pith. sign in

arxiv: 2605.23597 · v1 · pith:HZOXB3MUnew · submitted 2026-05-22 · 💻 cs.CL · cs.LG

Structure-Guided Entity Resolution: Fine-Tuning LLMs for Robust Name Matching in Complex Linguistic Contexts

Shivam Chourasia , Hitesh Kapoor , Nilesh Patil This is my paper

Pith reviewed 2026-05-25 04:24 UTC · model grok-4.3

classification 💻 cs.CL cs.LG
keywords entity resolutionname matchingLLM fine-tuningcurriculum learningmultilingual NLPIndian namesKYC
0
0 comments X

The pith

A two-phase curriculum fine-tunes LLMs to parse name structure before matching, reaching 99.02% accuracy on real Indian identity pairs.

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

The paper introduces Structure-Guided Entity Resolution, which trains an LLM first to understand the grammatical and semantic makeup of personal names and then to decide whether two names refer to the same person. This curriculum is tested on Indian records that contain transliteration differences, spelling variations, and cultural naming patterns. The resulting model records 99.02% accuracy and 0.994 F1 on a 50,000-pair held-out set and exceeds both GPT-4o few-shot prompting and single-stage fine-tuning. The system runs in production and processes matching for hundreds of millions of users. Readers care because reliable name unification supports identity verification tasks that must scale across linguistically diverse populations.

Core claim

SGER fine-tunes an LLM through a two-phase curriculum: first training the model to parse the grammatical and semantic structure of personal names, then optimizing it for binary entity matching. Evaluated on Indian identity data, the approach reaches 99.02% accuracy and an F1 of 0.994 on 50,000 held-out real-world pairs, outperforming GPT-4o few-shot prompting and single-stage fine-tuning baselines, and the resulting system is deployed in production serving 250M+ users.

What carries the argument

The two-phase curriculum that first teaches structure parsing of names and then optimizes the binary matching objective.

If this is right

  • The curriculum approach handles transliteration inconsistencies and naming variations more effectively than direct few-shot prompting.
  • High-precision matching becomes feasible for KYC compliance in large multilingual user bases.
  • Separating structure learning from the matching task improves results over single-stage fine-tuning on noisy records.
  • Production deployment demonstrates that the method scales to hundreds of millions of daily matching decisions.

Where Pith is reading between the lines

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

  • The same staged training pattern could be tested on other structured matching tasks such as address or product-name resolution.
  • Explicit structure awareness may lower the volume of labeled pairs needed to reach production accuracy in entity resolution.
  • The results point to possible gains if similar curricula are applied to other culturally variable text formats beyond names.

Load-bearing premise

The 50,000-pair held-out set is a representative, unbiased sample of real-world name-matching difficulty with no data leakage from training.

What would settle it

Collecting a fresh sample of 10,000 name pairs from the same sources after model deployment and measuring accuracy below 95% would falsify the reported performance.

Figures

Figures reproduced from arXiv: 2605.23597 by Hitesh Kapoor, Nilesh Patil, Shivam Chourasia.

Figure 1
Figure 1. Figure 1: Structure-Guided Entity Resolution (SGER) methodology. Phase 1 fine-tunes Llama 3 8B to parse names [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
read the original abstract

Matching person names across heterogeneous records is a core challenge in entity resolution, especially within linguistically and culturally complex environments. Variations in naming conventions, inconsistent transliteration across scripts, and frequent data entry errors make it difficult to unify user identities, an essential requirement for Know Your Customer (KYC) compliance. While Large Language Models have shown promise in understanding natural language, they often struggle with the structured ambiguity present in such domain-specific settings. This paper introduces Structure-Guided Entity Resolution (SGER), a novel framework that fine-tunes an LLM through a two-phase curriculum. The model is first trained to parse the grammatical and semantic structure of personal names, then optimized for the downstream task of binary entity matching. We evaluate SGER in the challenging context of Indian identity data, one of the most linguistically diverse and noisy environments globally. SGER achieves 99.02% accuracy and an F1 of 0.994 on a held-out set of 50,000 real-world pairs, outperforming GPT-4o few-shot prompting and single-stage fine-tuning baselines. The system is fully deployed in production at Dream11, the world's largest fantasy sports platform, serving 250M+ users. Our results demonstrate that curriculum-guided training enables robust, high-precision entity resolution in real-world multilingual systems at scale.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 0 minor

Summary. The paper introduces Structure-Guided Entity Resolution (SGER), a two-phase curriculum fine-tuning framework for LLMs that first parses grammatical and semantic structure of personal names and then optimizes for binary entity matching. It evaluates the approach on Indian identity data and reports 99.02% accuracy and 0.994 F1 on a held-out set of 50,000 real-world pairs, outperforming GPT-4o few-shot prompting and single-stage fine-tuning baselines, with the system deployed in production at Dream11 serving 250M+ users.

Significance. If the held-out evaluation is shown to be free of leakage and representative, the work would demonstrate that curriculum-guided fine-tuning can deliver high-precision name matching in linguistically diverse, noisy real-world settings. The reported production deployment provides concrete evidence of scalability and practical utility beyond academic benchmarks.

major comments (1)
  1. [Evaluation / Results section] The central performance claim (99.02% accuracy, F1 0.994 on 50k held-out pairs) is load-bearing for the paper's contribution, yet the evaluation section provides no description of the labeling process, train/test split methodology, entity-level deduplication, negative sampling strategy, or controls for data leakage. Without these details the independence of the test set cannot be verified and the reported metrics cannot be interpreted as evidence of generalization.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on the evaluation details. We agree that additional transparency is required to substantiate the reported results.

read point-by-point responses

  1. Referee: [Evaluation / Results section] The central performance claim (99.02% accuracy, F1 0.994 on 50k held-out pairs) is load-bearing for the paper's contribution, yet the evaluation section provides no description of the labeling process, train/test split methodology, entity-level deduplication, negative sampling strategy, or controls for data leakage. Without these details the independence of the test set cannot be verified and the reported metrics cannot be interpreted as evidence of generalization.

    Authors: We acknowledge that the manuscript does not currently provide these methodological details, which limits the ability to fully assess generalization. In the revised version we will expand the Evaluation section with a dedicated subsection on Dataset Construction and Evaluation Protocol. This will explicitly describe: the labeling process (expert annotation with agreement metrics), train/test split methodology (entity-level partitioning), entity-level deduplication steps, negative sampling strategy, and leakage controls (including verification that no shared entities or name structures cross splits). These additions will directly address the concern and allow readers to evaluate the independence of the held-out set. revision: yes

Circularity Check

0 steps flagged

No significant circularity; purely empirical evaluation with no derivations

full rationale

The paper describes a two-phase curriculum fine-tuning procedure for an LLM on name-matching data and reports accuracy/F1 on a held-out set of 50k pairs. No equations, mathematical derivations, fitted parameters renamed as predictions, or self-citation chains appear in the provided text. The central claims rest on external empirical benchmarks rather than any self-referential reduction (e.g., no Eq. X defined in terms of Y that is then 'predicted' from the same fit). This matches the default case of a self-contained empirical ML paper; the held-out set construction is an evaluation validity concern, not a circularity pattern under the enumerated kinds.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities; the approach rests on standard LLM fine-tuning assumptions not detailed here.

pith-pipeline@v0.9.0 · 5776 in / 1185 out tokens · 24249 ms · 2026-05-25T04:24:45.557441+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages

  1. [1]

    Computing Research Repository , volume =

    Fine-tuning Pre-trained Named Entity Recognition Models for Indian Languages , author =. Computing Research Repository , volume =. 2024 , url =

    work page 2024

  2. [2]

    ACM Computing Surveys , year =

    Vassilis Christophides and Vasilis Efthymiou and Themis Palpanas and George Papadakis and Kostas Stefanidis , title =. ACM Computing Surveys , year =. doi:10.1145/3418896 , url =

    work page doi:10.1145/3418896

  3. [3]

    Elmagarmid and Panagiotis G

    Ahmed K. Elmagarmid and Panagiotis G. Ipeirotis and Vassilios S. Verykios , title =. IEEE Transactions on Knowledge and Data Engineering , year =. doi:10.1109/TKDE.2007.250581 , url =

    work page doi:10.1109/tkde.2007.250581 2007

  4. [4]

    2023 , url =

    Tao Feng and Zifeng Wang and Jimeng Sun , journal =. 2023 , url =

    work page 2023

  5. [5]

    ACM Transactions on Knowledge Discovery from Data , year =

    Yifan Li and Cen Qu and Chao Li and Jia Wang , title =. ACM Transactions on Knowledge Discovery from Data , year =. doi:10.1145/3564752 , url =

    work page doi:10.1145/3564752

  6. [6]

    Computing Research Repository , volume =

    Pre-trained Language Models for Entity Matching: A Survey , author =. Computing Research Repository , volume =. 2023 , url =

    work page 2023

  7. [7]

    Computing Research Repository , volume =

    Fine-tuning Large Language Models for Entity Matching , author =. Computing Research Repository , volume =. 2024 , url =

    work page 2024

  8. [8]

    Computing Research Repository , volume =

    Disambiguate Entity Matching using Large Language Models through Relation Discovery , author =. Computing Research Repository , volume =. 2024 , url =

    work page 2024

  9. [9]

    Computing Research Repository , volume =

    Entity Matching using Large Language Models , author =. Computing Research Repository , volume =. 2024 , url =

    work page 2024

  10. [10]

    Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen

    Edward J. Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen. Computing Research Repository , volume =. 2021 , url =

    work page 2021

  11. [11]

    Computing Research Repository , volume =

    On Leveraging Large Language Models for Enhancing Entity Resolution: A Cost-efficient Approach , author =. Computing Research Repository , volume =. 2024 , url =

    work page 2024

  12. [12]

    Proceedings of the

    Cost-Effective In-Context Learning for Entity Resolution: A Design Space Exploration , author =. Proceedings of the. 2024 , pages =

    work page 2024

  13. [13]

    Beyond Full Fine-Tuning: Harnessing the Power of

    Chunlei Xin and Yaojie Lu and Hongyu Lin and Shuheng Zhou and Huijia Zhu and Weiqiang Wang and Zhongyi Liu and Xianpei Han and Le Sun , booktitle =. Beyond Full Fine-Tuning: Harnessing the Power of. 2024 , url =

    work page 2024

  14. [14]

    International Journal of Computer Vision , year =

    Curriculum Learning: A Survey , author =. International Journal of Computer Vision , year =. doi:10.1007/s11263-022-01611-x , url =

    work page doi:10.1007/s11263-022-01611-x

  15. [15]

    Computing Research Repository , volume =

    Efficient Memory Management for Large Language Model Serving with PagedAttention , author =. Computing Research Repository , volume =. 2023 , url =

    work page 2023

  16. [16]

    Computing Research Repository , volume =

    Learning Robust Named Entity Recognizers From Noisy Data With Retrieval Augmentation , author =. Computing Research Repository , volume =. 2024 , url =

    work page 2024

  17. [17]

    Khapra and Pratyush Kumar and Rudra Murthy V and Anoop Kunchukuttan , journal =

    Arnav Mhaske and Harshit Kedia and Sumanth Doddapaneni and Mitesh M. Khapra and Pratyush Kumar and Rudra Murthy V and Anoop Kunchukuttan , journal =. 2022 , url =

    work page 2022

  18. [18]

    Computing Research Repository , volume =

    JRC-Names: A freely available, highly multilingual named entity resource , author =. Computing Research Repository , volume =. 2013 , url =

    work page 2013

  19. [19]

    Fellegi and Alan B

    Ivan P. Fellegi and Alan B. Sunter , title =. Journal of the American Statistical Association , year =

    work page

  20. [20]

    Cohen and Pradeep Ravikumar and Stephen E

    William W. Cohen and Pradeep Ravikumar and Stephen E. Fienberg , title =. IJCAI Workshop on Information Integration on the Web (IIWeb) , year =

    work page

  21. [21]

    2012 , publisher =

    Peter Christen , title =. 2012 , publisher =

    work page 2012

  22. [22]

    Proceedings of NAACL-HLT , year =

    Jacob Devlin and Ming-Wei Chang and Kenton Lee and Kristina Toutanova , title =. Proceedings of NAACL-HLT , year =

    work page

  23. [23]

    2001 , howpublished =

    ISO 15919:2001---Information and documentation: Transliteration of Devanagari and related Indic scripts into Latin characters , author =. 2001 , howpublished =

    work page 2001

  24. [24]

    Journal of Big Data , volume =

    A survey on image data augmentation for deep learning , author =. Journal of Big Data , volume =. 2019 , publisher =

    work page 2019

  25. [25]

    Computing Research Repository , volume =

    The Llama 3 Herd of Models , author =. Computing Research Repository , volume =. 2024 , url =

    work page 2024