking row-wise associations.
    Protocol object introduced to fix bag-of-words loopholes in permutation-only probes; validated only inside this evaluation design.

reviewed 2026-07-31 · how reviews work

0 comments
Cite this review

Pith. "Pith review of TEmBed-T: A Multi-Dimensional Benchmark for Table-Level Embeddings." pith.science (2026). https://pith.science/paper/UMQXJ4CJ

@misc{pith2026260724130,
  author       = {Pith},
  title        = {Pith review of: TEmBed-T: A Multi-Dimensional Benchmark for Table-Level Embeddings},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UMQXJ4CJ}},
  note         = {Machine review of arXiv:2607.24130}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Tabular data is the dominant structured-data modality, and learning table representations has become a core research direction. Table-level embeddings in particular underpin a wide range of applications, including table retrieval, data lake discovery, and table classification. Despite their importance, there is still limited understanding of how different embedding approaches behave across tasks, making systematic evaluation and analysis essential. In this work, we introduce a systematic evaluation of table-level embeddings that captures several complementary properties required for downstream effectiveness. We realize this evaluation by extending TEmBed, a recently proposed testbed for tabular embeddings, whose table-level coverage is currently limited to a single retrieval task. An empirical study over the TEmBed model pool confirms that no single model excels across all tasks, demonstrating that table-level embedding quality cannot be reduced to retrieval alone.

Figures

Figures reproduced from arXiv: 2607.24130 by Ayeen Poostforoushan, Carsten Binnig, Liane Vogel.

Figure 1
Figure 1. Figure 1: Table Retrieval: MRR@10 per dataset at row_limit = 100. Spider: aggregated over its three splits. Dataset hetero￾geneity drives domain-specific ranking reversals between leading approaches. Error bars show bootstrapped 95% CIs. Granite-R2 (md) GritLM (md) Hashing HyTrel MiniLM (md) 0.0 0.2 0.4 0.6 0.8 1.0 MRR@10 0.56 0.70 0.56 0.72 0.21 0.32 0.02 0.02 0.50 0.54 Schema only (rl=0) Schema + 100 rows [PITH_F… view at source ↗
Figure 2
Figure 2. Figure 2: compares retrieval performance under schema-only and first-100-rows configurations, averaged over datasets. All ap￾proaches improve when row content is provided, confirming that cell values carry signal beyond headers. All transformer encoders gain substantially from row content, while MiniLM shows a markedly smaller gain. This asymmetry indicates that MiniLM’s lower-capacity representation already saturat… view at source ↗
Figure 3
Figure 3. Figure 3: Per-task scores for all five approaches, averaged over each task’s datasets. Markdown serialization was used for [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Table Type Detection: Macro-F1 per approach [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Table Shuffling: Accuracy heatmap across all variations v0–v14 (defined in Table 2) and approaches, averaged over [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Table Shuffling: Silhouette score per approach, aver [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Table Shuffling: Row vs. column perturbation ac [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: Cost–quality trade-off: Quality is the mean of each [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

21 extracted references · 3 linked inside Pith

  1. [1]

    Bhat, Vignesh P, Vishwajeet Kumar, Yushu Yang, Bhavani Iyer, Abraham Daniels, Rudra Murthy, Ken Barker, Martin Franz, Madison Lee, Todd Ward, Salim Roukos, David Cox, Luis A

    Parul Awasthy, Aashka Trivedi, Yulong Li, Meet Doshi, Riyaz A. Bhat, Vignesh P, Vishwajeet Kumar, Yushu Yang, Bhavani Iyer, Abraham Daniels, Rudra Murthy, Ken Barker, Martin Franz, Madison Lee, Todd Ward, Salim Roukos, David Cox, Luis A. Lastras, Jaydeep Sen, and Radu Florian. 2025. Granite Embedding R2 Models.CoRRabs/2508.21085 (2025). https://doi.org/10...

    Pith/arXiv arXiv doi:10.48550/arxiv.2508.21085 2025

  2. [2]

    Gilbert Badaro, Mohammed Saeed, and Paolo Papotti. 2023. Transformers for Tabular Data Representation: A Survey of Models and Applications.Trans. Assoc. Comput. Linguistics11 (2023), 227–249. https://doi.org/10.1162/TACL_A_00544

    doi:10.1162/tacl_a_00544 2023

  3. [3]

    Pei Chen, Soumajyoti Sarkar, Leonard Lausen, Balasubramaniam Srinivasan, Sheng Zha, Ruihong Huang, and George Karypis. 2023. HyTrel: Hypergraph- enhanced Tabular Data Representation Learning. InAdvances in Neural Infor- mation Processing Systems 36: Annual Conference on Neural Information Pro- cessing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, Dece...

    2023

  4. [4]

    Tianqi Chen and Carlos Guestrin. 2016. XGBoost: A Scalable Tree Boosting System. InProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, August 13-17, 2016, Balaji Krishnapuram, Mohak Shah, Alexander J. Smola, Charu C. Aggarwal, Dou Shen, and Rajeev Rastogi (Eds.). ACM, 785–794. https:/...

    arXiv 2016

  5. [5]

    Wenhu Chen, Ming-Wei Chang, Eva Schlinger, William Yang Wang, and William W. Cohen. 2021. Open Question Answering over Tables and Text. In9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net. https://openreview.net/forum? id=MmCRswl1UYl

    2021

  6. [6]

    Wenhu Chen, Hongmin Wang, Jianshu Chen, Yunkai Zhang, Hong Wang, Shiyang Li, Xiyou Zhou, and William Yang Wang. 2020. TabFact: A Large- scale Dataset for Table-based Fact Verification. In8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net. https://openreview.net/forum?id=rkeJRhNYDH

    2020

  7. [7]

    Tianji Cong, Madelon Hulsebos, Zhenjie Sun, Paul Groth, and H. V. Jagadish

  8. [8]

    Parameswaran, and Madelon Hulsebos

    Xingyu Ji, Parker Glenn, Aditya G. Parameswaran, and Madelon Hulsebos. 2024. TARGET: Benchmarking Table Retrieval for Generative Tasks.NeurIPS 2024 Third Table Representation Learning Workshop(2024). https://openreview.net/ pdf?id=gGGvnjFUfL

    2024

  9. [9]

    Jinyang Li, Binyuan Hui, Ge Qu, Jiaxi Yang, Binhua Li, Bowen Li, Bailin Wang, Bowen Qin, Ruiying Geng, Nan Huo, et al . 2024. Can llm already serve as a database interface? a big bench for large-scale database grounded text-to-sqls. Advances in Neural Information Processing Systems36 (2024)

    2024

  10. [10]

    Niklas Muennighoff, SU Hongjin, Liang Wang, Nan Yang, Furu Wei, Tao Yu, Amanpreet Singh, and Douwe Kiela. 2025. Generative representational instruc- tion tuning. InThe Thirteenth International Conference on Learning Representa- tions

    2025

  11. [11]

    Niklas Muennighoff, Nouamane Tazi, Loïc Magne, and Nils Reimers. 2023. Mteb: Massive text embedding benchmark. InProceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics. 2014–2037

    2023

  12. [12]

    Linyong Nan, Chiachun Hsieh, Ziming Mao, Xi Victoria Lin, Neha Verma, Rui Zhang, Wojciech Kryscinski, Hailey Schoelkopf, Riley Kong, Xiangru Tang, Mutethia Mutuma, Ben Rosand, Isabel Trindade, Renusree Bandaru, Jacob Cun- ningham, Caiming Xiong, and Dragomir R. Radev. 2022. FeTaQA: Free-form Table Question Answering.Trans. Assoc. Comput. Linguistics10 (20...

    doi:10.1162/tacl_a_00446 2022

  13. [13]

    Pedregosa, G

    F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cour- napeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine Learning in Python.Journal of Machine Learning Research12 (2011), 2825–2830

    2011

  14. [14]

    Ralph Peeters, Alexander Brinkmann, and Christian Bizer. 2024. The Web Data Commons Schema.org Table Corpora. InCompanion Proceedings of the ACM on Web Conference 2024, WWW 2024, Singapore, Singapore, May 13-17, 2024, Tat-Seng Chua, Chong-Wah Ngo, Roy Ka-Wei Lee, Ravi Kumar, and Hady W. Lauw (Eds.). ACM, 1079–1082. https://doi.org/10.1145/3589335.3651441

    arXiv 2024

  15. [15]

    Kavitha Srinivas, Julian Dolby, Ibrahim Abdelaziz, Oktie Hassanzadeh, Har- sha Kokel, Aamod Khatiwada, Tejaswini Pedapati, Subhajit Chaudhury, and Horst Samulowitz. 2023. LakeBench: Benchmarks for Data Discovery over Data Lakes.CoRRabs/2307.04217 (2023). https://doi.org/10.48550/ARXIV.2307.04217 arXiv:2307.04217

    Pith/arXiv arXiv doi:10.48550/arxiv.2307.04217 2023

  16. [16]

    D’Souza, Sola Shirai, Oktie Has- sanzadeh, and Horst Samulowitz

    Liane Vogel, Kavitha Srinivas, Niharika S. D’Souza, Sola Shirai, Oktie Has- sanzadeh, and Horst Samulowitz. 2026. Towards Universal Tabular Embed- dings: A Benchmark Across Data Tasks.CoRRabs/2604.21696 (2026). https: //doi.org/10.48550/ARXIV.2604.21696 arXiv:2604.21696

    Pith/arXiv arXiv doi:10.48550/arxiv.2604.21696 2026

  17. [17]

    Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou. 2020. MiniLM: Deep Self-Attention Distillation for Task-Agnostic Compression of Pre- Trained Transformers. InAdvances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, Hugo Larochelle,...

    2020

  18. [18]

    Ali Younes, Saeed Ghoorchian, Maximilian Schambach, and Johannes Höhne

  19. [19]

    Tao Yu, Rui Zhang, Kai Yang, Michihiro Yasunaga, Dongxu Wang, Zifan Li, James Ma, Irene Li, Qingning Yao, Shanelle Roman, Zilin Zhang, and Dragomir R. Radev. 2018. Spider: A Large-Scale Human-Labeled Dataset for Complex and Cross-Domain Semantic Parsing and Text-to-SQL Task. InProceedings of the 2018 Conference on Empirical Methods in Natural Language Pro...

    arXiv 2018

  20. [2023]

    VLDB Endow.17, 4 (2023), 849–862

    Observatory: Characterizing Embeddings of Relational Tables.Proc. VLDB Endow.17, 4 (2023), 849–862. https://doi.org/10.14778/3636218.3636237

    arXiv 2023

  21. [2026]

    Unified Evaluation of Table Embedding Methods Across Multiple Bench- mark Scenarios.3rd DATA-FM Workshop at ICLR(2026)

    2026

This paper was first reviewed by grok-4.5 on July 31, 2026.