Pith. sign in

REVIEW 4 major objections 4 minor 55 references

Column annotation can be made more accurate and fully interpretable by materializing it as an LLM-induced semantic decision tree with evolving local classifiers.

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 ↗

T0 review · deepseek-v4-flash

2026-08-01 03:00 UTC pith:QZ44PQFF

load-bearing objection SymCA is a genuinely new interpretable column-annotation pipeline, but the headline accuracy margin is undercut by an undisclosed choice about whether it used TabEmb's embeddings in the main experiments. the 4 major comments →

arxiv 2607.25228 v1 pith:QZ44PQFF submitted 2026-07-28 cs.CL

Interpretable Column Annotation with LLM-Symbolized Decision Process Materialization

Mengqi Wang (1) , Jianwei Wang (1) , Qing Liu (2) , Xiwei Xu (2) , Zhenchang Xing (2) , Michael Bain (1) , Liming Zhu (2) , Wenjie Zhang (1) ((1) UNSW Sydney
show 2 more authors
(2) Data61 CSIRO)
This is my paper
classification cs.CL
keywords column type annotationcolumn property annotationsemantic skeletonhypernym hierarchyLLMinterpretabilityrandom foresttable understanding
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 paper sets out to show that column type and column property annotation — deciding what a table column means and how two columns are related — does not have to choose between accuracy and interpretability. It proposes SymCA, which first asks a large language model to organize the target labels into a hypernym-inspired tree, then builds a small random-forest classifier at each internal node over explicit, named evidence operators. Each column is predicted by routing it down this tree, so every answer comes with a visible root-to-leaf path and the exact operator values that justified it. The authors report that this approach beats the strongest existing baselines by an average of 6.42% in Micro-F1 and 11.03% in Macro-F1 across five public benchmarks, and that it degrades more gracefully than neural baselines when training data is scarce. If correct, the result matters because it offers data integration and schema matching a way to get both higher accuracy and human-auditable reasoning.

Core claim

On its own terms, the paper's central claim is that column annotation can be materialized as a global-to-local symbolic decision process, and that doing so improves both accuracy and interpretability. SymCA induces a tree-structured semantic skeleton over the label space by asking an LLM to group labels by hypernym-inspired meaning, selecting a robust skeleton via a Minimum Bayes Risk consensus over multiple generated candidates. It then turns each internal node of that tree into a predictive substrate: a random forest that classifies columns among the node's children using explicit scalar operators (value profiles, surface patterns, contextual statistics, and semantic similarity). Over seve

What carries the argument

The central object is the hypernym-inspired semantic skeleton: a tree over the label vocabulary, induced by LLM reasoning and selected by MBR consensus, in which each internal node groups semantically related labels. Each internal node is carried by a predictive substrate — a random forest trained on a node-specific set of scalar, deterministic operators covering value profiles, surface patterns, contextual statistics, and semantic similarity. The LLM proposes modifications to these operators based on a substrate evidence profile of confusions and misclassifications, while a UCB-style multi-armed bandit scheduler decides which substrate gets the next retraining trial. This machinery decompos

Load-bearing premise

The reported margin over baselines assumes all methods were trained and tuned under the same subsampled training conditions; if a baseline was disadvantaged by the reduced training scale or by not being retuned on the same validation splits, the 6.42% and 11.03% advantages would shrink.

What would settle it

Retrain every baseline on the exact subsampled training folds used for SymCA, tuning each baseline's hyperparameters on the same official validation split; if the strongest baseline closes the 6.42-point Micro-F1 gap, the comparative claim fails. Separately, after skeleton induction, scramble the leaf labels while keeping the tree structure: if accuracy stays high, the hierarchy is not carrying the claimed semantic signal.

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

If this is right

  • Every prediction becomes inspectable: each annotation is a root-to-leaf path with per-node probabilities and named operators, so users can see why a column was labeled as a URL rather than an email address.
  • The label hierarchy is reusable: because the skeleton is induced from label names alone, a new label vocabulary can be organized without retraining encoders, which suits evolving ontologies.
  • Low-resource settings benefit most: on training subsets of 25% to 100%, SymCA degrades more gracefully than representation-learning baselines, suggesting it fits annotation tasks with limited labeled columns.
  • Rare labels gain disproportionately: the larger Macro-F1 improvement indicates the hierarchy helps infrequent classes survive class imbalance.
  • LLM cost stays bounded: the LLM is invoked during skeleton induction and operator evolution rather than per prediction, so token consumption on large benchmarks is lower than direct prompting.

Where Pith is reading between the lines

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

  • Because the skeleton depends only on label names, the same global-to-local machinery could be applied to other structured label spaces — such as event typing, relation extraction, or ontology alignment — where an LLM can organize labels without task-specific training; the paper does not test this transfer.
  • The operator-evolution loop can be viewed as automated feature engineering: operators like url_like_ratio are explicit and potentially portable across datasets, a transfer the paper leaves unexamined.
  • The substrate evidence profiles identify exactly which child pairs are confused, so a natural extension is an active-learning loop that requests labels only for regions the tree finds hardest; the paper does not explore this.
  • The MBR consensus is computed over five candidates; a direct extension would measure how the selected skeleton and downstream accuracy change as the candidate count grows, which the paper does not report.

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

4 major / 4 minor

Summary. The paper proposes SymCA, a two-stage column annotation framework. A global skeleton induction stage uses LLM-generated candidate hypernym trees, scores them with a MiniLM-based semantic cost plus a structural regularizer, and selects a consensus tree via MBR decoding. A local substrate evolution stage trains a random forest at each internal node over explicit symbolic operators, lets an LLM propose node-specific operator insert/update/remove operations, and schedules retraining with a UCB-style multi-armed bandit. The authors evaluate Micro-F1 and Macro-F1 on five CTA/CPA benchmarks and report that SymCA outperforms the strongest baseline by 6.42% Micro-F1 and 11.03% Macro-F1 on average, along with ablations, robustness to different LLM backbones, training-set size, evolution rounds, an interpretability case study, and efficiency comparisons.

Significance. If the empirical claims are fully supported, SymCA is a meaningful contribution to column annotation. It offers a transparent root-to-leaf decision path with local random forests and named operators, a label-hierarchy induction method designed to reduce LLM generation variance, and evidence that hierarchical decomposition plus adaptive operators can improve accuracy, especially under limited supervision. The source-code link is a strength, and the two lemmas in Section 3 are simple but correctly proved. However, the headline comparison is currently under-specified in ways that affect the validity of the claimed 6.42%/11.03% advantage, particularly the unresolved use of TabEmb features inside SymCA and the lack of a demonstrated baseline-retuning protocol on the subsampled ST splits.

major comments (4)
  1. [§4.1 Implementation details; Table 2] The implementation details state that each internal-node random forest 'may additionally incorporate TabEmb [16] features if provided when selected through LLM-guided evolution.' Since TabEmb is itself one of the baselines in Table 2, the paper must disclose whether the Table 2 runs used this optional feature. If enabled, SymCA is not being compared purely as the symbolic skeleton-plus-evolution pipeline: it can borrow a competitive baseline's learned representation. The reported 6.42% Micro-F1 / 11.03% Macro-F1 advantage over baselines would then not isolate the symbolic decision process. The 'w/o skeleton' and 'w/o evolution' ablations in Table 3 do not resolve this, because they would inherit the same feature flag. Please report the main results with the feature disabled and enabled, or state explicitly that it was disabled for all Table 2 and Table 3 runs.
  2. [§4.1 Datasets; Table 1] The ST-CTA and ST-CPA training sets are subsampled 'while preserving full label coverage,' but the paper does not state whether TabEmb, Watchog, Doduo, and REVEAL+ were retrained and retuned on exactly these subsampled folds or trained on the original training sets. The central average-gain claim can be affected by this asymmetry: methods with different data requirements or hyperparameter sensitivity may be under-tuned on the reduced folds. Additionally, T2D-CPA has only 63 test instances, so Macro-F1 differences of about 10 points may be within noise. The central claim needs (a) an explicit baseline-retuning protocol on the same subsampled splits, and (b) confidence intervals or multiple-seed results, at least for T2D-CPA.
  3. [§4.1/§4.3; Tables 2-4] The main tables report single runs. Module 1 uses stochastic LLM generation and Module 2 uses LLM-proposed operator modifications with UCB-based exploration, so run-to-run variance is expected. Without variance estimates, the 'robust' claims in Table 4 and Figure 4 cannot be evaluated, and small differences such as DeepSeek 79.56 vs SymCA 79.51 Macro-F1 on ST-CTA are uninterpretable. Please report mean±std over at least 3-5 independent runs for Tables 2-4 and Figure 4, or justify single-run determinism.
  4. [§3.1.2 and Table 3] The 'w/o skeleton' ablation removes the entire hierarchical decision process and replaces it with a single flat evolving classifier over all labels. This changes both the architecture and the source of the label-structure signal, because the MiniLM-based semantic cost in Cost_sem(T) is also removed. The ablation therefore does not isolate the contribution of the induced hypernym skeleton per se. The conclusion that 'semantic decomposition' is responsible for the gain would be better supported by a control that keeps the same skeleton but replaces the MBR/label-affinity selection with a simpler or independently sourced hierarchy, or by an ablation that varies only the skeleton source.
minor comments (4)
  1. [§3.2.1] The per-label isotonic calibration function g_y is not described concretely. Please specify how g_y is fit, on which split it is fit, and how it interacts with the hierarchical path scores.
  2. [§4.1 Implementation details] The relevance and redundancy thresholds epsilon_v=0.01 and gamma_v=0.92 are named, but the gate criterion (executability, relevance, non-redundancy) is only cited to [34]. Please provide explicit formulas or a precise procedural definition.
  3. [Table 2] The 'Deepseek (0-shot)' row contains formatting irregularities in the T2D-CPA and average columns; please clean the table so all numbers are consistently aligned.
  4. [Figure 4] The text says non-monotonicity is expected because evolution states are selected using validation evidence rather than test performance. Please clarify that the plotted curves are per-round test F1 values and that the final reported model is the validation-selected state.

Circularity Check

0 steps flagged

No significant circularity: SymCA's reported gains are an empirical comparison, not a derivation from its own fitted inputs.

full rationale

I walked SymCA's derivation chain. Module 1 selects a label-space skeleton via MBR consensus over LLM-generated candidates weighted by a label-embedding affinity cost; this is an unsupervised structure over label names and is not defined in terms of the downstream F1 predictions. Module 2 trains per-node random forests on table instances with symbolic operators; no fitted parameter is renamed as a prediction, and no equation reduces the claimed outcome to an input. The paper's self-citations appear in related-work and motivation passages but are not load-bearing: the method's key components (MBR, random forests, isotonic calibration, benchmark datasets) cite external, standard sources. The sentence about optionally incorporating TabEmb features is an unresolved experimental-reporting ambiguity that could affect the fairness of the baseline comparison, but it does not make any stated derivation or claimed result equivalent to its own inputs by construction. The ablation design also does not equate the skeleton benefit to the skeleton-selection objective; it is a structural comparison, not a tautology. I therefore find no derivation-level circularity.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 1 invented entities

SymCA introduces no new physical entities, but its central empirical result rests on several unstated or unreported hyperparameters (κ, λ, η, β, α, γ, ε, evolution rounds, retraining budget) and on domain assumptions about label-embedding similarity, LLM candidate quality, and fair subsampling. These are the choices that the claimed 6.42%/11.03% margin depends on.

free parameters (6)
  • granularity target κ and strength λ in Cost_str(T) = not stated in main text
    λ>0 and κ proportional to |L| shape the skeleton-selection objective; values are chosen by the authors and not reported, so the MBR selection depends on untested constants.
  • weight concentration η in Eq. 1 = not stated
    Controls how sharply candidate weights concentrate; affects which skeleton is selected through the MBR objective.
  • exploration strength β in UCB selection = not stated
    Balances exploration and exploitation in substrate selection; affects how the retraining budget is spent and thus the evolution gains.
  • reward gain scaling α = not stated
    Scales accuracy-gain bonus in bandit rewards; used in Lemma 2 and the scheduler.
  • redundancy threshold γ_v and relevance threshold ε_v = 0.92 / 0.01
    Gates that accept or reject LLM-proposed operators; thresholds chosen by hand with no sensitivity analysis.
  • number of evolution rounds R, max operator modifications, and retraining budget B = 5 rounds / 4 mods / 6 trials
    Empirically chosen budget; Figure 4 shows convergence on two datasets, but the main results depend on these constants.
axioms (5)
  • domain assumption all-MiniLM-L6-v2 cosine similarity is a valid semantic-affinity measure for label names (§3.1.2)
    The entire skeleton-quality cost, and hence skeleton selection, treats clipped cosine similarity of label-name embeddings as semantic affinity; no calibration against human judgment.
  • domain assumption LLM-generated hypernym trees (K=5 candidates) span the space of reasonable label hierarchies (§3.1.1)
    MBR consensus can only select a skeleton among generated candidates; if the LLM's candidate set is biased or incomplete, the consensus inherits the bias.
  • domain assumption Random-forest feature importances over the operator set provide a faithful interpretation of local decisions (§4.4)
    The interpretability analysis presents operator importance as evidence of how predictions are made; RF importances are known to be imperfect proxies.
  • domain assumption The subsampled SOTAB training sets preserve the label-conditional data distributions needed for fair CTA/CPA comparisons (§4.1)
    Subsampling 'while preserving full label coverage' can still change class priors and value distributions; baseline comparisons assume this does not disadvantage them.
  • domain assumption DeepSeek-V4-Pro is treated as a fixed but fully accessible LLM oracle (§3.1, §3.2, §4.1)
    All LLM-dependent outputs (skeletons, operator proposals) inherit whatever instabilities or update drift that model may have; reproducibility assumes the model remains accessible and stable.
invented entities (1)
  • predictive substrate S_v^(r) = (O_v^(r), f_v^(r), H_v^(<r)) independent evidence
    purpose: Executable per-node classifier plus operator set and evolution memory that routes instances down the skeleton.
    The substrate is a defined algorithmic object whose behavior is falsifiable through the reported test performance and the walked-through URL example; it is not a new physical entity.

pith-pipeline@v1.3.0-alltime-deepseek · 19153 in / 9010 out tokens · 77181 ms · 2026-08-01T03:00:55.163640+00:00 · methodology

0 comments
read the original abstract

Column annotation (CA), including column type annotation (CTA) and column property annotation (CPA), aims to identify the meanings of table columns and the semantic relationships among them. Recent CA methods usually use various neural models to learn column representations and directly map them to label categories, thereby (1) sacrificing model interpretability and adaptivity, and (2) overlooking rich label semantics and ultimately limiting accuracy. To address these limitations, we propose SymCA, an LLM-empowered interpretable CA framework that materializes column annotation as a global-to-local symbolic decision process. SymCA consists of two components: (1) global skeleton induction, which constructs a semantic skeleton over the label space, and (2) local substrate evolution, which evolves predictive substrates within the skeleton. Specifically, to exploit label semantics while preserving an interpretable decision process, the global skeleton induction module leverages LLMs to generate candidate hypernym-inspired tree-structured semantic skeletons and employs a Minimum Bayes Risk (MBR)-based consensus strategy to select a robust skeleton against generation variance. Since different internal nodes require different evidence to distinguish among their child nodes, the local substrate evolution module materializes each internal node as an executable and evolvable predictive substrate. Over multiple evolution rounds, each substrate trains an interpretable random forest classifier with the current operator set, leverages the LLM to propose node-specific operator modifications, and uses an exploration-exploitation strategy to prioritize promising substrates. Extensive experiments demonstrate that SymCA is accurate, robust, and interpretable, outperforming the strongest baselines by an average of 6.42% in Micro-F1 and 11.03% in Macro-F1.

Figures

Figures reproduced from arXiv: 2607.25228 by (2) Data61, CSIRO), Jianwei Wang (1), Liming Zhu (2), Mengqi Wang (1), Michael Bain (1), Qing Liu (2), Wenjie Zhang (1) ((1) UNSW Sydney, Xiwei Xu (2), Zhenchang Xing (2).

Figure 1
Figure 1. Figure 1: A motivating example of column annotation with [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of SymCA. Module 1 induces a global semantic skeleton over the label space, while Module 2 materializes [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Evolution round-level test Micro-/Macro-F1. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Example of CTA prediction. Micro-/Macro-F1 reaching 81.79%/79.49%. On ST-CPA, most im￾provements occur within the first two rounds, after which both metrics remain stable. Such non-monotonicity is expected because evolution states are selected using validation evidence rather than test performance. Overall, both tasks converge within five rounds, supporting the evolution round used in the main experiments.… view at source ↗
Figure 6
Figure 6. Figure 6: Efficiency comparison across CTA and CPA datasets. [PITH_FULL_IMAGE:figures/full_fig_p013_6.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

55 extracted references · 7 linked inside Pith

  1. [1]

    Nora Abdelmageed, Jiaoyan Chen, Vincenzo Cutrona, Vasilis Efthymiou, Ok- tie Hassanzadeh, Madelon Hulsebos, Ernesto Jiménez-Ruiz, Juan Sequeda, and Kavitha Srinivas. 2022. Results of semtab 2022.Semantic Web Challenge on Tabular Data to Knowledge Graph Matching3320 (2022)

    2022

  2. [2]

    Peter Auer, Nicolo Cesa-Bianchi, and Paul Fischer. 2002. Finite-time analysis of the multiarmed bandit problem.Machine learning47, 2 (2002), 235–256

    2002

  3. [3]

    Leo Breiman. 2001. Random forests.Machine learning45, 1 (2001), 5–32

    2001

  4. [4]

    Michael J Cafarella, Alon Y Halevy, Daisy Zhe Wang, Eugene Wu, and Yang Zhang. 2008. WebTables: Exploring the power of tables on the web.Proc. VLDB Endow.1, 1 (2008), 538–549

    2008

  5. [5]

    Hanzhu Chen, Xu Shen, Qitan Lv, Jie Wang, Xiaoqi Ni, and Jieping Ye. 2024. SAC-KG: Exploiting large language models as skilled automatic constructors for domain knowledge graph. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 4345–4360

    2024

  6. [6]

    Jiaoyan Chen, Ernesto Jiménez-Ruiz, Ian Horrocks, and Charles Sutton. 2019. Colnet: Embedding the semantics of web tables for column type prediction. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 33. 29–36

    2019

  7. [7]

    Mengshi Chen, Yuxiang Sun, Tengchao Li, Jianwei Wang, Kai Wang, Xuemin Lin, Ying Zhang, and Wenjie Zhang. 2026. Empowering tabular data preparation with language models: Why and how?. InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 228–246

    2026

  8. [8]

    Qian Chen, Jianwei Wang, and Wenjie Zhang. 2026. Collaborative Large and Small Language Models for Accurate and Scalable Data Repair.arXiv preprint arXiv:2606.17582(2026)

    Pith/arXiv arXiv 2026

  9. [9]

    Arnak Dalalyan and Alexandre B Tsybakov. 2008. Aggregation by exponential weighting, sharp PAC-Bayesian bounds and sparsity.Machine Learning72, 1 (2008), 39–61

    2008

  10. [10]

    Sanjoy Dasgupta. 2016. A cost function for similarity-based hierarchical clus- tering. InProceedings of the forty-eighth annual ACM symposium on Theory of Computing. 118–127

    2016

  11. [11]

    Xiang Deng, Huan Sun, Alyssa Lees, You Wu, and Cong Yu. 2022. Turl: Table understanding through representation learning.ACM SIGMOD Record51, 1 (2022), 33–40

    2022

  12. [12]

    Zhihao Ding, Yongkang Sun, and Jieming Shi. 2025. Retrieve-and-verify: A table context selection framework for accurate column annotations.Proceedings of the ACM on Management of Data3, 6 (2025), 1–27

    2025

  13. [13]

    Vasilis Efthymiou, Oktie Hassanzadeh, Mariano Rodriguez-Muro, and Vassilis Christophides. 2017. Matching web tables with knowledge base entities: from entity lookups to entity embeddings. InInternational Semantic Web Conference. Springer, 260–277

    2017

  14. [14]

    Benjamin Feuer, Yurong Liu, Chinmay Hegde, and Juliana Freire. 2024. ArcheType: A Novel Framework for Open-Source Column Type Annotation Using Large Language Models.Proc. VLDB Endow.17, 9 (May 2024), 2279–2292. doi:10.14778/ 3665844.3665857

    arXiv 2024

  15. [15]

    Ramanathan V Guha, Dan Brickley, and Steve Macbeth. 2016. Schema. org: evolution of structured data on the web.Commun. ACM59, 2 (2016), 44–51

    2016

  16. [16]

    Ehsan Hoseinzade, Ke Wang, and Anandharaju Durai Raju. 2026. TabEmb: Joint Semantic-Structure Embedding for Table Annotation. InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 16620–16631

    2026

  17. [17]

    Madelon Hulsebos, Çagatay Demiralp, and Paul Groth. 2023. Gittables: A large- scale corpus of relational tables.Proceedings of the ACM on Management of Data 1, 1 (2023), 1–17

    2023

  18. [18]

    Madelon Hulsebos, Kevin Hu, Michiel Bakker, Emanuel Zgraggen, Arvind Satya- narayan, Tim Kraska, Çagatay Demiralp, and César Hidalgo. 2019. Sherlock: A deep learning approach to semantic data type detection. InProceedings of the 25th ACM SIGKDD International Conference on knowledge discovery & data mining. 1500–1508

    2019

  19. [19]

    Rongjie Jiang, Jianwei Wang, Gengda Zhao, Chengyang Luo, Kai Wang, and Wenjie Zhang. 2026. Advancing Multimodal Agent Reasoning with Long-Term Neuro-Symbolic Memory.arXiv preprint arXiv:2603.15280(2026)

    arXiv 2026

  20. [20]

    Ricardo Knauer, Mario Koddenbrock, Raphael Wallsberger, Nicholas M Brisson, Georg N Duda, Deborah Falla, David W Evans, and Erik Rodner. 2025. ’Oh LLM, I’m Asking Thee, Please Give Me a Decision Tree’: Zero-Shot Decision Tree Induction and Embedding with Large Language Models. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data...

    2025

  21. [21]

    Keti Korini and Christian Bizer. 2023. Column type annotation using chatgpt. arXiv preprint arXiv:2306.00745(2023)

    Pith/arXiv arXiv 2023

  22. [22]

    Keti Korini and Christian Bizer. 2024. Column property annotation using large language models. InEuropean Semantic Web Conference. Springer, 61–70

    2024

  23. [23]

    Keti Korini, Ralph Peeters, and Christian Bizer. 2022. SOTAB: The WDC Schema. org table annotation benchmark. InCEUR Workshop Proceedings, Vol. 3320. RWTH Aachen, 14–19

    2022

  24. [24]

    Keti Korini, Ralph Peeters, and Christian Bizer. 2023. SOTAB V2: The WDC Schema.org Table Annotation Benchmark. https://webdatacommons.org/ structureddata/sotab/v2/. Accessed: 2026-07-15

    2023

  25. [25]

    Shankar Kumar and Bill Byrne. 2004. Minimum bayes-risk decoding for statistical machine translation. InProceedings of the Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics: HLT-NAACL 2004. 169–176

    2004

  26. [26]

    Xunhao Lai, Weiqi Xu, Yufeng Yang, Qiaorui Chen, Yang Xu, Lunbin Zeng, Xiaolong Li, Haohai Sun, Haichao Zhu, Vito Zhang, et al. 2026. MiniMax Sparse Attention.arXiv preprint arXiv:2606.13392(2026)

    Pith/arXiv arXiv 2026

  27. [27]

    Jens Lehmann, Robert Isele, Max Jakob, Anja Jentzsch, Dimitris Kontokostas, Pablo N Mendes, Sebastian Hellmann, Mohamed Morsey, Patrick Van Kleef, Sören Auer, et al. 2015. Dbpedia–a large-scale, multilingual knowledge base extracted from wikipedia.Semantic web6, 2 (2015), 167–195

    2015

  28. [28]

    Girija Limaye, Sunita Sarawagi, and Soumen Chakrabarti. 2010. Annotating and searching web tables using entities, types and relationships.Proceedings of the VLDB Endowment3, 1-2 (2010), 1338–1347

    2010

  29. [29]

    Shiyuan Liu, Jianwei Wang, Xuemin Lin, Lu Qin, Wenjie Zhang, and Ying Zhang

  30. [30]

    Xusheng Luo, Kangqi Luo, Xianyang Chen, and Kenny Zhu. 2018. Cross-lingual entity linking for web tables. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 32

    2018

  31. [31]

    Zhengjie Miao and Jin Wang. 2023. Watchog: A light-weight contrastive learning based framework for column annotation.Proceedings of the ACM on Management of Data1, 4 (2023), 1–24

    2023

  32. [32]

    Belinda Mo, Kyssen Yu, Joshua Kazdan, Proud Mpala, Lisa Yu, Charilaos Kanat- soulis, and Sanmi Koyejo. 2026. Kggen: Extracting knowledge graphs from plain text with language models.Advances in Neural Information Processing Systems 38 (2026), 30092–30115

    2026

  33. [33]

    Alexandru Niculescu-Mizil and Rich Caruana. 2005. Predicting good probabilities with supervised learning. InProceedings of the 22nd international conference on Machine learning. 625–632

    2005

  34. [34]

    Hanchuan Peng, Fuhui Long, and Chris Ding. 2005. Feature selection based on mutual information criteria of max-dependency, max-relevance, and min- redundancy.IEEE Transactions on pattern analysis and machine intelligence27, 8 (2005), 1226–1238

    2005

  35. [35]

    Minh Pham, Suresh Alse, Craig A Knoblock, and Pedro Szekely. 2016. Semantic labeling: a domain-independent approach. InInternational Semantic Web Confer- ence. Springer, 446–462

    2016

  36. [36]

    S Krishnamurthy Ramnandan, Amol Mittal, Craig A Knoblock, and Pedro Szekely

  37. [37]

    Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. InProceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP). 3982–3992

    2019

  38. [38]

    Dominique Ritze and Christian Bizer. 2017. Matching web tables to dbpedia-a feature utility study.context42, 41 (2017), 210–221

    2017

  39. [39]

    David F Robinson and Leslie R Foulds. 1981. Comparison of phylogenetic trees. Mathematical biosciences53, 1-2 (1981), 131–147

    1981

  40. [40]

    Yoshihiko Suhara, Jinfeng Li, Yuliang Li, Dan Zhang, Çağatay Demiralp, Chen Chen, and Wang-Chiew Tan. 2022. Annotating columns with pre-trained lan- guage models. InProceedings of the 2022 International Conference on Management of Data. 1493–1503

    2022

  41. [41]

    Yushi Sun, Xujia Li, Nan Tang, Quanqing Xu, Chuanhui Yang, and Lei Chen

  42. [42]

    Xingyu Tan, Xiaoyang Wang, Qing Liu, Xiwei Xu, Xin Yuan, Liming Zhu, and Wenjie Zhang. 2026. PrivGemo: Privacy-Preserving Dual-Tower Graph Retrieval for Empowering LLM Reasoning with Memory Augmentation.arXiv preprint arXiv:2601.08739(2026)

    arXiv 2026

  43. [43]

    Jianwei Wang, Mengqi Wang, Yinsi Zhou, Zhenchang Xing, Qing Liu, Xiwei Xu, Binghao Li, Serkan Saydam, Wenjie Zhang, and Liming Zhu. 2026. LLM- based HSE compliance assessment: Benchmark, performance, and advancements. InInternational Conference on Intelligent Systems and Knowledge Engineering. Springer, 507–536

    2026

  44. [44]

    LakeHopper: Cross Data Lakes Column Type Annotation through Model Adaptation.arXiv preprint arXiv:2602.08793(2026)

    arXiv 2026

  45. [45]

    Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou

  46. [46]

    Anyi Xu, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, Chenchen Ling, et al. 2026. Deepseek- v4: Towards highly efficient million-token context intelligence.arXiv preprint arXiv:2606.19348(2026)

    arXiv 2026

  47. [47]

    Mengqi Wang, Jianwei Wang, Qing Liu, Xiwei Xu, Zhenchang Xing, Liming Zhu, and Wenjie Zhang. 2025. Ensembling LLM-Induced Decision Trees for Explainable and Robust Error Detection.arXiv preprint arXiv:2512.07246(2025)

    Pith/arXiv arXiv 2025

  48. [48]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. 2025. Qwen3 technical report.arXiv preprint arXiv:2505.09388(2025)

    Pith/arXiv arXiv 2025

  49. [49]

    Dan Zhang, Yoshihiko Suhara, Jinfeng Li, Madelon Hulsebos, Çağatay Demiralp, and Wang-Chiew Tan. 2019. Sato: Contextual semantic type detection in tables. arXiv preprint arXiv:1911.06311(2019)

    Pith/arXiv arXiv 2019

  50. [50]

    Mo Zhou, Jianwei Wang, Kai Wang, Helen Paik, Ying Zhang, and Wenjie Zhang

  51. [51]

    Yuanyuan Xu, Wenjie Zhang, Yin Chen, Xuemin Lin, and Ying Zhang. 2026. Self- evolving agents as dynamic graph transformation: A survey and new perspective. Interpretable Column Annotation with LLM-Symbolized Decision Process Materialization Conference’17, July 2017, Washington, DC, USA (2026)

    2026

  52. [55]

    Conference’17, July 2017, Washington, DC, USA Mengqi Wang et al

    Multi-Perspective Evidence Synthesis and Reasoning for Unsupervised Multimodal Entity Linking.arXiv preprint arXiv:2604.20283(2026). Conference’17, July 2017, Washington, DC, USA Mengqi Wang et al. A LLM Prompt Templates Prompt Design A: LLM-Guided Skeleton Induction (1) Instruction:Organize the given CTA label vocabulary into a hypernym-inspired semantic...

    Pith/arXiv arXiv 2026

  53. [2015]

    InEuropean semantic web conference

    Assigning semantic labels to data sources. InEuropean semantic web conference. Springer, 403–417

  54. [2020]

    Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers.Advances in neural information processing systems33 (2020), 5776–5788

    2020

  55. [2026]

    HyperJoin: LLM-augmented Hypergraph Link Prediction for Joinable Table Discovery.arXiv preprint arXiv:2601.01015(2026)

    arXiv 2026