Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

TransClean: Finding False Positives in Multi-Source Entity Matching under Real-World Conditions via Transitive Consistency

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read False positives in entity matching can be found by testing the matcher on the transitive pairs its own matches imply, yielding an average +24.42 F1 gain.

desk verdict A sensible cleanup heuristic with a real practical motivation, but the headline F1 number is not supported by the paper's own table and the central proxy assumption needs a direct test before I'd trust it. read the letter →

arxiv 2506.04006 v1 pith:CMCCRJZB submitted 2025-06-04 cs.DB cs.AIcs.LG

Fernando de Meer Pardo , Branka Hadji Misheva , Martin Braschler , Kurt Stockinger This is my paper
classification cs.DBcs.AIcs.LG
keywords entitymatchingtransitiveconsistencyfalsepositivedetectionmulti-sourcedataintegrationrecordlinkagelabel-efficientcleanupgraphpruningLLMpseudo-labeling
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 introduces TransClean, a cleanup method for entity matching in large multi-source datasets where ground-truth labels are scarce. It claims that when a pairwise matching model links records A-B and B-C, the model's prediction on the implied pair A-C exposes whether the links are correct: negative predictions on such transitive pairs flag components that contain false positive matches. TransClean repeatedly prunes the minimum edge cuts of the most inconsistent components, re-fines the model on a small labeled budget, and then restores edges that remain transitively consistent. The paper reports an average +24.42 F1 improvement over plain pairwise matching across five multi-source benchmarks. If correct, the method gives practitioners a label-free quality signal for a matching and a way to clean it under a limited labeling budget.

What carries the argument

Transitive Consistency is the central measure: a pairwise matching model is transitively consistent on a matching graph if it predicts Match for every non-adjacent pair of records in the same connected component of the graph. The machinery is this consistency check plus the algorithmic loop around it: negative transitive predictions identify suspect components, minimum edge cuts provide candidate false-positive edges to label or prune, and the positive-versus-negative transitive prediction count decides which components to break up without labeling.

What would settle it

Take any dataset with known ground truth, run a pairwise matcher, and for each component count negative transitive predictions against the actual false-positive edges in that component; if components with many negative transitive predictions regularly contain no false-positive edges, the pruning rule would remove true positives and the proxy claim fails.

Watch

Extended reading notes

Core claim

The central claim is that positive and negative predictions on transitive matches—pairs of records that are connected indirectly through predicted matches but were never directly scored—serve as a proxy for the true and false positives of a matching. A model that says A matches B and B matches C but scores A-C as NoMatch is internally inconsistent, and the paper argues such inconsistencies concentrate where false positive edges have wrongly glued record groups together. TransClean operationalizes this by sorting components by negative transitive predictions, labeling edges from minimum edge cuts and shortest paths within a budget, pruning the cuts of components whose negative transitive predictions outnumber positive ones, and finally adding back removed edges whose reintroduction would not create new negative transitive predictions. The authors show the proxy quantities track true and false positive counts during cleanup and report an average +24.42 F1 improvement across datasets and base matchers.

Load-bearing premise

The method assumes that a negative prediction on an implied record pair reliably marks a false positive edge somewhere in that component; if a strong model simply misjudges hard true transitive pairs, the pruning removes true positives instead of false positives.

Editorial extensions

If this is right

  • On unlabeled multi-source data, the count of negative transitive predictions can be read directly as an estimate of matching quality, replacing manual inspection of entire matchings.
  • A practitioner can combine TransClean with an existing pairwise matcher rather than retraining an end-to-end system, because the cleanup operates on the matcher's output graph.
  • The method's improvements concentrate in settings where the base matcher already performs well; for low-recall matchings with few false positives, there is little to clean.
  • Using LLM pseudo-labels in place of human labeling preserves most of the gain on datasets where the base matcher is strong, but removes more true positives on weak matchings.

Reading between the lines

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

  • The transitive-consistency signal is not specific to entity matching: any pairwise classifier whose output induces connected components, such as coreference resolution, duplicate detection, or knowledge-graph linking, could use negative transitive predictions as a free quality audit.
  • A testable extension is to measure how the correlation between negative transitive predictions and false-positive edges degrades as base-model quality decreases; the paper's Camera and Monitor results suggest the proxy weakens precisely when the matcher's recall is very low.
  • The evolution of the positive and negative transitive prediction counts could serve as a stopping criterion for label acquisition, spending the labeling budget only while the inconsistency signal still moves.
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

5 major / 6 minor

Summary. The paper introduces Transitive Consistency, a measure based on a pairwise matching model's predictions on implied (transitive) record pairs within its own matching graph, and TransClean, an iterative cleanup algorithm that uses this measure plus a limited labeling budget to remove false-positive matches while preserving true positives. The method is evaluated on five multi-source datasets with two base matchers (DistilBERT and CLER), using both manual labels and LLM pseudo-labels, and the authors report an average +24.42 F1 improvement over pairwise matching when transitive matches are included in the evaluation.

Significance. The problem addressed is practically important: real-world multi-source entity matching produces matchings that are difficult to evaluate without large labeling efforts, and a lightweight method that removes false positives using the model's own transitive predictions would be valuable. The paper makes a concrete algorithmic proposal with clear pseudocode (Algorithms 1–3), uses a very large synthetic benchmark, compares against a state-of-the-art method (CLER), and reports runtime figures. These are genuine strengths. However, the central proxy assumption—that negative transitive predictions reliably indicate false-positive edges—is not validated beyond a single illustrative plot, and the paper's own Section 4.2.4 concedes a failure mode for this assumption. The headline quantitative claim is not reproducible from Table 3, and several CLER runs are excluded without a full discussion. If the proxy assumption and the reported gains can be substantiated with additional analysis, the work would be a useful contribution; in its current form the evidence is not sufficient.

major comments (5)
  1. [§4.1 and §4.2.4] The load-bearing premise is that negative transitive predictions are "very likely" in components with false-positive matches (Section 4.1) and that pruning minimum edge cuts of such components removes substantially more false positives than true positives (Section 4.2.2). This premise is contradicted by the paper's own concession in Section 4.2.4 that adding a true positive edge can increase the number of negative transitive predictions because the model may misclassify true transitive matches, especially in edge-case groups. Under that failure mode, an all-true-positive component can look transitively inconsistent, and Algorithm 2's pruning will remove true positives. Table 3 shows this is not hypothetical: the LLM-labeling runs on Camera, Monitor, and WDC remove 30.61–87.59% of true positives and reduce the F1 score below the Pre-TransClean value, and Section 5.3.2 explains these cases by saying transitive matches are "not a good indication" there—which is precisely the proxy assumption the method needs. The paper should provide a quantitative analysis of, for example, the precision of negative transitive predictions as an indicator of false-positive edges, broken down by dataset and by component type, rather than relying on a single motivating example.
  2. [Algorithm 2, line 8] There is a direct inconsistency between the algorithm and the prose. Section 4.2.3 states that TransClean should iteratively remove minimum edge cuts from components "with more negative transitive predictions than positive ones," and the while-loop comment in Algorithm 2 says pruning continues "until the positive transitive predictions outnumber the negatives." However, line 8 of Algorithm 2 tests if |Pos_tr| > |Neg_tr| and then prunes that component. As printed, the loop prunes components that are already transitively consistent according to the stated criterion and stops when every remaining component has at least as many negative as positive transitive predictions—the opposite of the intended behavior. This must be corrected, and the experiments need to be rerun with the corrected condition, or the prose must be changed to match the algorithm; otherwise the reported results cannot be attributed to the method as described.
  3. [§5.3.2, Table 3] The headline average +24.42 F1 improvement is not reproducible from the data presented. The seven manual-labeling rows in Table 3 yield an average improvement of about 20.93 (69.43, 5.05, 6.21, 0.33, 0.01, 3.24, 62.23), and the six LLM-labeling rows yield a different average (about 15.37). If a different subset of rows was used to compute +24.42, the selection rule must be stated. The same paragraph also says TransClean "improves the matchings produced by all pairwise models," but CLER is not applied on MusicBrainz (training timeout), Monitor (perfect precision, zero false positives), or WDC (zero false positives), so the claim covers only a subset of the experimental matrix. In addition, no error bars, standard deviations, or multiple-seed runs are reported for any of the F1 values, so it is unclear whether the improvements are statistically significant, especially in the low-recall Camera and Monitor cases.
  4. [§5.1.4, §5.3.1, Table 3] The evaluation metric counts all implied transitive pairs as part of the matching, which can make Pre-TransClean F1 collapse to 0.04 (Synthetic Companies with DistilBERT) or increase relative to pairwise F1 (Synthetic Companies with CLER). This choice makes the F1 improvements partly an artifact of the evaluation convention, because pruning a few false-positive edges can eliminate a large number of erroneous transitive pairs. The paper should report, in addition to the transitive-inclusive F1, the standard pairwise precision/recall/F1 on the final predicted edge set and compare those numbers with the base matcher's pairwise scores. Without this, it is difficult to assess how much of the reported gain corresponds to real improvement in the predicted matching as opposed to the change in the evaluation definition. The paper should also justify the fixed choices of n=5, S=50, and the 50/50 budget split; these are free parameters in Algorithms 1–3 and no sensitivity analysis is provided.
  5. [§5.2.3, Figure 7] The paper claims that positive and negative transitive predictions "are correlated" with true and false positives, and Figure 7 is presented as visual evidence of this. No correlation coefficient or other quantitative measure is given, and the claim is central to the method's motivation. A simple numerical analysis—e.g., per-component correlation between the number of negative transitive predictions and the number of false-positive edges, or precision/recall of using a negative transitive prediction to flag a component—would strengthen the paper substantially and would also address the concern raised in Section 4.2.4 about true positives creating negative transitive predictions.
minor comments (6)
  1. [§5.2.2] There is a typo: "haven not" should be "have not."
  2. [§5.3.2, Table 3] The Ditto marks in Table 3 make the repeated values clear, but the notation "10K budget(– used)" for the WDC Products DistilBERT run is confusing; please clarify whether the labeling budget was actually consumed and, if not, how that affects the comparison with the manual-labeling setup.
  3. [§5.3.2] The LLM-labeling runs on Camera, Monitor, and WDC are negative results (F1 decreases below Pre-TransClean), yet the abstract and conclusions emphasize only the positive average gain. These negative results should be reported in the abstract or at least prominently in the conclusions, not only in the body of Section 5.3.2.
  4. [§5.2.3, Figure 7] Figure 7 is difficult to read in print because the bars are grouped and the legend is dense; please provide a numerical table of the quantities shown (false positives, true positives, labeled pairs, positive transitive predictions, negative transitive predictions) for each TransClean step.
  5. [§5.1.2] The paper uses DeepSeek LLM 7B Base for pseudo-labeling, but no reference or version identifier is given for this model; please add a citation or URL so that the labeling procedure is reproducible.
  6. [General] No code or reproduction package is mentioned in the paper. Given the algorithmic nature of the contribution, a public implementation and configuration details would be essential for reproducibility, especially because the pseudocode in Algorithm 2 currently does not match the prose (see Major Comment 2).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: TransClean's transitive-consistency proxy is an empirically falsifiable heuristic, tested against ground truth, not a derivation that reduces to its own inputs.

full rationale

TransClean's load-bearing assertion is that negative transitive predictions are 'very likely in components/record groups with false positive matches' and 'can be used as a proxy of true and false positives of a matching' (Section 4.1). This is an empirical correlation claim between f_theta's decisions on implied pairs and ground-truth label status, and the paper validates it externally: Figure 7 compares the proxy counts with measured true/false positive counts, and Table 3 reports Post-TransClean F1 against ground truth. The evaluation loop is broken by manual/LLM labels and by final ground-truth scoring, so the result does not reduce by construction to the model's own predictions. The paper itself identifies the premise's failure mode in Section 4.2.4: 'adding a true positive edge to G_TC may lead to an increase in the number of negative transitive predictions of f_theta_TC in G_TC because the model may wrongly predict true positive transitive matches, as is particularly likely in record groups with edge cases.' This concession, together with the degraded LLM-labeling runs on Camera, Monitor, and WDC, is evidence of brittleness or correctness risk rather than circularity. No equation defines the proxy in terms of the target F1, no fitted parameter is renamed as a prediction, and no load-bearing argument rests on a self-citation: the cited prior work [12] supplies the Synthetic Companies benchmark and a blocking heuristic, but TransClean's performance is measured against that benchmark's ground truth and against external benchmarks. I also note an apparent typo in Algorithm 2 line 8, where the printed pruning condition |Pos_tr| > |Neg_tr| is the reverse of the 'more negative than positive' criterion stated in Section 4.2.3; this is an implementation/correctness concern, not a circularity. On balance, the derivation chain is self-contained and externally evaluated, so the appropriate circularity score is 0.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The method depends on a chain of modeling choices: the true matching is an equivalence relation; the model's negative transitive predictions are evidence of false positive edges; minimum cuts concentrate false positives; LLM pseudo-labels approximate manual labels; and several hyperparameters (budget, n, S, split) are chosen by hand. These are not derived and are not validated by external theory.

free parameters (7)
  • Labeling budget LB_Total = 1K or 10K per dataset
    Set by hand in Section 5.1.4; directly controls how many pairs are labeled and finetuned.
  • Number of initial finetuning steps n = 5
    Section 5.1.4; no sensitivity analysis is provided.
  • Component size threshold S = 50
    Section 5.1.4; controls when components are broken down and when transitive matches are evaluated.
  • Initial budget split = LB_Total/2
    Section 5.1.4; the allocation between initial steps and later steps is arbitrary.
  • Pruning criterion = more negative than positive transitive predictions (per text)
    Threshold for removing minimum edge cuts; no theoretical or empirical justification. Algorithm 2 pseudocode may invert this criterion.
  • Blocking top-k = 10
    Used for both models in Section 5.1.3; affects candidate pairs and recall.
  • RandomPaths selection
    Algorithm 1 calls RandomPaths but does not specify how many paths or which paths are selected; this affects which edges get labeled.
assumptions (5)
  • domain assumption True entity matching is transitive: if A matches B and B matches C, then A matches C, so the output record groups are cliques.
    Invoked in Section 4, where matching output is described as a series of cliques and transitive matches are defined from connected components.
  • ad hoc to paper A NoMatch prediction on a transitive pair implies at least one false positive edge on the path connecting that pair.
    Section 4.1 states negative transitive predictions are very likely in components with false positive matches; this is the basis for pruning. It is a heuristic, and Section 4.2.4 concedes true positive edges can also create negative transitive predictions.
  • ad hoc to paper Minimum edge cuts of transitively inconsistent components are enriched in false positives.
    Used in Algorithms 1 and 2 to select edges for labeling and pruning without proof or separate validation.
  • domain assumption DeepSeek 7B pseudo-labels are accurate enough to replace manual labels in cleanup loops.
    Section 4.2.5 uses LLM labels in Algorithms 1 and 2; the paper reports a performance drop but does not quantify label accuracy or LLM configuration.
  • ad hoc to paper The same matching model f_theta can reliably judge its own transitive matches before and after finetuning.
    Transitive Consistency computations and edge recovery in Algorithms 1 to 3 use f_theta's own predictions as evidence about f_theta's own errors.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TransClean: Finding False Positives in Multi-Source Entity Matching under Real-World Conditions via Transitive Consistency." pith.science (2026). https://pith.science/paper/CMCCRJZB

@misc{pith2026250604006,
  author       = {Pith},
  title        = {Pith review of: TransClean: Finding False Positives in Multi-Source Entity Matching under Real-World Conditions via Transitive Consistency},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CMCCRJZB}},
  note         = {Machine review of arXiv:2506.04006}
}
read the original abstract

We present TransClean, a method for detecting false positive predictions of entity matching algorithms under real-world conditions characterized by large-scale, noisy, and unlabeled multi-source datasets that undergo distributional shifts. TransClean is explicitly designed to operate with multiple data sources in an efficient, robust and fast manner while accounting for edge cases and requiring limited manual labeling. TransClean leverages the Transitive Consistency of a matching, a measure of the consistency of a pairwise matching model f_theta on the matching it produces G_f_theta, based both on its predictions on directly evaluated record pairs and its predictions on implied record pairs. TransClean iteratively modifies a matching through gradually removing false positive matches while removing as few true positive matches as possible. In each of these steps, the estimation of the Transitive Consistency is exclusively done through model evaluations and produces quantities that can be used as proxies of the amounts of true and false positives in the matching while not requiring any manual labeling, producing an estimate of the quality of the matching and indicating which record groups are likely to contain false positives. In our experiments, we compare combining TransClean with a naively trained pairwise matching model (DistilBERT) and with a state-of-the-art end-to-end matching method (CLER) and illustrate the flexibility of TransClean in being able to detect most of the false positives of either setup across a variety of datasets. Our experiments show that TransClean induces an average +24.42 F1 score improvement for entity matching in a multi-source setting when compared to traditional pair-wise matching algorithms.

Figures

Figures reproduced from arXiv: 2506.04006 by the authors.

Figure 1
Figure 1. Example of a matching involving three records [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Example record group implied by a set of pair [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The component in this figure has two true posi [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: If we consider to recover the edge between records [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Evolution of the component of Figure 2 over the [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Aggregate pairwise predictions on the Synthetic [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Visualization of relevant metrics of a matching on the Synthetic Companies dataset and their evolution over [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Entity Resolution in Practice: Lessons from a Self-Serve Pipeline

    cs.LG 2026-07 conditional novelty 4.0 of 10

    On six public benchmarks, the paper shows no single entity-matching algorithm dominates and recommends an automatic matcher tournament, separate precision/recall mechanisms, and verified-merge clustering to stop casca...

Reference graph

Works this paper leans on

60 extracted references · 38 canonical work pages · cited by 1 Pith paper

  1. [1]

    Kalashnikov, and Sharad Mehrotra

    Yasser Altowim, Dmitri V. Kalashnikov, and Sharad Mehrotra. 2018. Pro- gressER: Adaptive Progressive Approach to Relational Entity Resolution. ACM Trans. Knowl. Discov. Data12, 3, Article 33 (March 2018), 45 pages. https://doi.org/10.1145/3154410

    work page doi:10.1145/3154410 2018

  2. [2]

    Nils Barlaug. 2020. Tailoring Entity Matching for Industrial Settings. InPro- ceedings of the 29th ACM International Conference on Information & Knowledge Management (CIKM ’20). Association for Computing Machinery, New York, NY, USA, 3217–3220. https://doi.org/10.1145/3340531.3418514

    work page arXiv 2020

  3. [3]

    Indrajit Bhattacharya and Lise Getoor. 2007. Collective Entity Resolution in Relational Data.ACM Transactions on Knowledge Discovery from Data (TKDD) 1, 1 (2007), 5–es. https://doi.org/10.1145/1217299.1217304

    arXiv 2007

  4. [4]

    Lukas Bischof, Stefan Teodoropol, Rudolf M Füchslin, and Kurt Stockinger

    work page

  5. [5]

    Ursin Brunner and Kurt Stockinger. 2019. Entity matching on unstructured data: an active learning approach. In2019 6th Swiss Conference on Data Science (SDS). IEEE, 97–102

    work page 2019

  6. [6]

    Ursin Brunner and Kurt Stockinger. 2020. Entity matching with transformer architectures - a step forward in data integration.23rd International Conference on Extending Database Technology, Copenhagen, 30 March - 2 April 2020(2020). https://doi.org/10.21256/ZHAW-19637

    work page doi:10.21256/zhaw-19637 2020

  7. [7]

    Zhaoqiang Chen, Qun Chen, Boyi Hou, Murtadha Ahmed, and Zhanhuai Li. 2018. Improving Machine-based Entity Resolution with Limited Human Effort: A Risk Perspective.arXiv e-prints, Article arXiv:1805.12502 (May 2018), arXiv:1805.12502 pages. https://doi.org/10.48550/arXiv.1805.12502 arXiv:cs.DB/1805.12502

    work page Pith review arXiv doi:10.48550/arxiv.1805.12502 2018

  8. [8]

    Zhaoqiang Chen, Qun Chen, Boyi Hou, Tianyi Duan, Zhanhuai Li, and Guoliang Li. 2019. Towards Interpretable and Learnable Risk Analy- sis for Entity Resolution.arXiv e-prints, Article arXiv:1912.02947 (Dec. 2019), arXiv:1912.02947 pages. https://doi.org/10.48550/arXiv.1912.02947 arXiv:cs.DB/1912.02947

    work page Pith review arXiv doi:10.48550/arxiv.1912.02947 2019

Show all 60 references
  1. [9]

    Vassilis Christophides, Vasilis Efthymiou, Themis Palpanas, George Papadakis, and Kostas Stefanidis. 2020. An Overview of End-to-End Entity Resolution for Big Data.ACM Comput. Surv.53, 6, Article 127 (Dec. 2020), 42 pages. https://doi.org/10.1145/3418896

    2020 doi

  2. [10]

    Valter Crescenzi, Andrea De Angelis, Donatella Firmani, Maurizio Mazzei, Paolo Merialdo, Federico Piai, and Divesh Srivastava. 2021. Alaska: A Flexible Benchmark for Data Integration Tasks.arXiv e-prints, Article arXiv:2101.11259 (Jan. 2021), arXiv:2101.11259 pages. https://do...

    2021 doi arXiv

  3. [11]

    C., Chaitanya Gokhale, Pradap Konda, Yash Govind, and Derek Paulsen

    Sanjib Das, AnHai Doan, Paul Suganthan G. C., Chaitanya Gokhale, Pradap Konda, Yash Govind, and Derek Paulsen. [n.d.]. The Magellan Data Repository. https://sites.google.com/site/anhaidgroup/projects/data

  4. [12]

    Fernando de Meer Pardo, Claude Lehmann, Dennis Gehrig, Andrea Nagy, Stefano Nicoli, Branka Hadji Misheva, Martin Braschler, and Kurt Stockinger

  5. [13]

    H L Dunn. 1946. Record linkage.Am. J. Public Health Nations. Health36, 12 (Dec. 1946), 1412–1416

    1946

  6. [14]

    GraLMatch: Matching Groups of Entities with Graphs and Language Models. InProceedings 28th International Conference on Extending Database Technology, EDBT 2025, Barcelona, Spain, March 25-28, 2025, Alkis Simitsis, Bettina Kemme, Anna Queralt, Oscar Romero, and Petar Jovanovic ...

    2025 doi

  7. [15]

    Fellegi and Alan B

    Ivan P. Fellegi and Alan B. Sunter. 1969. A Theory for Record Linkage.J. Amer. Statist. Assoc.64 (1969), 1183–1210. https://api.semanticscholar.org/CorpusID: 17349112

    1969

  8. [16]

    Muhammad Ebraheem, Saravanan Thirumuruganathan, Shafiq Joty, Mourad Ouzzani, and Nan Tang. 2018. Distributed representations of tuples for entity resolution.Proc. VLDB Endow.11, 11 (July 2018), 1454–1467. https://doi.org/ 10.14778/3236187.3236198

    2018

  9. [17]

    Cheng Fu, Xianpei Han, Le Sun, Bo Chen, Wei Zhang, Suhui Wu, and Hao Kong. 2019. End-to-End Multi-Perspective Matching for Entity Res- olution. InInternational Joint Conference on Artificial Intelligence. https: //api.semanticscholar.org/CorpusID:199466280

    2019

  10. [18]

    Cheng Fu, Xianpei Han, Jiaming He, and Le Sun. 2020. Hierarchical Matching Network for Heterogeneous Entity Resolution. InProceedings of the Twenty- Ninth International Joint Conference on Artificial Intelligence, IJCAI-20, Chris- tian Bessiere (Ed.). International Joint Confe...

    2020 doi

  11. [19]

    Bar Genossar, Avigdor Gal, and Roee Shraga. 2023. The Battleship Ap- proach to the Low Resource Entity Matching Problem.arXiv e-prints, Article arXiv:2311.15685 (Nov. 2023), arXiv:2311.15685 pages. https://doi.org/10. 48550/arXiv.2311.15685 arXiv:cs.DB/2311.15685

    2023 doi arXiv

  12. [20]

    Congcong Ge, Pengfei Wang, Lu Chen, Xiaoze Liu, Baihua Zheng, and Yunjun Gao. 2023. CollaborEM: A Self-Supervised Entity Matching Framework Using Multi-Features Collaboration.IEEE Transactions on Knowledge and Data Engineering35, 12 (2023), 12139–12152. https://doi.org/10.1109...

    2023 doi

  13. [21]

    Boyi Hou, Qun Chen, Zhaoqiang Chen, Youcef Nafa, and Zhanhuai Li. 2020. r-HUMO: A Risk-Aware Human-Machine Cooperation Framework for Entity Resolution with Quality Guarantees.IEEE Transactions on Knowledge and Data Engineering32, 2 (2020), 347–359. https://doi.org/10.1109/TKDE...

    2020

  14. [22]

    Miller, and Hyun Chul Lee

    Oktie Hassanzadeh, Fei Chiang, Renée J. Miller, and Hyun Chul Lee. 2009. Framework for Evaluating Clustering Algorithms in Duplicate Detection.Proc. VLDB Endow.2 (2009), 1282–1293. https://api.semanticscholar.org/CorpusID: 13970666

    2009

  15. [23]

    Jungo Kasai, Kun Qian, Sairam Gurajada, Yunyao Li, and Lucian Popa. 2019. Low-resource Deep Entity Resolution with Transfer and Active Learning. arXiv e-prints, Article arXiv:1906.08042 (June 2019), arXiv:1906.08042 pages. https://doi.org/10.48550/arXiv.1906.08042 arXiv:cs.DB/...

    2019 doi arXiv

  16. [24]

    Kalashnikov and Sharad Mehrotra

    Dmitri V. Kalashnikov and Sharad Mehrotra. 2006. Domain-independent data cleaning via analysis of entity-relationship graph.ACM Trans. Database Syst. 31, 2 (June 2006), 716–767. https://doi.org/10.1145/1138394.1138401

    2006

  17. [25]

    Hanna Köpcke, Andreas Thor, and Erhard Rahm. 2010. Evaluation of entity resolution approaches on real-world match problems.Proc. VLDB Endow.3, 1–2 (Sept. 2010), 484–493. https://doi.org/10.14778/1920841.1920904

    2010

  18. [26]

    C., AnHai Doan, Adel Ardalan, Jeffrey R

    Pradap Konda, Sanjib Das, Paul Suganthan G. C., AnHai Doan, Adel Ardalan, Jeffrey R. Ballard, Han Li, Fatemah Panahi, Haojun Zhang, Jeff Naughton, Shishir Prasad, Ganesh Krishnan, Rohit Deep, and Vijay Raghavendra. 2016. Magellan: Toward Building Entity Matching Management Sys...

    2016

  19. [27]

    Bing Li, Wei Wang, Yifang Sun, Linhan Zhang, Muhammad Asif Ali, and Yi Wang. 2020. GraphER: Token-Centric Entity Resolution with Graph Convolu- tional Neural Networks.Proceedings of the AAAI Conference on Artificial Intelli- gence34, 05 (April 2020), 8172–8179. https://doi.org...

    2020 doi

  20. [28]

    Pigi Kouki, Jay Pujara, Christopher Marcum, Laura Koehly, and Lise Getoor

  21. [29]

    Yuliang Li, Jinfeng Li, Yoshihiko Suhara, Jin Wang, Wataru Hirota, and Wang- Chiew Tan. 2021. Deep Entity Matching: Challenges and Opportunities.J. Data and Information Quality13, 1, Article 1 (jan 2021), 17 pages. https: //doi.org/10.1145/3431816

    2021 doi

  22. [30]

    Jakub Maciejewski, Konstantinos Nikoletos, George Papadakis, and Yannis Velegrakis. 2025. Progressive Entity Matching: A Design Space Exploration. Proc. ACM Manag. Data3, 1, Article 65 (Feb. 2025), 25 pages. https://doi.org/ 10.1145/3709715

    2025 doi

  23. [31]

    Yuliang Li, Jinfeng Li, Yoshi Suhara, An Hai Doan, and Wang Chiew Tan. 2023. Effective entity matching with transformers.VLDB Journal(1 2023), 1–21. https://doi.org/10.1007/S00778-023-00779-Z/TABLES/14

    2023 doi

  24. [32]

    Avanika Narayan, Ines Chami, Laurel Orr, Simran Arora, and Christopher Ré. 2022. Can Foundation Models Wrangle Your Data?arXiv e-prints, Article arXiv:2205.09911 (May 2022), arXiv:2205.09911 pages. https://doi.org/10. 48550/arXiv.2205.09911 arXiv:cs.LG/2205.09911

    2022 doi arXiv

  25. [33]

    Markus Nentwig, Anika Groß, and Erhard Rahm. 2016. Holistic Entity Cluster- ing for Linked Data. In2016 IEEE 16th International Conference on Data Mining Workshops (ICDMW). 194–201. https://doi.org/10.1109/ICDMW.2016.0035

    2016

  26. [34]

    Zhengjie Miao, Yuliang Li, and Xiaolan Wang. 2021. Rotom: A meta-learned data augmentation framework for entity matching, data cleaning, text clas- sification, and beyond. InProceedings of the 2021 International Conference on Management of Data. 1303–1316

    2021

  27. [35]

    George Papadakis, George Mandilaras, Luca Gagliardelli, Giovanni Simonini, Emmanouil Thanos, George Giannakopoulos, Sonia Bergamaschi, Themis Palpanas, and Manolis Koubarakis. 2020. Three-dimensional Entity Resolution with JedAI.Information Systems93 (05 2020), 101565. https:/...

    2020

  28. [36]

    Ralph Peeters and Christian Bizer. 2022. Supervised Contrastive Learning for Product Matching.Companion Proceedings of the Web Conference 2022(2022)

    2022

  29. [37]

    Rahul Pandey, Hemant Purohit, Carlos Castillo, and Valerie L. Shalin. 2022. Modeling and mitigating human annotation errors to design efficient stream processing systems with human-in-the-loop machine learning.International Journal of Human-Computer Studies160 (2022), 102772. ...

    2022

  30. [38]

    Ralph Peeters, Aaron Steiner, and Christian Bizer. 2023. Entity Matching using Large Language Models.arXiv e-prints, Article arXiv:2310.11244 (Oct. 2023), arXiv:2310.11244 pages. https://doi.org/10.48550/arXiv.2310.11244 arXiv:cs.CL/2310.11244

    2023 doi arXiv

  31. [39]

    Anna Primpeli and Christian Bizer. 2021. Graph-Boosted Active Learning for Multi-source Entity Resolution. InThe Semantic Web – ISWC 2021: 20th International Semantic Web Conference, ISWC 2021, Virtual Event, October 24–28, 2021, Proceedings. Springer-Verlag, Berlin, Heidelber...

    2021 doi

  32. [40]

    Ralph Peeters, Reng Chiz Der, and Christian Bizer. 2023. WDC Products: A Multi-Dimensional Entity Matching Benchmark.arXiv e-prints, Article arXiv:2301.09521 (Jan. 2023), arXiv:2301.09521 pages. https://doi.org/10.48550/ arXiv.2301.09521 arXiv:cs.LG/2301.09521

    2023 doi arXiv

  33. [41]

    2017.Comparative Evaluation of Distributed Clustering Schemes for Multi-source Entity Resolution

    Alieh Saeedi, Eric Peukert, and Erhard Rahm. 2017.Comparative Evaluation of Distributed Clustering Schemes for Multi-source Entity Resolution. Springer In- ternational Publishing, 278–293. https://doi.org/10.1007/978-3-319-66917-5_ 19

    2017 doi

  34. [42]

    Alieh Saeedi, Eric Peukert, and Erhard Rahm. 2018. Using Link Features for Entity Clustering in Knowledge Graphs. InThe Semantic Web: 15th In- ternational Conference, ESWC 2018, Heraklion, Crete, Greece, June 3–7, 2018, Proceedings. Springer-Verlag, Berlin, Heidelberg, 576–592...

    2018

  35. [43]

    Alieh Saeedi, Markus Nentwig, Eric Peukert, and Erhard Rahm. 2018. Scalable Matching and Clustering of Entities with FAMER.Complex Systems Informatics and Modeling Quarterly16 (Oct. 2018), 61–83. https://doi.org/10.7250/csimq. 2018-16.04

    2018 doi

  36. [44]

    Giovanni Simonini, Luca Zecchini, Sonia Bergamaschi, and Felix Naumann

  37. [45]

    Saravanan Thirumuruganathan, Han Li, Nan Tang, Mourad Ouzzani, Yash Govind, Derek Paulsen, Glenn Fung, and AnHai Doan. 2021. Deep Learning for Blocking in Entity Matching: A Design Space Exploration.Proc. VLDB Endow.14, 11 (jul 2021), 2459–2472. https://doi.org/10.14778/347624...

    2021

  38. [46]

    Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. 2019. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. arXiv e-prints, Article arXiv:1910.01108 (Oct. 2019), arXiv:1910.01108 pages. https://doi.org/10.48550/arXiv.1910.01108 arXiv:cs....

    2019 doi arXiv

  39. [47]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Atten- tion Is All You Need.arXiv e-prints, Article arXiv:1706.03762 (June 2017), arXiv:1706.03762 pages. https://doi.org/10.48550/arXiv.1706.037...

    2017 doi arXiv

  40. [48]

    Runhui Wang, Yuliang Li, and Jin Wang. 2022. Sudowoodo: Contrastive Self-supervised Learning for Multi-purpose Data Integration and Preparation. arXiv:cs.DB/2207.04122

    2022 arXiv

  41. [49]

    Tianshu Wang, Xiaoyang Chen, Hongyu Lin, Xuanang Chen, Xianpei Han, Hao Wang, Zhenyu Zeng, and Le Sun. 2024. Match, Compare, or Select? An Investigation of Large Language Models for Entity Matching.arXiv e- prints, Article arXiv:2405.16884 (May 2024), arXiv:2405.16884 pages. h...

    2024 doi arXiv

  42. [50]

    Venkata Vamsikrishna Meduri, Lucian Popa, Prithviraj Sen, and Mohamed Sarwat. 2020. A Comprehensive Benchmark Framework for Active Learning Methods in Entity Matching.arXiv e-prints, Article arXiv:2003.13114 (March 2020), arXiv:2003.13114 pages. https://doi.org/10.48550/arXiv....

    2020 doi arXiv

  43. [51]

    Robert Wrembel. 2022. Data Integration, Cleaning, and Deduplication: Re- search Versus Industrial Projects. InInformation Integration and Web Intelli- gence, Eric Pardede, Pari Delir Haghighi, Ismail Khalil, and Gabriele Kotsis (Eds.). Springer Nature Switzerland, Cham, 3–17

    2022

  44. [52]

    Shiwen Wu, Qiyu Wu, Honghua Dong, Wen Hua, and Xiaofang Zhou. 2024. Blocker and Matcher Can Mutually Benefit: A Co-Learning Framework for Low-Resource Entity Resolution.Proceedings of the VLDB Endowment17 (01 2024), 292–304. https://doi.org/10.14778/3632093.3632096

    2024

  45. [53]

    Dezhong Yao, Yuhong Gu, Gao Cong, Hai Jin, and Xinqiao Lv. 2022. Entity Resolution with Hierarchical Graph Attention Networks.Proceedings of the 2022 International Conference on Management of Data(2022)

    2022

  46. [54]

    Steven Euijong Whang, David Marmaros, and Hector Garcia-Molina. 2013. Pay-As-You-Go Entity Resolution.IEEE Transactions on Knowledge and Data Engineering25, 5 (2013), 1111–1124. https://doi.org/10.1109/TKDE.2012.43

    2013 doi

  47. [55]

    Zeyu Zhang, Paul Groth, Iacer Calixto, and Sebastian Schelter. 2025. A Deep Dive Into Cross-Dataset Entity Matching with Large and Small Language Models. https://doi.org/10.48786/EDBT.2025.75

    2025 doi

  48. [56]

    Chen Zhao and Yeye He. 2019. Auto-EM: End-to-end Fuzzy Entity-Matching using Pre-trained Deep Models and Transfer Learning. InThe World Wide Web Conference (WWW ’19). Association for Computing Machinery, New York, NY, USA, 2413–2424. https://doi.org/10.1145/3308558.3313578

    2019

  49. [58]

    Wei Zhang, Hao Wei, Bunyamin Sisman, Xin Luna Dong, Christos Falout- sos, and David Page. 2019. AutoBlock: A Hands-off Blocking Frame- work for Entity Matching.arXiv e-prints, Article arXiv:1912.03417 (Dec. 2019), arXiv:1912.03417 pages. https://doi.org/10.48550/arXiv.1912.034...

    2019 doi arXiv

  50. [2017]

    In2017 IEEE Interna- tional Conference on Data Mining (ICDM)

    Collective Entity Resolution in Familial Networks. In2017 IEEE Interna- tional Conference on Data Mining (ICDM). 227–236. https://doi.org/10.1109/ ICDM.2017.32

    2017

  51. [2022]

    VLDB Endow.15, 7 (March 2022), 1506–1518

    Entity resolution on-demand.Proc. VLDB Endow.15, 7 (March 2022), 1506–1518. https://doi.org/10.14778/3523210.3523226

    2022

  52. [2025]

    Hybrid quantum neural networks show strongly reduced need for free parameters in entity matching.Scientific Reports15, 1 (2025), 4318

    2025

Pith tools

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