Pith. sign in

REVIEW 2 major objections 6 minor 35 references

Template-Based Schema Matching of Multi-Layout Tenancy Schedules:A Comparative Study of a Template-Based Hybrid Matcher and the ALITE Full Disjunction Model

T0 review · 2 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a template-based hybrid matcher can align multi-layout tenancy schedules to one business-approved target schema, reaching F1 0.881 with 45.7% nulls, versus F1 0.712 and 75.6% nulls for a full-disjunction baseline.

desk verdict Real artifacts and a plausible domain, but the headline F1 gap is an artifact of tuning and a mismatched baseline. read the letter →

arxiv 2507.02020 v1 pith:7YS22BHE submitted 2025-07-02 cs.DB

Tim Uilkema , Yao Ma , Seyed Sahand Mohammadi Ziabari , Joep van Vliet This is my paper
classification cs.DB
keywords schemamatchingtenancyscheduletemplate-basedintegrationhybridmatcherfulldisjunctionHungarianalgorithmusabilitymulti-layouttables
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 tries to establish that a business-first, template-based matcher can replace both manual per-layout mapping and fully automated table union when integrating tenancy schedules from different real estate firms. Instead of retaining every column from every layout, it aligns each source column to a predefined target schema that a business has already approved. On five real tenancy-schedule layouts, the matcher reaches an F1 score of 0.881 with a null percentage of 45.7% when per-attribute weights are tuned, whereas the full-disjunction baseline achieves F1 0.712 with 75.6% nulls and produces schemas that are far larger and less usable. If the result holds, real estate firms can onboard new document layouts automatically while keeping the output in a form their downstream systems can consume directly.

What carries the argument

The mechanism is a three-stage matching pipeline with a predefined target schema as the anchor. Each source column is scored against every target attribute using schema-based metrics (Levenshtein distance, Jaccard token overlap, synonym matching) and instance-based metrics (numeric and date type checks, numerical range comparison against expected means and interquartile ranges, and a Kolmogorov–Smirnov test comparing the column distribution to the attribute's expected distribution). The metrics are combined as a weighted sum with a tunable blend factor $\alpha$ and threshold $\theta$, and the final column-to-attribute assignment is solved as a maximum-weight bipartite matching problem with the Hungarian algorithm, producing a globally optimal one-to-one mapping among all pairs above threshold.

What would settle it

Run the optimized configuration on a manually labeled tenancy schedule from a sixth real estate firm and compare F1; if it falls substantially below 0.881, the reported gain is partly an artifact of tuning on the evaluation set.

Watch

Extended reading notes

Core claim

The central discovery is that combining a fixed business-approved target schema with hybrid column matching outperforms completeness-first integration for multi-layout tenancy schedules. Concretely, the paper reports that per-attribute weight optimization with default $\alpha=0.5$ and $\theta=0.5$ yields precision 0.926, recall 0.840, and F1 0.881 for the hybrid matcher, while a full-disjunction baseline using semantic column embeddings reaches F1 0.712 and leaves 75.6% of cells null. The matched output stays compact at 17 columns versus about 88 for full disjunction, and the null percentage drops to 45.7%. The paper further finds that instance-based metrics carry more weight than schema-based metrics for DATE, DECIMAL, and STRING attributes, with the largest gap for DATE attributes, supporting the claim that value distributions and data types resolve ambiguities that headers alone cannot.

Load-bearing premise

The load-bearing premise is that grid-search tuning on the manually labeled ground truth is an acceptable way to report performance; if the same ground truth is used for both tuning and scoring, the reported F1 of 0.881 may be an in-sample fit rather than a generalizable gain over the baseline.

Editorial extensions

If this is right

  • Incoming tenancy schedules with new layouts can be mapped to the target schema without building a separate template for each layout, removing the main scalability bottleneck in current document integration.
  • Downstream applications receive a 17-column schema with 45.7% nulls instead of an 88-column union with 75.6% nulls, so less post-processing is needed before property management and financial reporting.
  • Per-attribute weight tuning matters more than global parameter tuning: it raises F1 from 0.800 to 0.881, indicating that different attribute types need different matching signals.
  • Because the target schema is business-approved and extensible, new attributes can be added without redesigning the matching logic for existing layouts.

Reading between the lines

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

  • The accuracy comparison is not apples-to-apples because the hybrid matcher is scored on full target-schema mappings while the baseline is scored on 20 manually grouped column sets; the robust conclusion is that a template-based pipeline produces a more usable schema, while the exact F1 gap should be read with that asymmetry in mind.
  • The dominance of instance-based weights for DATE and DECIMAL attributes suggests the same design would transfer to other numeric-heavy document families, such as invoices, where value ranges can disambiguate headers that no longer look alike.
  • Because weights were tuned and scored on the same ground truth, a held-out layout or cross-validation would be the decisive test of whether 0.881 generalizes; until then the reported number is best understood as an upper bound of what the tuned matcher can do on these five layouts.
  • The one-to-one Hungarian constraint rules out composite mappings such as merging office area and archive area into one leased-area attribute; relaxing it to many-to-one assignments is the paper's stated next step and would broaden practical coverage.
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 proposes a template-based hybrid schema matcher that aligns multi-layout tenancy schedules (five real-estate layouts, 4–6 schedules each) to a fixed target schema derived from CBRE's Data Requirements Document (DRD). The matcher combines schema-based metrics (Levenshtein, Jaccard, synonym matching) and instance-based metrics (data type, range, KS-test) with a Hungarian-algorithm global assignment. The evaluation reports F1 scores for four matcher configurations, with a default-configuration F1 of 0.778 and a per-attribute grid-search-optimized F1 of 0.881, and compares these against an ALITE-style full-disjunction baseline that achieves F1 0.712 on a separate ground truth of 20 semantically similar column sets. The paper claims substantial improvements in both alignment accuracy and schema usability (17 columns, 45.7% nulls vs. 88 columns, 75.6% nulls). Sections 4 and 5 of the manuscript are identical in content.

Significance. If the central claim were properly established, the paper would make a useful contribution to domain-specific schema matching: it addresses the multi-layout-to-one-schema scenario with a business-defined target schema, a combination rarely studied in the schema-matching literature. The practical motivation (tenancy schedules, real-estate data integration) is clear and relevant, and the pipeline is described in enough detail to be reimplementable. The availability of the code, ground-truth mappings, and data on GitHub is a strength, as is the honest reporting of the default configuration's F1 of 0.778 and the error analysis acknowledging that null percentage can be misleading. However, the headline comparison is currently undermined by the evaluation design: the optimized F1 is an in-sample fit on the same ground truth used for scoring, and the two F1 numbers being compared are computed on different annotation objects. Until these issues are resolved, the claimed 'substantial improvements' are not supported.

major comments (2)
  1. [Sections 3.4, 4.1, 5.2, Table 5] The headline F1 of 0.881 is obtained by grid-search optimizing per-attribute metric weights on the same manually labeled ground truth that is then used to compute the F1 score. Section 3.4 states that the grid search uses the ground truth mapping, and Sections 4.1 and 5.1 state that these mappings 'serve as a benchmark for evaluating the accuracy of schema matching and training individual metric weights.' No held-out split or cross-validation is described anywhere in the paper. The optimized configuration is therefore an in-sample fit, and with only 4–6 schedules per layout and 17 target attributes, this can materially inflate F1. The default configuration (alpha=0.5, theta=0.5) at F1 0.778 is not affected by this issue, but the abstract and Section 6.3 present the 0.881 value as the main result. To support generalization claims, the authors should either use a held-out split at the layout level, perform cross-validation, or clearly report the optimized numbers as in-sample tuning results and base any comparative claim on the default configuration or on a validation set.
  2. [Sections 3.5, 5.2, Table 5 footnote; Abstract] The two F1 numbers compared in the abstract and Section 6.3 are computed on different ground truths and different matching tasks. The hybrid matcher is evaluated against column-to-attribute mappings to the DRD target schema, while the ALITE baseline is evaluated on a manually constructed ground truth of 20 semantically similar column sets (column-to-column grouping). Section 3.5 explicitly states that this baseline evaluation 'is not done for direct comparison but to verify that its alignment accuracy is acceptable,' and Section 5.2 states that the F1 of the FD baseline 'is not directly compared in terms of alignment accuracy.' Nevertheless, the abstract presents both numbers (0.881 vs. 0.712) as evidence of 'substantial improvements,' and Section 6.3 repeats this comparison. Because the matching definitions and annotation objects differ, these two F1 scores are not comparable. The usability comparison (17 vs. 88/89 columns, 45.7% vs. 75.6% nulls) is also confounded by the different objective functions and reference schemas. The authors should either compute the baseline on the same column-to-attribute ground truth or remove the comparative interpretation of these numbers from the abstract and conclusions.
minor comments (6)
  1. [Sections 4 and 5] Sections 4 and 5 are identical in text and tables: Table 3 equals Table 5, Table 4 equals Table 6, and Sections 4.1–4.4 duplicate Sections 5.1–5.4 verbatim. This duplication should be removed; only one evaluation section should remain.
  2. [Section 4.3/5.3, Wilcoxon test] The Wilcoxon signed-rank test is reported with a test statistic of 3.0 and p=0.00049, but the manuscript does not state the sample size, the number of attributes, or the definition of the paired observations used in the test. Please provide these details so the statistical claim can be assessed.
  3. [Appendix B.2, KS-test] The KS-test description says the source column is compared to 'a normal distribution generated using the target's mean and IQR,' but the formula given is the standard two-sample KS statistic between empirical CDFs. The relationship between the normal reference distribution and the empirical CDF of the target should be clarified, and a citation for this usage should be provided.
  4. [Section 5.4/4.4] The sentence 'Per-attribute weights would reveal that the columns denotes "Office Area"' is grammatically incorrect and unclear; the authors likely mean that per-attribute weight optimization would correct this mapping. Please rephrase.
  5. [Section 3.5.2] The reproducibility statement gives the URL https://github.com/TUilkema, which appears to be a user profile rather than a specific repository path; the exact repository URL should be provided.
  6. [General] The paper's own error analysis in Section 5.4 notes that the overall null percentage can be misleading because some target attributes are rarely present in any source schedule; the abstract and Section 6.3 still emphasize the null-percentage difference without this caveat. Please reconcile these statements.

Circularity Check

1 steps flagged · score 6.0 of 10

The headline F1 of 0.881 is the in-sample optimum of a grid search against the same ground truth used for evaluation, so the reported peak is a fitted value rather than an independent prediction.

  1. fitted input called prediction [Sections 3.1.4, 3.4, 4.1, and Table 5 / Section 5.2]
    "Ground truth mapping serves two purposes. First, it is the basis for supervised learning or rule calibration in our hybrid matcher. ... Secondly, the ground truth is essential for our evaluation stage. ... To systematically explore the optimal settings for these parameters, this research utilizes a grid search optimization using the ground truth mapping. ... The F1-score ... served as the main metric for optimizing parameters and weights. ... The highest performance was achieved with per-attribute weight optimization, reaching an F1-score of 0.881."

    The per-attribute metric weights are chosen by maximizing F1 on the same manually created column-to-attribute ground truth that is then used as the evaluation benchmark. Section 3.1.4 explicitly assigns this ground truth the double role of rule calibration and evaluation, and Section 3.4 states that the objective of hyperparameter optimization is to maximize F1 across the dataset. Consequently, the reported peak F1 of 0.881 is the value of the optimization objective at the fitted configuration on the calibration data, not an unbiased estimate for unseen tenancy schedules.

full rationale

The only clear circularity is the in-sample optimization loop. The paper is transparent that the ground truth is used both for training/calibration and for evaluation, and that grid search maximizes F1 over the whole dataset; this makes the peak F1 of 0.881 an in-sample fit. No load-bearing self-citation is present: the ALITE baseline is external and the authors provide a public repository, so reproducibility is not the issue. An additional validity concern, not a circularity, is that the ALITE F1 (0.712) is computed on a separate ground truth of 20 semantically similar column sets and the paper itself states that this 'is not done for direct comparison'; nevertheless the abstract compares it to the hybrid F1 as evidence of 'substantial improvements.' That mismatch weakens the head-to-head comparison but does not itself make the derivation circular. The default-configuration result (0.778) and the usability metrics provide some independent content, so the appropriate score is partial circularity at 6 rather than 8 or 10.

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

The ledger shows the empirical result depends on several manually set or tuned quantities: the alpha blend, threshold, per-attribute metric weights, and the hand-written synonyms and ranges in the DRD target schema. No new formal entities are introduced. The main burden is that the tuned weights and the evaluation labels come from the same ground truth, which is the circularity flagged in this report.

free parameters (4)
  • alpha (α) = 0.75 optimized, 0.5 default
    Balances schema-based versus instance-based scores; chosen by grid search to maximize F1 on the ground truth.
  • threshold (θ) = 0.51 optimized, 0.5 default
    Minimum hybrid similarity for a column-attribute assignment; chosen by grid search on the evaluation ground truth.
  • Per-attribute metric weights = Seven weights per attribute in [0,1], exact values not reported in the paper
    Grid-searched per target attribute against the manually labeled ground truth to maximize F1.
  • Target schema synonyms and typical value ranges = Examples: rent between €50,000 and €2,000,000; office area between 200 and 5,000 sqm
    Hand-specified business knowledge from the CBRE DRD, used in synonym, range, and KS-test metrics; these constants influence matching but are not fitted with uncertainty.
assumptions (5)
  • domain assumption Clean, pre-extracted tabular input with no OCR noise
    Stated in the abstract and Section 6.4; the pipeline starts from CSV headers and values, so noisy extraction would break the matching.
  • domain assumption One-to-one column-to-attribute mapping is sufficient
    Section 3.2.3 imposes one-to-one matching via the Hungarian algorithm; composite fields such as Office plus Archive to Leased Area are explicitly excluded.
  • domain assumption Manual ground-truth mapping is correct and complete
    Section 3.1.4; all F1 scores and all optimized weights depend on the correctness of these manual labels.
  • domain assumption Anonymization preserves data types and distributions
    Section 3.1.2; instance-based matching relies on data types, ranges, and distributions, so anonymization must not alter those characteristics.
  • domain assumption Five layouts with 4-6 schedules each are representative
    Section 6.4 acknowledges the small number of layouts; the generalization claim depends on this representativeness.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Template-Based Schema Matching of Multi-Layout Tenancy Schedules:A Comparative Study of a Template-Based Hybrid Matcher and the ALITE Full Disjunction Model." pith.science (2026). https://pith.science/paper/7YS22BHE

@misc{pith2026250702020,
  author       = {Pith},
  title        = {Pith review of: Template-Based Schema Matching of Multi-Layout Tenancy Schedules:A Comparative Study of a Template-Based Hybrid Matcher and the ALITE Full Disjunction Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7YS22BHE}},
  note         = {Machine review of arXiv:2507.02020}
}
read the original abstract

The lack of standardized tabular formats for tenancy schedules across real estate firms creates significant inefficiencies in data integration. Existing automated integration methods, such as Full Disjunction (FD)-based models like ALITE, prioritize completeness but result in schema bloat, sparse attributes and limited business usability. We propose a novel hybrid, template-based schema matcher that aligns multi-layout tenancy schedules to a predefined target schema. The matcher combines schema (Jaccard, Levenshtein) and instance-based metrics (data types, distributions) with globally optimal assignments determined via the Hungarian Algorithm. Evaluation against a manually labeled ground truth demonstrates substantial improvements, with grid search optimization yielding a peak F1-score of 0.881 and an overall null percentage of 45.7%. On a separate ground truth of 20 semantically similar column sets, ALITE achieves an F1-score of 0.712 and 75.6% nulls. These results suggest that combining structured business knowledge with hybrid matching can yield more usable and business-aligned schema mappings. The approach assumes cleanly extracted tabular input, future work could explore extending the matcher to support complex, composite tables.

Figures

Figures reproduced from arXiv: 2507.02020 by the authors.

Figure 1
Figure 1. High-level overview of methodology Each of the following subsections provides insight into a spe￾cific stage of the pipeline, highlighting both the methodological choices made and the practical implementation details to support reproducibility. 3.1 Dataset and Target Schema 3.1.1 Source and Layout Diversity. To evaluate the robustness of the proposed hybrid schema matcher across real-world variability, a diverse dat… view at source ↗
Figure 2
Figure 2. Detailed illustration of schema matching pipeline [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 20 canonical work pages

  1. [1]

    Md Asif-Ur-Rahman, Bayzid Ashik Hossain, Michael Bewong, Md Zahidul Islam, Yanchang Zhao, Jeremy Groves, and Rory Judith. 2023. A semi-automated hybrid schema matching framework for vegetation data integration. Expert Systems 12 with Applications 229 (Nov. 2023), 120405. doi:10.1016/j.eswa.2023.120405

    work page arXiv 2023

  2. [2]

    David Aumueller, Hong-Hai Do, Sabine Massmann, and Erhard Rahm. 2005. Schema and ontology matching with COMA++. In Proceedings of the 2005 ACM SIGMOD international conference on Management of data . ACM, Baltimore Mary- land, 906–908. doi:10.1145/1066157.1066283

    arXiv 2005

  3. [3]

    Dipali Baviskar, Swati Ahirrao, Vidyasagar Potdar, and Ketan Kotecha. 2021. Efficient Automated Processing of the Unstructured Documents Using Artificial Intelligence: A Systematic Literature Review and Future Directions. IEEE Access 9 (2021), 72894–72936. doi:10.1109/ACCESS.2021.3072900

    arXiv 2021

  4. [4]

    Douglas Burdick, Marina Danilevsky, Alexandre V Evfimievski, Yannis Katsis, and Nancy Wang. 2020. Table extraction and understanding for scientific and enterprise applications. Proceedings of the VLDB Endowment 13, 12 (Aug. 2020), 3433–3436. doi:10.14778/3415478.3415563

    arXiv 2020

  5. [5]

    Riccardo Cappuzzo, Aimee Coelho, Felix Lefebvre, Paolo Papotti, and Gael Varo- quaux. 2025. Retrieve, Merge, Predict: Augmenting Tables with Data Lakes. doi:10.48550/arXiv.2402.06282 arXiv:2402.06282 [cs]

    work page Pith review arXiv doi:10.48550/arxiv.2402.06282 2025

  6. [6]

    Zewen Chi, Heyan Huang, Heng-Da Xu, Houjin Yu, Wanxuan Yin, and Xian-Ling Mao. 2019. Complicated Table Structure Recognition. doi:10.48550/arXiv.1908. 04729 arXiv:1908.04729 [cs]

    arXiv doi:10.48550/arxiv.1908 2019

  7. [7]

    Lingxi Cui, Huan Li, Ke Chen, Lidan Shou, and Gang Chen. 2024. Tabular Data Augmentation for Machine Learning: Progress and Prospects of Embracing Generative AI. doi:10.48550/arXiv.2407.21523 arXiv:2407.21523 [cs]

    arXiv doi:10.48550/arxiv.2407.21523 2024

  8. [8]

    Rosen, Gerbrand Ceder, Kristin A

    John Dagdelen, Alexander Dunn, Sanghoon Lee, Nicholas Walker, Andrew S. Rosen, Gerbrand Ceder, Kristin A. Persson, and Anubhav Jain. 2024. Structured information extraction from scientific text with large language models. Nature Communications 15, 1 (Feb. 2024), 1418. doi:10.1038/s41467-024-45563-x

    doi:10.1038/s41467-024-45563-x 2024

Show all 35 references
  1. [9]

    Hong-Hai Do and Erhard Rahm. 2002. COMA — A system for flexible combination of schema matching approaches. In VLDB ’02: Proceedings of the 28th Interna- tional Conference on Very Large Databases . Elsevier, 610–621. doi:10.1016/B978- 155860869-6/50060-3

    2002 doi

  2. [10]

    Anhai Doan, Pedro Domingos, and Alon Halevy. 2001. Reconciling Schemas of Disparate Data Sources: A Machine-Learning Approach

    2001

  3. [11]

    Yuyang Dong, Chuan Xiao, Takuma Nozawa, Masafumi Enomoto, and Masafumi Oyamada. 2023. DeepJoin: Joinable Table Discovery with Pre-trained Language Models. doi:10.48550/arXiv.2212.07588 arXiv:2212.07588 [cs]

    2023 doi arXiv

  4. [12]

    Azka Gilani, Shah Rukh Qasim, Imran Malik, and Faisal Shafait. 2017. Table Detection Using Deep Learning. In 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR) . IEEE, Kyoto, 771–776. doi:10.1109/ ICDAR.2017.131

    2017

  5. [13]

    Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. 2022. LayoutLMv3: Pre-training for Document AI with Unified Text and Image Masking. doi:10. 48550/arXiv.2204.08387 arXiv:2204.08387 [cs]

    2022 doi arXiv

  6. [14]

    Daomin Ji, Hui Luo, Zhifeng Bao, and Shane Culpepper. 2025. Table Integra- tion in Data Lakes Unleashed: Pairwise Integrability Judgment, Integrable Set Discovery, and Multi-Tuple Conflict Resolution. doi:10.48550/arXiv.2412.00324 arXiv:2412.00324 [cs]

    2025 doi arXiv

  7. [15]

    Aamod Khatiwada, Roee Shraga, Wolfgang Gatterbauer, and Renée J. Miller. 2022. Integrating Data Lake Tables. Proceedings of the VLDB Endowment 16, 4 (Dec. 2022), 932–945. doi:10.14778/3574245.3574274

    2022

  8. [16]

    Nitish Korula. 2010. Maximum Weight Matching in Bipartite Graphs

    2010

  9. [17]

    Christos Koutras, Kyriakos Psarakis, George Siachamis, Andra Ionescu, Marios Fragkoulis, Angela Bonifati, and Asterios Katsifodimos. 2021. Valentine in action: matching tabular data at scale. Proceedings of the VLDB Endowment 14, 12 (July 2021), 2871–2874. doi:10.14778/3476311.3476366

    2021

  10. [18]

    Christos Koutras, George Siachamis, Andra Ionescu, Kyriakos Psarakis, Jerry Brons, Marios Fragkoulis, Christoph Lofi, Angela Bonifati, and Asterios Katsifodi- mos. 2021. Valentine: Evaluating Matching Techniques for Dataset Discovery. In 2021 IEEE 37th International Conference...

    2021

  11. [19]

    Diego Leon. 2021. Extracting Information From PDF Invoices Using Deep Learn- ing

    2021

  12. [20]

    Jayant Madhavan, Philip A Bernstein, and Erhard Rahm. [n. d.]. Generic Schema Matching with Cupid. ([n. d.])

  13. [21]

    Madonna University and Adanma Cecilia Eberendu. 2016. Unstructured Data: an overview of the data of Big Data. International Journal of Computer Trends and Technology 38, 1 (Aug. 2016), 46–50. doi:10.14445/22312803/IJCTT-V38P109

    2016 doi

  14. [22]

    Melnik, H

    S. Melnik, H. Garcia-Molina, and E. Rahm. 2002. Similarity flooding: a versatile graph matching algorithm and its application to schema matching. InProceedings 18th International Conference on Data Engineering . IEEE Comput. Soc, San Jose, CA, USA, 117–128. doi:10.1109/ICDE.20...

    2002

  15. [23]

    Renke J Miller. 1998. Using schematically heterogeneous structures. ACM SIGMOD Record (June 1998), 189–200. doi:10.1145/276305.276322

    1998

  16. [24]

    Suyash Mishra and Anuranjan Misra. 2017. Structured and Unstructured Big Data Analytics. In 2017 International Conference on Current Trends in Computer, Electrical, Electronics and Communication (CTCEEC) . IEEE, Mysore, 740–746. doi:10.1109/CTCEEC.2017.8454999

    2017

  17. [25]

    Pu, and Renée J

    Fatemeh Nargesian, Erkang Zhu, Ken Q. Pu, and Renée J. Miller. 2018. Table union search on open data. Proceedings of the VLDB Endowment 11, 7 (March 2018), 813–825. doi:10.14778/3192965.3192973

    2018

  18. [26]

    Ahmed Nassar, Nikolaos Livathinos, Maksym Lysak, and Peter Staar. 2022. Table- Former: Table Structure Understanding with Transformers. doi:10.48550/arXiv. 2203.01017 arXiv:2203.01017 [cs]

    2022 doi arXiv

  19. [27]

    Shubham Paliwal, Vishwanath D, Rohit Rahul, Monika Sharma, and Lovekesh Vig

  20. [28]

    Balachandar Paulraj. 2024. Ext-NoSQL: A NoSQL Schematics through JSON. In 2024 IEEE 12th International Conference on Intelligent Systems (IS) . IEEE, Varna, Bulgaria, 1–7. doi:10.1109/IS61756.2024.10705262

    2024

  21. [29]

    David Pinto, Andrew McCallum, Xing Wei, and W Bruce Croft. [n. d.]. Table Extraction Using Conditional Random Fields. ([n. d.])

  22. [30]

    Bernstein

    Erhard Rahm and Philip A. Bernstein. 2001. A survey of approaches to automatic schema matching. The VLDB Journal 10, 4 (Dec. 2001), 334–350. doi:10.1007/ s007780100057

    2001

  23. [31]

    Thomas Saout, Frédéric Lardeux, and Frédéric Saubion. 2024. An Overview of Data Extraction From Invoices. IEEE Access 12 (2024), 19872–19886. doi:10.1109/ ACCESS.2024.3360528

    2024

  24. [32]

    Mayank Singh, Barnopriyo Barua, Priyank Palod, Manvi Garg, Sidhartha Sata- pathy, Samuel Bushi, Kumar Ayush, Krishna Sai Rohith, Tulasi Gamidi, Pawan Goyal, and Animesh Mukherjee. [n. d.]. OCR++: A Robust Framework For Infor- mation Extraction from Scholarly Articles. ([n. d.])

  25. [33]

    Brandon Smock, Rohith Pesala, and Robin Abraham. 2022. PubTables-1M: To- wards comprehensive table extraction from unstructured documents. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . IEEE, New Orleans, LA, USA, 4624–4632. doi:10.1109/CVPR526...

    2022

  26. [34]

    Vijaya Bhaskara Reddy Soperla. 2025. AUTOMATED DATA MAPPING FOR ENTERPRISE DATA INTEGRATION: LEVERAGING MACHINE LEARNING FOR EFFICIENT DATA TRANSFORMATION. INTERNATIONAL JOURNAL OF COMPUTER ENGINEERING AND TECHNOLOGY 16, 1 (Feb. 2025), 3883–3898. doi:10.34218/IJCET_16_01_268 1...

    2025 doi

  27. [2020]

    doi:10.48550/arXiv.2001.01469 arXiv:2001.01469 [cs]

    TableNet: Deep Learning model for end-to-end Table detection and Tabular data extraction from Scanned Document Images. doi:10.48550/arXiv.2001.01469 arXiv:2001.01469 [cs]

    2001 doi arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.