Pith. sign in

REVIEW 3 major objections 7 minor 86 references

Beyond Scale and Generation: Understanding Language Model-based Entity Matching

T0 review · 3 major / 7 minor · reviewed 2026-07-31 · grok-4.5

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

Pith's one-line read When architecture, pretraining objective, and size are separated, generative entity matchers beat cross-encoders mainly under distribution shift—not by default.

desk verdict Solid factorial EM study: generative wins are mostly under shift, not universal—and the head-initialization confound is real but disclosed, not fatal. read the letter →

arxiv 2607.24688 v1 pith:XTSUBT22 submitted 2026-07-27 cs.DB cs.CLcs.LG

Zeyu Zhang , Xue Li , Iacer Calixto , Paul Groth , Sebastian Schelter This is my paper · ORCID
classification cs.DBcs.CLcs.LG
keywords entitymatchinglanguagemodelsbi-encodercross-encodergenerativematcherdistributionshiftcross-datasettransferfactorialevaluation
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

Entity matching asks whether two records name the same real-world thing. Language models can do this as bi-encoders (encode each record alone), cross-encoders (encode the pair together), or generative matchers (generate yes/no). Earlier work often mixed architecture with backbone, pretraining style, and size, so it was unclear what actually drove gains. This paper runs a controlled factorial study inside one model family—three architectures, three pretraining variants, three sizes, nine datasets, 1,215 fine-tunes—plus transfer and cost checks. It finds that bi-encoders need embedding-oriented initialization; joint pair encoding keeps a lasting edge over independent encoding; generative matchers are not universally better than cross-encoders, with gains concentrated when schemas or datasets shift; and bigger models can amplify shortcut learning instead of helping. The practical upshot is architecture choice should track data shift and deployment cost, not a story that bigger or more generative is always better.

What carries the argument

A full factorial design over three matcher architectures × three Qwen3 pretraining variants (base, instruction-tuned, embedding) × three sizes × nine datasets × five seeds (1,215 fine-tunes), with cross-dataset F1 retention and cost/throughput comparisons used to attribute performance to representation geometry, joint encoding, prediction head, scale, and shift rather than to architecture labels alone.

What would settle it

Repeat the same factorial design on another model family with matched base/embedding/instruction variants and sizes: if generative matchers then win on ordinary within-dataset splits without schema or cross-dataset shift, or if non-embedding bi-encoders fully catch embedding variants under full fine-tuning, the central attribution claim fails.

Watch

Extended reading notes

Core claim

Once matcher architecture is disentangled from model variant and size within a single family, generative matchers do not universally outperform cross-encoders. Their advantages concentrate under distribution shift—unseen schema combinations and cross-dataset transfer—while bi-encoder quality is dominated by embedding-oriented initialization, joint pair encoding retains a persistent advantage over independent encoding even at 8B, and scaling can worsen precision via shortcut learning rather than improve matching.

Load-bearing premise

The rankings and practical advice are assumed to carry over from one decoder-only family, LoRA fine-tuning, fixed prompts and serialization, and nine curated benchmark splits to other model families, full fine-tuning, and real production workloads.

Editorial extensions

If this is right

  • For in-distribution supervised pairwise matching, a medium-scale cross-encoder is a strong default on effectiveness–throughput grounds.
  • Prefer generative matchers when deployment faces unseen schema combinations or cross-dataset transfer.
  • Bi-encoders should start from embedding-oriented checkpoints and are most attractive when record embeddings can be cached after blocking.
  • Scaling model size is secondary to representative, high-coverage fine-tuning data; larger models can over-rely on unstable cues.
  • Future EM benchmarks should report shift in attribute combinations and attribute–label relationships, plus cross-dataset retention, not only in-domain F1.

Reading between the lines

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

  • Many leaderboard gains attributed to ‘LLMs for EM’ may shrink once backbone, variant, and size are held fixed—architecture bake-offs without that control are easy to misread.
  • The same shift-focused evaluation pattern would likely reorder claims in neighbouring integration tasks (schema matching, blocking, entity linking) where joint vs independent encoding also matters.
  • If shortcut amplification with scale is common, EM training sets may need explicit hard negative design around attribute-agreement cues, not just more parameters.
  • Cost-aware systems that route easy in-distribution pairs to cross-encoders and shifted pairs to generative matchers follow directly from the paper’s split of strengths.
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

3 major / 7 minor

Summary. The manuscript presents a controlled factorial study of language-model-based pairwise entity matching over bi-encoder, cross-encoder, and generative architectures; three Qwen3 variants; three sizes; nine datasets; and five seeds, totaling 1,215 LoRA fine-tuning runs. It additionally studies cross-dataset retention and training/inference cost. The authors find that embedding-oriented initialization is crucial for bi-encoders, joint encoding retains an advantage over independent encoding, generative matching is not uniformly superior but performs relatively better under schema and attribute-agreement shift and in cross-dataset transfer, and larger models can be more vulnerable to shortcut learning. Rebalanced versions of two MaChAmp datasets are used to attribute the architecture gaps to distribution shift.

Significance. If the conclusions hold, this is a useful corrective to the assumption that entity matching improves monotonically with generation or scale. The study is unusually systematic: it uses one decoder-only family, multiple pretraining variants and sizes, five seeds, controlled learning-rate selection, representation-geometry analysis, a limited full-fine-tuning control, counterfactual rebalanced splits, transfer-retention matrices, systems measurements, and released code/data. These are substantial strengths. The results could inform both architecture selection and future benchmark construction, provided the head/decision-policy comparison and the rebalanced-split interventions are made more rigorous.

major comments (3)
  1. [§2, §3.2, §5.3, §6.2] The central cross-encoder/generative comparison changes both head provenance and decision policy. The cross-encoder uses a randomly initialized classification head and a validation-selected threshold (§2, §3.2), whereas the generative matcher reuses a pretrained LM head and an untuned parsed yes/no decision. Consequently, Tables 5–6 and §6 cannot distinguish robustness of conditional generation from pretrained-head reuse or validation-threshold miscalibration under shift; the high-precision/low-recall pattern on sh makes this especially plausible. Please compare matched decision rules—e.g., fixed 0.5 for both and the same validation-tuned threshold applied to cross-encoder probability and normalized first-token yes/no probability—and, if feasible, a binary scorer initialized from the LM-head yes/no weights. Otherwise, present the result as a bundled-pipeline comparison and temper the cau
  2. [§5.3, Tables 5–6] The rebalanced sh/sr datasets are the main interventional evidence for the two shift explanations, but their construction is not specified. Please describe the sampling/reassignment algorithm; train/validation/test sizes and positive rates; entity disjointness and pair overlap; source, length, and attribute distributions; and how the quoted unseen-combination/title-collision rates were achieved. It is important to rule out that the near-ceiling results in Tables 5–6 arise from making the test partitions easier or otherwise changing more than the targeted shift. Results over several independently constructed rebalanced splits, ideally with a control rebalance that preserves the shift, would substantially strengthen the causal claim.
  3. [§5.3–§6, Table 4, Tables 5–6, Figures 3–4] Several decisive cells are highly variable—for example, cross-encoder F1 on sh is 62.8±16.9 at 4B and 65.1±14.4 at 8B, and on sr is 85.0±11.4 at 8B—yet the architecture, scaling, and transfer conclusions are drawn from means without confidence intervals or paired tests. Please show per-seed distributions and analyze paired dataset/seed outcomes with confidence intervals, hierarchical bootstrap, or a mixed-effects model. Figure 4's retention means/medians also need uncertainty that accounts for dependence among source–target pairs, together with robustness to unstable or bimodal seeds.
minor comments (7)
  1. [§5.2, Table 3] The text says the “best-performing” generative variant is Base, but Table 3 gives Embedding a nominally higher macro-F1 (86.4 versus 86.3), while Base has more wins. Please state the selection criterion and preferably show that downstream conclusions are insensitive to this tie.
  2. [§3.2, §5.1] The Qwen3 Base, Instruction, and Embedding checkpoints may differ in post-training data and recipe as well as objective. Referring to the effect simply as a “pretraining objective” is stronger than the design supports; “checkpoint/model variant” would be more accurate unless the additional confounders are discussed.
  3. [§5.1, §8, Limitations] The full-fine-tuning control covers only 0.6B bi-encoders on three datasets, so it does not literally “exclude” LoRA as a cause. More generally, the main cross-encoder/generative conclusions should be explicitly scoped to Qwen3, LoRA rank 8, and the fixed serialization/prompt protocol, particularly in the abstract and §8 recommendations.
  4. [Figure 2] The reported Spearman correlation ρ=0.59 is computed over 27 dataset–variant pairs that are not independent. Please report a p-value or confidence interval using a dataset-blocked resampling procedure, or otherwise present the result as descriptive.
  5. [§3.4, §5.3] GPT-4.1-mini differs in scale, training data, and alignment from the fine-tuned Qwen3 models. It is useful as a contemporary reference point, but the text should avoid using it as direct evidence for a generative-architecture advantage.
  6. [§7, Table 7] Table 7 should state whether source-B embeddings are precomputed, how many unique records are encoded, and whether embedding storage and cache-update costs are included. These assumptions materially affect the claimed benefit as k grows.
  7. [General presentation] There are several presentation issues: “efficieny” (§1), “from a difference source dataset” (§4.1), “This shift concerns on” (§5.3), “may to” (§6.2), and a broken space in “high-perfo rmance” (§3). Figures 4 and 5 would also benefit from larger labels and clearer panel captions.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: factorial EM study evaluates held-out and cross-dataset labels; rebalancing and retention are experimental controls, not definitional reductions.

full rationale

This paper is a controlled empirical comparison (3 architectures × 3 variants × 3 sizes × 9 datasets × 5 seeds = 1,215 runs), not a first-principles derivation. Load-bearing claims—embedding-variant geometry for bi-encoders, persistent joint-encoding advantage, generative gains concentrated under schema/class-conditional shift, cross-dataset retention, and cost frontiers—are measured against held-out test labels and external baselines (SBERT, Ditto, GPT-4.1-mini). Retention is F1(train=s,test=t)/F1(train=t,test=t), a descriptive transfer ratio, not a quantity forced by fitting the same target. Rebalanced sh/sr splits are constructed from stated mismatch statistics (unseen attribute-combination rate; title-collision rates among negatives) and used as interventions that close architecture gaps—proper controls, not fitted inputs renamed as predictions. Selecting the best variant per architecture from Table 3 before later sections is standard sequential analysis, not self-definitional. Self-citations to the authors’ prior EM work are background, not uniqueness theorems or load-bearing premises. Confounds raised elsewhere (e.g., pretrained LM head vs. from-scratch classification head) are validity concerns, not circularity. No step reduces a claimed prediction to its inputs by construction.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

Load-bearing content is experimental protocol and domain framing, not new physical entities. Claims rest on standard EM pairwise formulation, the representativeness of Qwen3 variants as proxies for pretraining objectives, LoRA as adequate adaptation, fixed serialization/instructions, and the constructed distribution-shift diagnostics on sh/sr.

free parameters (5)
  • per-configuration learning rate = selected per architecture×variant×size (not listed numerically in text)
    Chosen from five candidates on an 800-pair auxiliary mix; used for all dataset runs of that architecture/variant/size.
  • bi-encoder contrastive margin m = 0.3
    Margin in Lbi = y d^2 + (1-y) max(0, m-d)^2; fixed design choice affecting representation learning.
  • LoRA rank and alpha = rank=8, alpha=16, dropout=0.05
    Controls adapter capacity for all main runs; full-FT only in a small control.
  • training budget cap = max 400 steps; train/val capped 8000/2000 pairs
    Standardizes optimization across datasets; can under/over-fit relative to full-data training.
  • decision threshold τ / output parse rules = τ on validation; gen max 5 tokens, first word mapped case-insensitively
    Binary decisions from scores or first generated token yes/no; invalid generative outputs treated as invalid.
assumptions (6)
  • domain assumption Pairwise EM is binary classification/scoring over candidate pairs after blocking, with records linearized to text.
    §2 problem setup; standard Magellan-style framing underlying all metrics.
  • domain assumption Qwen3 base / instruct / embedding checkpoints at 0.6B, 4B, 8B adequately represent pretraining-objective and scale factors for decoder-only LLM matchers.
    §3.2; single-family design is what enables factorial control but limits external validity.
  • domain assumption LoRA fine-tuning yields architecture comparisons comparable to full fine-tuning for the main claims.
    §3.3 cites prior EM LoRA work; authors only partially stress-test with full FT on three bi-encoder cases.
  • ad hoc to paper Fixed prompts, attribute:value serialization, and yes/no labels do not drive the architecture ranking.
    Limitations explicitly leave prompt/serialization/label-word effects unisolated.
  • ad hoc to paper Rebalanced sh/sr splits isolate unseen attribute combinations and title-agreement shift as the causes of generative vs cross-encoder divergence.
    §5.3; strong internal evidence but still a constructed intervention on two datasets.
  • standard math Statistical comparisons (macro F1, Wilcoxon over nine dataset means, five seeds) suffice to rank architectures.
    Table 3 and repeated-seed tables; conventional experimental statistics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Scale and Generation: Understanding Language Model-based Entity Matching." pith.science (2026). https://pith.science/paper/XTSUBT22

@misc{pith2026260724688,
  author       = {Pith},
  title        = {Pith review of: Beyond Scale and Generation: Understanding Language Model-based Entity Matching},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XTSUBT22}},
  note         = {Machine review of arXiv:2607.24688}
}
read the original abstract

Entity matching identifies records that refer to the same real-world entity. Language models can be adapted to this task through bi-encoder, cross-encoder, and generative matcher architectures. However, prior studies often conflate matcher architecture with differences in model backbone, model variant(reflecting different pretraining objectives), and model size, making it difficult to isolate the sources of performance gains. We address this issue through a controlled factorial study spanning three matcher architectures, three model variants and three model sizes from the Qwen3 family, and nine datasets, totaling 1,215 fine-tuning runs. We also evaluate cross-dataset transferability and computational cost. Our results show that model variant is critical for bi-encoders: embedding-oriented variants provide stronger initialization and more favorable representation geometry predictive of downstream matching performance. Cross-encoders retain a consistent advantage over bi-encoders because they jointly encode record pairs rather than representing each record independently, although larger models partially narrow this gap. Generative matchers do not universally outperform cross-encoders. Instead, their advantages concentrate under distribution shift, including subtle unseen differences in record schemas and cross-dataset transfer. We further find that larger models rely more heavily on shortcut learning and therefore do not necessarily perform better. These findings clarify the factors underlying performance differences across matcher architectures and motivate future research and benchmark designs that better disentangle architectural choices from model-level factors while explicitly evaluating distribution shift and cross-dataset transferability. We release our experimental results, code, training scripts, and evaluation data at https://github.com/Jantory/llm-trained-matcher.

Figures

Figures reproduced from arXiv: 2607.24688 by the authors.

Figure 1
Figure 1. Three LLM-based architectures for entity matching. Bi-encoders encode the two records independently and compare [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Pretrained representation geometry predicts bi [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Cross-encoders vs generative matchers. (a) Mean [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Cross-dataset F1 and retention at 8B. Higher retention indicates that fine-tuning on the source preserves more of the [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Effectiveness–efficiency frontiers. The panels report [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

86 extracted references · 7 linked inside Pith

  1. [1]

    Mehdi Akbarian Rastaghi, Ehsan Kamalloo, and Davood Rafiei. 2022. Probing the robustness of pre-trained language models for entity matching. InProceedings of the 31st ACM International Conference on Information & Knowledge Management. 3786–3790

    2022

  2. [2]

    Tiago Brasileiro Araújo, Vasilis Efthymiou, Vassilis Christophides, Evaggelia Pitoura, and Kostas Stefanidis. 2025. TREATS: Fairness-aware entity resolution over streaming data.Information Systems129 (2025), 102506

    2025

  3. [3]

    Tiago Brasileiro Araújo, Vasilis Efthymiou, and Kostas Stefanidis. 2026. X- TREATS: Integrating explainability and fairness into streaming entity resolution. Information Sciences(2026), 123801. Beyond Scale and Generation: Understanding Language Model-based Entity Matching

    2026

  4. [4]

    Abhishek Arora, Emily Silcock, Melissa Dell, and Leander Heldring. 2024. Con- trastive entity coreference and disambiguation for historical texts. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 6174–6186

    2024

  5. [5]

    Oleh Astappiev, Franziska Neuhof, Marco Fisichella, and George Papadakis. 2026. SMBench: No-code benchmarking of learning-based entity matching.Information Systems(2026), 102711

    2026

  6. [6]

    Lorenzo Balzotti, Donatella Firmani, Luca Gagliardelli, and Giovanni Si- monini. 2026. Entity Resolution via Batched Oracle Queries.arXiv preprint arXiv:2606.24407(2026)

    arXiv 2026

  7. [7]

    Nils Barlaug and Jon Atle Gulla. 2021. Neural networks for entity matching: A survey.ACM Transactions on Knowledge Discovery from Data (TKDD)15, 3 (2021), 1–37

    2021

  8. [8]

    Alexander Brinkmann, Roee Shraga, and Christina Bizer. 2024. Sc-block: Su- pervised contrastive blocking within entity resolution pipelines. InEuropean Semantic Web Conference. Springer, 121–142

    2024

Show all 86 references
  1. [9]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners.Advances in neural information processing systems33 (2020), 1877–1901

    2020

  2. [10]

    Ursin Brunner and Kurt Stockinger. 2020. Entity Matching with Transformer Architectures-A Step Forward in Data Integration.. InEDBT. 463–473

    2020

  3. [11]

    Kittayaporn Chantaranimi and Juggapong Natwichai. 2023. Evaluation of Can- didate Pair Generation Strategies in Entity Matching. InInternational Conference on P2P, Parallel, Grid, Cloud and Internet Computing. Springer, 116–128

    2023

  4. [12]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human...

    2019

  5. [13]

    Xin Luna Dong and Theodoros Rekatsinas. 2018. Data integration and machine learning: A natural synergy. InProceedings of the 2018 international conference on management of data. 1645–1650

    2018

  6. [14]

    Meihao Fan, Xiaoyue Han, Ju Fan, Chengliang Chai, Nan Tang, Guoliang Li, and Xiaoyong Du. 2024. Cost-effective in-context learning for entity resolution: A design space exploration. In2024 IEEE 40th International Conference on Data Engineering (ICDE). IEEE, 3696–3709

    2024

  7. [15]

    Luca Gagliardelli, George Papadakis, Giovanni Simonini, Sonia Bergamaschi, and Themis Palpanas. 2024. GSM: A generalized approach to Supervised Meta- blocking for scalable entity resolution.Information Systems120 (2024), 102307

    2024

  8. [16]

    Robert Geirhos, Jörn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. 2020. Shortcut learning in deep neural networks.Nature Machine Intelligence2, 11 (2020), 665–673

    2020

  9. [17]

    Stefan Hegselmann, Alejandro Buendia, Hunter Lang, Monica Agrawal, Xiaoyi Jiang, and David Sontag. 2023. Tabllm: Few-shot classification of tabular data with large language models. InInternational conference on artificial intelligence and statistics. PMLR, 5549–5581

    2023

  10. [18]

    Mahmoud Mohamed Ashour Hussein. 2025. LLM-Enhanced Entity Matching: Comparative Analysis of traditional and modern techniques. (2025)

    2025

  11. [19]

    Shashank Mohan Jain. 2022. Hugging face. InIntroduction to transformers for NLP: With the hugging face library and models to solve problems. Springer, 51–67

    2022

  12. [20]

    Di Jin, Bunyamin Sisman, Hao Wei, Xin Luna Dong, and Danai Koutra. 2021. Deep transfer learning for multi-source entity linkage via domain adaptation. Proceedings of the VLDB Endowment15, 3 (2021), 465–477

    2021

  13. [21]

    Muhammad Ebraheem Saravanan Thirumuruganathan Shafiq Joty and Mourad Ouzzani Nan Tang. 2018. Distributed Representations of Tuples for Entity Resolution.Proceedings of the VLDB Endowment11, 11 (2018)

    2018

  14. [22]

    Dimitrios Karapiperis, Leonidas Akritidis, and Panayiotis Bozanis. 2026. The Impact of Fine-Tuning on Entity Resolution: An Experimental Evaluation. Knowledge-Based Systems(2026), 115427

    2026

  15. [23]

    Dimitrios Karapiperis, George Papadakis, and Vassilios Verykios. 2025. SPER: Accelerating Progressive Entity Resolution via Stochastic Bipartite Maximization. arXiv preprint arXiv:2512.23491(2025)

    2025

  16. [24]

    Jungo Kasai, Kun Qian, Sairam Gurajada, Yunyao Li, and Lucian Popa. 2019. Low- resource deep entity resolution with transfer and active learning. InProceedings of the 57th annual meeting of the association for computational linguistics. 5851– 5861

    2019

  17. [25]

    Nishadi Kirielle, Peter Christen, and Thilina Ranbaduge. 2022. TransER: Homo- geneous Transfer Learning for Entity Resolution.. InEDBT. 2–118

    2022

  18. [26]

    2018.Magellan: Toward building entity matching management systems

    Pradap Venkatramanan Konda. 2018.Magellan: Toward building entity matching management systems. The University of Wisconsin-Madison

    2018

  19. [27]

    Evgeny Krivosheev, Katsiaryna Mirylenka, Mattia Atzeni, and Paolo Scotton

  20. [28]

    Yuliang Li, Jinfeng Li, Yoshihiko Suhara, AnHai Doan, and Wang-Chiew Tan

  21. [29]

    Yuliang Li, Jinfeng Li, Yoshi Suhara, AnHai Doan, and Wang-Chiew Tan. 2023. Effective entity matching with transformers.VLDB Journal International Journal on Very Large Data Bases32, 6 (2023), 1215

    2023

  22. [30]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach.arXiv preprint arXiv:1907.11692 (2019)

    2019 arXiv

  23. [31]

    Jakub Maciejewski, Konstantinos Nikoletos, George Papadakis, and Yannis Vele- grakis. 2025. Progressive entity matching: A design space exploration.Proceedings of the ACM on Management of Data3, 1 (2025), 1–25

    2025

  24. [32]

    Sourab Mangrulkar, Sylvain Gugger, Lysandre Debut, Younes Belkada, Sayak Paul, and Benjamin Bossan. 2022. Peft: State-of-the-art parameter-efficient fine- tuning methods. (2022)

    2022

  25. [33]

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

    2021

  26. [34]

    Mohammad Hossein Moslemi, Amir Mousavi, Behshid Behkamal, and Mostafa Milani. 2026. Heterogeneity in entity matching: A survey and experimental analysis.Data & Knowledge Engineering(2026), 102575

    2026

  27. [35]

    Sidharth Mudgal, Han Li, Theodoros Rekatsinas, AnHai Doan, Youngchoon Park, Ganesh Krishnan, Rohit Deep, Esteban Arcaute, and Vijay Raghavendra. 2018. Deep learning for entity matching: A design space exploration. InProceedings of the 2018 international conference on managemen...

    2018

  28. [36]

    Konstantinos Nikoletos, Vasilis Efthymiou, George Papadakis, and Kostas Ste- fanidis. 2025. Auto-Configuring Entity Resolution Pipelines.IEEE Access(2025)

    2025

  29. [37]

    Matteo Paganelli, Francesco Del Buono, Andrea Baraldi, Francesco Guerra, et al

  30. [38]

    George Papadakis, Nishadi Kirielle, Peter Christen, and Themis Palpanas. 2024. A critical re-evaluation of record linkage benchmarks for learning-based matching algorithms. In2024 IEEE 40th International Conference on Data Engineering (ICDE). IEEE, 3435–3448

    2024

  31. [39]

    George Papadakis, Dimitrios Skoutas, Emmanouil Thanos, and Themis Palpanas

  32. [40]

    Ralph Peeters and Christian Bizer. 2022. Supervised contrastive learning for product matching. InCompanion Proceedings of the Web Conference 2022. 248– 251

    2022

  33. [41]

    Ralph Peeters and Christian Bizer. 2023. Using chatgpt for entity matching. In European Conference on Advances in Databases and Information Systems. Springer, 221–230

    2023

  34. [42]

    Ralph Peeters, Reng Chiz Der, and Christian Bizer. 2023. WDC products: A multi-dimensional entity matching benchmark.arXiv preprint arXiv:2301.09521 (2023)

    2023 arXiv

  35. [43]

    Blocking and filtering techniques for entity resolution: A survey.ACM Computing Surveys (CSUR)53, 2 (2020), 1–42

    2020

  36. [44]

    Nicholas Pulsone, Gregory Goren, and Roee Shraga. 2026. Understanding Domain-Aware Distribution Alignment in Budgeted Entity Matching.arXiv preprint arXiv:2606.27342(2026)

    2026 arXiv

  37. [45]

    Nicholas Pulsone, Roee Shraga, and Gregory Goren. 2026. BEACON: Budget- Aware Entity Matching Across Domains.Proceedings of the ACM on Management of Data4, 3 (SIGMOD (2026), 1–26

    2026

  38. [46]

    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-IJCN...

    2019

  39. [47]

    Ralph Peeters, Aaron Steiner, and Christian Bizer. 2025. Entity matching using large language models.OpenProceedings2 (2025), 529–541

    2025

  40. [48]

    Avi Schwarzschild, Zhili Feng, Pratyush Maini, Zachary C Lipton, and J Zico Kolter. 2024. Rethinking llm memorization through the lens of adversarial compression.Advances in Neural Information Processing Systems37 (2024), 56244–56267

    2024

  41. [49]

    Nima Shahbazi, Nikola Danevski, Fatemeh Nargesian, Abolfazl Asudeh, and Divesh Srivastava. 2023. Through the Fairness Lens: Experimental Analysis and Evaluation of Entity Matching.Proceedings of the VLDB Endowment16, 11 (2023), 3279–3292

    2023

  42. [50]

    Aaron Steiner and Christian Bizer. 2026. Labeling Training Data for Entity Matching Using Large Language Models.arXiv preprint arXiv:2606.28823(2026)

    2026 arXiv

  43. [51]

    Qian Ruan, Dachuan Shi, and Thomas Bauernhansl. 2025. Fine-tuning large lan- guage models with contrastive margin ranking loss for selective entity matching in product data integration.Advanced Engineering Informatics67 (2025), 103538

    2025

  44. [52]

    Yafeng Tang, Zheng Liang, Hongzhi Wang, Xiaoou Ding, Tianyu Mu, and Huan Hu. 2025. Description-Similarity Rules: Towards Flexible Feature Engineering for Entity Matching. In2025 IEEE 41st International Conference on Data Engineering (ICDE). IEEE, 1–15. Zeyu Zhang, Xue Li, Iace...

    2025

  45. [53]

    Tommaso Teofili, Donatella Firmani, Nick Koudas, Paolo Merialdo, and Divesh Srivastava. 2026. Can we trust LLM Self-Explanations for Entity Resolution? arXiv preprint arXiv:2606.01210(2026)

    2026 arXiv

  46. [54]

    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.Proceedings of the VLDB Endowment14, 11 (2021), 2459–2472

    2021

  47. [55]

    Aaron Steiner, Ralph Peeters, and Christian Bizer. 2025. Fine-tuning large lan- guage models for entity matching. In2025 IEEE 41st International Conference on Data Engineering Workshops (ICDEW). IEEE, 9–17

    2025

  48. [56]

    Jianhong Tu, Ju Fan, Nan Tang, Peng Wang, Chengliang Chai, Guoliang Li, Ruixue Fan, and Xiaoyong Du. 2022. Domain adaptation for deep entity resolution. In Proceedings of the 2022 international conference on management of data. 443–457

    2022

  49. [57]

    Jianhong Tu, Ju Fan, Nan Tang, Peng Wang, Guoliang Li, Xiaoyong Du, Xiaofeng Jia, and Song Gao. 2023. Unicorn: A unified multi-tasking model for supporting matching tasks in data integration.Proceedings of the ACM on Management of Data1, 1 (2023), 1–26

    2023

  50. [58]

    David Vos, Till Döhmen, and Sebastian Schelter. 2022. Towards parameter- efficient automation of data wrangling tasks with prefix-tuning. InNeurIPS 2022 First Table Representation Workshop

    2022

  51. [59]

    Mohamed Trabelsi, Jeff Heflin, and Jin Cao. 2022. Dame: Domain adaptation for matching entities. InProceedings of the Fifteenth ACM International Conference on Web Search and Data Mining. 1016–1024

    2022

  52. [60]

    Jin Wang, Yuliang Li, and Wataru Hirota. 2021. Machamp: A Generalized Entity Matching Benchmark. InCIKM

    2021

  53. [61]

    Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024. Improving text embeddings with large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 11897–11916

    2024

  54. [62]

    Pengfei Wang, Xiaocan Zeng, Lu Chen, Fan Ye, Yuren Mao, Junhao Zhu, and Yunjun Gao. 2022. PromptEM: prompt-tuning for low-resource generalized entity matching.Proceedings of the VLDB Endowment16, 2 (2022), 369–378

    2022

  55. [63]

    Jiannan Wang, Guoliang Li, Jeffrey Xu Yu, and Jianhua Feng. 2011. Entity matching: How similar is similar.Proceedings of the VLDB Endowment4, 10 (2011), 622–633

    2011

  56. [64]

    Runhui Wang, Yuliang Li, and Jin Wang. 2023. Sudowoodo: Contrastive self- supervised learning for multi-purpose data integration and preparation. In2023 IEEE 39th International Conference on Data Engineering (ICDE). IEEE, 1502–1515

    2023

  57. [65]

    Tianshu Wang, Xiaoyang Chen, Hongyu Lin, Xuanang Chen, Xianpei Han, Le Sun, Hao Wang, and Zhenyu Zeng. 2025. Match, compare, or select? an investigation of large language models for entity matching. InProceedings of the 31st International Conference on Computational Linguistic...

    2025

  58. [66]

    Tianshu Wang, Hongyu Lin, Cheng Fu, Xianpei Han, Le Sun, Feiyu Xiong, Hui Chen, Minlong Lu, and Xiuwen Zhu. [n.d.]. Bridging the Gap between Reality and Ideality of Entity Matching: A Revisiting and Benchmark Re-Construction. ([n. d.])

  59. [67]

    Runhui Wang, Luyang Kong, Yefan Tao, Andrew Borthwick, Davor Golac, Henrik Johnson, Shadie Hijazi, Dong Deng, and Yongfeng Zhang. 2024. Neural locality sensitive hashing for entity blocking. InProceedings of the 2024 SIAM International Conference on Data Mining (SDM). SIAM, 887–895

    2024

  60. [68]

    Ledell Wu, Fabio Petroni, Martin Josifoski, Sebastian Riedel, and Luke Zettle- moyer. 2020. Scalable zero-shot entity linking with dense entity retrieval. In Proceedings of the 2020 conference on empirical methods in natural language pro- cessing (EMNLP). 6397–6407

    2020

  61. [69]

    Renzhi Wu, Alexander Bendeck, Xu Chu, and Yeye He. 2023. Ground truth inference for weakly supervised entity matching.Proceedings of the ACM on Management of Data1, 1 (2023), 1–28

    2023

  62. [70]

    Junjie Xing, Yeye He, Mengyu Zhou, Haoyu Dong, Shi Han, Dongmei Zhang, and Surajit Chaudhuri. 2025. Table-LLM-Specialist: Language Model Specialists for Tables using Iterative Fine-tuning. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing...

    2025

  63. [71]

    Steven Euijong Whang and Hector Garcia-Molina. 2014. Incremental entity resolution on rules and data.The VLDB journal23, 1 (2014), 77–102

    2014

  64. [72]

    Dezhong Yao, Yuhong Gu, Gao Cong, Hai Jin, and Xinqiao Lv. 2022. Entity resolution with hierarchical graph attention networks. InProceedings of the 2022 International Conference on Management of Data. 429–442

    2022

  65. [73]

    Haoteng Yin, Jinha Kim, Prashant Mathur, Krishanu Sarker, and Vidit Bansal

  66. [74]

    Yu Yuan, Lili Zhao, Kai Zhang, Guangting Zheng, and Qi Liu. 2024. Do llms over- come shortcut learning? an evaluation of shortcut challenges in large language models. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 12188–12200

    2024

  67. [75]

    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)

    2025 arXiv

  68. [76]

    Alexandros Zeakis, George Papadakis, Dimitrios Skoutas, and Manolis Koubarakis. 2025. An in-depth analysis of pre-trained embeddings for entity resolution: A. Zeakis et al.The VLDB Journal34, 1 (2025), 5

    2025

  69. [77]

    Alexandros Zeakis, George Papadakis, Dimitrios Skoutas, and Manolis Koubarakis. 2026. DistillER: Knowledge Distillation in Entity Resolution with Large Language Models.arXiv preprint arXiv:2602.05452(2026)

    2026

  70. [78]

    Zeyu Zhang, Paul Groth, Iacer Calixto, and Sebastian Schelter. 2024. Directions towards efficient and automated data wrangling with large language models. In 2024 IEEE 40th International Conference on Data Engineering Workshops (ICDEW). IEEE, 301–304

    2024

  71. [79]

    Zeyu Zhang, Paul Groth, Iacer Calixto, and Sebastian Schelter. 2025. A Deep Dive Into Cross-Dataset Entity Matching with Large and Small Language Models.. In EDBT. 922–934

    2025

  72. [80]

    Alexandros Zeakis, George Papadakis, Dimitrios Skoutas, and Manolis Koubarakis. 2025. AvengER: Ensembling and Fine-Tuning LLMs for SELECT Prompts in Entity Resolution. InEuropean Semantic Web Conference. Springer, 301–320

    2025

  73. [81]

    Lingfeng Zhong, Jia Wu, Qian Li, Hao Peng, and Xindong Wu. 2023. A compre- hensive survey on automatic knowledge graph construction.Comput. Surveys 56, 4 (2023), 1–62

    2023

  74. [85]

    Zhanfang Zhao, Sung-Kook Han, and In-Mi So. 2018. Architecture of knowl- edge graph construction techniques.International Journal of Pure and Applied Mathematics118, 19 (2018), 1869–1883

    2018

  75. [2020]

    Deep entity matching with pre-trained language models.Proceedings of the VLDB Endowment14, 1 (2020), 50–60

    2020

  76. [2022]

    Analyzing how BERT performs entity matching.Proceedings of the VLDB Endowment15, 8 (2022), 1726–1738

    2022

  77. [2023]

    In2023 IEEE International Conference on Big Data (BigData)

    Graph neural networks for entity matching. In2023 IEEE International Conference on Big Data (BigData). IEEE, 6212–6214

  78. [2025]

    InFindings of the Association for Computational Linguistics: NAACL 2025

    How to Talk to Language Models: Serialization Strategies for Structured Entity Matching. InFindings of the Association for Computational Linguistics: NAACL 2025. 7836–7850

    2025

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.