HomoEnsNER: Does Language Alignment Outperform Architectural Complexity in Gujarati Named Entity Recognition?

Chandrakant K. Bhogayata

REVIEW 4 major objections 5 minor 22 references

HomoEnsNER: Does Language Alignment Outperform Architectural Complexity in Gujarati Named Entity Recognition?

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read Majority voting over five fine-tuned GujaratiBERT models achieves the highest entity-level F1 (0.8442) on Gujarati NER, beating both the single-model baseline (0.8347) and every heterogeneous alternative (lowest 0.7855).

desk verdict A clean, honestly limited empirical comparison of homogeneous vs heterogeneous ensembling for Gujarati NER; the headline gap is real but rests on single runs, so the broader claim outruns the evidence. read the letter →

arxiv 2608.03105 v3 pith:MNQLPIQF submitted 2026-08-04 cs.CL

Chandrakant K. Bhogayata This is my paper · ORCID

classification cs.CL

keywords GujaratiNERhomogeneousensemblelanguagealignmentBERTmajorityvotinglow-resourceNLPNaamapadamdatasetnamedentityrecognition

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 asks whether, for a low-resource morphologically rich language like Gujarati, an ensemble of identical language-aligned encoders beats ensembles built from diverse architectures. It answers yes: five independently fine-tuned GujaratiBERT models combined by majority voting (HomoEnsNER) reach entity-level F1 0.8442 on the Naamapadam Gujarati test split, exceeding the single GujaratiBERT baseline (0.8347) and every heterogeneous alternative, including mixtures with MuRIL, IndicBERT, mBERT, BiLSTM, and CRF, and a stacked BiLSTM-CRF-GujaratiBERT architecture (lowest 0.7855). A sympathetic reader would care because the result suggests that for low-resource Indic languages, spending budget on multiple fine-tuned instances of one strong monolingual encoder may be more effective than engineering architectural diversity.

What carries the argument

The central object is HomoEnsNER: a homogeneous ensemble of five GujaratiBERT models, each fine-tuned on the same Naamapadam Gujarati training split with a different random seed and a learning rate sampled from 1.5e-05 to 3.5e-05, combined by majority voting. The comparison set includes a single GujaratiBERT baseline, three heterogeneous ensembles mixing GujaratiBERT with multilingual encoders (MuRIL-base, MuRIL-large, IndicBERT, mBERT), two ensembles mixing it with classical sequence models (BiLSTM on full GujaratiBERT embeddings, CRF on PCA-reduced 32-dimensional embeddings), and a stacked BiLSTM-CRF-GujaratiBERT model. The mechanism that carries the argument is the controlled contrast: the only configuration that varies model instances while holding language alignment constant is the one that beats the baseline.

What would settle it

Run each of the eight configurations across, say, ten seeds and report the mean and standard deviation of entity-level F1; if the 0.0095 gap between HomoEnsNER and the single baseline shrinks to within noise, or if a co-trained BiLSTM-CRF-GujaratiBERT variant with the encoder unfrozen and jointly tuned crosses 0.8442, the paper's central conclusion would not stand.

Watch

Extended reading notes

Core claim

HomoEnsNER, a homogeneous ensemble of five independently fine-tuned GujaratiBERT models aggregated by majority voting at the token or entity-span level, is the only configuration among the eight tested to improve on the single fine-tuned GujaratiBERT baseline, achieving entity-level F1 0.8442 versus 0.8347. Every heterogeneous configuration—GujaratiBERT with MuRIL-base and MuRIL-large, with IndicBERT and mBERT, with BiLSTM and CRF, with CRF alone, and a stacked BiLSTM-CRF-GujaratiBERT architecture—falls below the baseline, with degradation roughly proportional to how far the added component's language alignment or training regime departs from GujaratiBERT itself. The paper interprets this as evidence that, under a fixed 10-epoch budget, the reliability gained from language-aligned monolingual pretraining outweighs the complementary error patterns that architectural diversity is generally expected to provide.

Load-bearing premise

The heterogeneous baselines are fair representatives of architectural diversity: in particular, that training the BiLSTM and CRF components on frozen or PCA-reduced GujaratiBERT features, and running each configuration once, gives those designs a fair chance against the homogeneous ensemble.

Editorial extensions

If this is right

For Gujarati NER under the stated budget, fine-tuning several instances of one monolingual encoder and combining them via majority voting is a better use of compute than assembling heterogeneous encoders or adding sequence-modeling layers.
The advantage of architectural heterogeneity in ensembling appears conditional on having multiple individually strong, comparably reliable components, a condition that does not hold for Gujarati, where GujaratiBERT is the only genuinely strong option.
Adding multilingual encoders trained across many languages degrades ensemble performance more than adding other BERT-style models, because multilingual dilution weakens language alignment.
Jointly stacked BiLSTM-CRF layers on GujaratiBERT perform worst, likely because their slower convergence (about 6 hours 37 minutes with early stopping at epoch 2, versus 35-40 minutes per encoder) interferes with GujaratiBERT's own fine-tuning under the fixed budget.

Reading between the lines

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

Beyond the paper: if the pattern generalizes to other low-resource Indic languages with a strong monolingual encoder, the cheapest path to an NER performance gain is replication (seed and learning-rate jitter) plus majority voting, not architecture search.
Beyond the paper: the 0.0095 F1 advantage over the baseline comes from a single training run per configuration, so a multi-seed variance analysis could easily place that particular gap within noise, even though the ordering of the heterogeneous alternatives is more pronounced.
Beyond the paper: a natural testable extension is to grow the ensemble beyond five members and to replace fixed voting with learned weights; the paper's own results suggest the voting mechanism matters less than component quality, which a learned-weight version could test directly.

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.

Desk Editor's Note

read the letter

Worth a look if you do low-resource Indic NER. The paper does something genuinely new: it compares a five-member majority-vote ensemble of one monolingual encoder (GujaratiBERT) against six heterogeneous alternatives and a stacked BiLSTM-CRF on the Gujarati Naamapadam split, and it reports a clean ranking. HomoEnsNER edges the single baseline by 0.0095 F1 and beats every heterogeneous config. The method is simple and the practical recipe—fine-tune several instances of the strongest monolingual encoder, majority vote—is sensible for a tight GPU budget.

Credit where due: the paper is readable, the related-work table is useful, and the limitations section is unusually honest. It explicitly flags single runs, no variance, a single language/dataset, and the fact that the CRF/BiLSTM baselines ran on frozen or PCA-reduced features. So the authors are not hiding the confounds.

The soft spots are real and structural. The 0.0095 gain over baseline is the entire basis of the 'language alignment beats architectural complexity' claim, and with one run per configuration the ranking could easily be noise. The heterogeneous baselines are handicapped by design: the CRF gets 32-dim PCA features, the BiLSTM trains on frozen embeddings, and the budget is the same for much slower components. So the comparison tests 'architectural diversity under a fixed, possibly ungenerous budget' more than architectural diversity per se. The paper's own discussion half concedes this. The widening to other Indic languages is speculation.

Still, the paper is internally consistent, the data are public, and the central observation—homogeneous majority voting helps a little; the tested heterogeneous additions hurt—is plausible even if not proven. The abstract overreaches, but the body is careful enough that a referee could push for variance estimates and fairer baselines. For a reading group it's a decent case study in experimental design; I'd cite it as a data point with a caveat.

Recommend sending to peer review. The question is real, the experiment is reproducible, and the flaws are fixable in revision rather than fatal. A serious referee could get a solid paper out of this.

Referee Report

4 major / 5 minor

Summary. The paper proposes HomoEnsNER, a homogeneous ensemble of five independently fine-tuned GujaratiBERT models combined by majority voting, and compares it with a single GujaratiBERT baseline and six alternatives that introduce architectural or representational diversity (other multilingual encoders, classical BiLSTM/CRF components, and a stacked BiLSTM-CRF-GujaratiBERT model). All configurations are trained on the Naamapadam Gujarati NER split under a fixed epoch budget and evaluated by entity-level F1 on the held-out test split. The reported results show HomoEnsNER at 0.8442 F1, the single baseline at 0.8347, and all heterogeneous alternatives lower (0.7855 to 0.8322). The paper interprets this ranking as evidence that language-aligned homogeneous ensembling is more effective and more budget-conscious than architectural diversity for low-resource Indian-language NER.

Significance. If the ranking were robust, the result would be practically useful: it gives low-resource NER practitioners a simple recipe (multiple fine-tuned instances of one monolingual encoder plus majority voting) and raises a scientifically interesting question about when ensemble diversity helps. The paper's strengths are its clear experimental setup, the use of an external held-out test set (Naamapadam Gujarati), and the unusually explicit Limitations section, which acknowledges the single-run design and the frozen-feature treatment of the classical components. However, the central claim is currently supported only by point estimates from a single training run per configuration, and several design choices handicap the heterogeneous baselines. The significance of the paper therefore depends on whether the authors can supply the missing variance estimates and fairer comparisons.

major comments (4)

[Section 5.6 and Table 5] The central ranking rests on single-run point estimates. The gap between Model 2 (0.8442) and Model 1 (0.8347) is 0.0095 F1, and between Model 1 and Model 3 it is 0.0025 F1; with one seed per configuration these differences are within the range of typical run-to-run noise for transformer fine-tuning, as the manuscript itself concedes in Section 5.6. The abstract's claim that language alignment is 'a more effective, budget-conscious ensembling strategy' is load-bearing and requires either repeated runs (e.g., at least five seeds per configuration with mean and standard deviation) or a paired significance test over the test set. Without such evidence, the conclusion that HomoEnsNER outperforms the baseline is not statistically supported.
[Sections 3.4.1 and 3.3] HomoEnsNER varies both the random seed and the learning rate (1.5e-05 to 3.5e-05) across its five members, while the baseline is a single run at 2.0e-05. The observed 0.0095 F1 gain therefore conflates ensembling with hyperparameter and seed exploration. To isolate the effect of the ensemble itself, the authors should either compare against the best single GujaratiBERT model selected from the same learning-rate range, or construct the homogeneous ensemble from five runs at the same learning rate and report the variance. As written, the experiment cannot distinguish 'language alignment helps ensembling' from 'one of five fine-tuning runs happened to be better than one baseline run.'
[Sections 3.4.3, 3.4.4, and 5.6] The heterogeneous baselines are handicapped by design. The BiLSTM component in Models 4 and 8 operates on frozen GujaratiBERT embeddings, the CRF in Model 7 operates on PCA-reduced static features, and the jointly trained stacked Model 8 stops at epoch 2 after 6 hours 37 minutes of training. The paper's central claim that 'language alignment outperforms architectural complexity' is therefore not a clean test of architectural diversity: it conflates diversity with a particular, weaker training regime. The claim should either be rephrased to say that this specific implementation of heterogeneous components underperformed, or the classical components should be given co-adaptive training (within a comparable compute budget) before drawing the general conclusion.
[Sections 3.3, 3.8, and 5.5] The 'consistent budget' is defined by epochs and early stopping, not by compute or wall-clock time. The BiLSTM component took roughly 6 hours 37 minutes (with early stopping at epoch 2) while each pretrained-encoder fine-tuning run took 35 to 40 minutes; under the same definition of budget, the five-member HomoEnsNER requires approximately five encoder runs, so the 'budget-conscious' recommendation needs an explicit compute or time comparison. Reporting GPU-hours or total training time for each configuration would make the practical claim testable and would also clarify whether the heterogeneous models were simply undertrained relative to the homogeneous ensemble.

minor comments (5)

[Section 3.5] The voting description says predictions are combined 'at each token position (or entity span),' but the evaluation is entity-level; please specify whether voting is performed at the token level and then converted to spans, or directly at the span level, because this affects how ties are resolved.
[Section 4.4] The discussion of the voting mechanism is confounded: majority versus weighted voting was assigned by configuration family rather than varied within the same ensemble, so the statement that 'the degree of language alignment and reliability of the component models, rather than the voting mechanism per se, is the dominant factor' is not directly supported by the design.
[Table 4, row 8] For Model 8, the learning rate entry '1.0e-05 / 1.0e-03' should clarify which learning rate applies to the GujaratiBERT encoder and which to the BiLSTM/CRF layers, since the Section 3.6 text does not explain the two values.
[Reference [13]] Reference [13] describes L3Cube-HindBERT and DevBERT for Hindi and Marathi, but the paper uses it to support GujaratiBERT; please verify the citation and add the correct reference for the GujaratiBERT model (e.g., the l3cube-pune/gujarati-bert model card or the corresponding publication).
[Section 5.4] The sentence describing 'the weighted-voting configuration built from other pretrained encoders alone (Model 7, GujaratiBERT + CRF)' appears to conflate two different configurations; Model 7 is a heterogeneous ensemble with a classical CRF, not an encoder-only weighted-voting ensemble.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claim is an empirical ranking measured on an external held-out test split, with no fitted parameter renamed as a prediction and no load-bearing self-citation chain.

full rationale

This paper makes no derivation that reduces to its own inputs. The central comparison — HomoEnsNER (F1 0.8442) versus a single GujaratiBERT baseline (0.8347) and six heterogeneous alternatives — is obtained by fine-tuning all eight configurations under a stated budget and evaluating them on the held-out Gujarati Naamapadam test split (Section 3.7, Table 5). There is no fitted constant that is later called a prediction: the ensemble's majority voting and the individual models' hyperparameters are specified a priori (Sections 3.5, 3.6), and the test-set F1 values are external measurements, not recovered from the training objective by construction. The only citations that motivate the choice of GujaratiBERT ([9], [13]) are external prior works by other authors and are not load-bearing in the sense of forcing the experimental outcome; the paper's conclusion is an interpretation of measured results, not a theorem derived from those citations. The acknowledged limitations in Section 5.6 — single dataset, single language, single training run per configuration, no variance estimates, and frozen or PCA-reduced features for classical components — are threats to statistical robustness and experimental fairness, but they are not circularity: they do not make the reported ranking true by definition. Accordingly, the appropriate circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new entity, particle, force, or conserved quantity. It combines existing pretrained models, so the invented-entities ledger is empty. The free parameters and axioms above capture the hand-chosen design elements and unstated assumptions that the central comparison depends on.

free parameters (3)

CRF input PCA dimension = 32
Chosen by hand; reduces the CRF component to static low-dimensional features, which affects Model 7 and the fairness of the heterogeneous comparison.
HomoEnsNER learning-rate range = 1.5e-05 to 3.5e-05
Chosen by hand to encourage diversity among the five ensemble members; exact per-member values are not reported, so the result depends on this unspecified sampling.
Weighted voting weights = Validation F1 per component
Used for Models 5 through 7 and fit to validation data; a standard model-selection step, but a fitted input to the reported test results.

assumptions (4)

domain assumption GujaratiBERT is the appropriate strong monolingual encoder for Gujarati
The paper relies on citations [9] and [13] to assume that monolingual GujaratiBERT is the best available encoder; if another encoder were stronger, the ensemble comparison could change. Section 3.2.
domain assumption Naamapadam Gujarati training annotations are reliable despite automatic projection
The training split is automatically labeled via Samanantar projection, and the paper does not assess label noise. Section 3.1.
ad hoc to paper One training run per configuration is representative
The ranking in Table 5 is based on single point estimates per configuration; the paper admits it does not report variance across independent runs. Section 5.6.
ad hoc to paper Frozen or PCA-reduced classical components fairly represent architectural diversity
The BiLSTM and CRF components are not jointly optimized with the encoder in the ensemble settings, so the heterogeneous baselines are constrained in ways that likely lower their scores. Sections 3.4.3 and 3.4.4.

how reviews work

0 comments

Cite this review

Pith. "Pith review of HomoEnsNER: Does Language Alignment Outperform Architectural Complexity in Gujarati Named Entity Recognition?." pith.science (2026). https://pith.science/paper/MNQLPIQF

@misc{pith2026260803105,
  author       = {Pith},
  title        = {Pith review of: HomoEnsNER: Does Language Alignment Outperform Architectural Complexity in Gujarati Named Entity Recognition?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MNQLPIQF}},
  note         = {Machine review of arXiv:2608.03105}
}

read the original abstract

Named Entity Recognition (NER) for Gujarati remains underexplored, hindered by the absence of capitalization cues, rich morphology, lexical ambiguity, and free word order. Prior ensemble work has emphasized architectural diversity by combining heterogeneous classifiers, multilingual encoders, or classical sequence models, rather than exploiting language-aligned monolingual pretraining. This study asks whether, for a low-resource, morphologically rich language like Gujarati, a homogeneous ensemble of a single monolingual encoder outperforms such architectural diversity. We propose HomoEnsNER, a homogeneous ensemble of five independently fine-tuned GujaratiBERT models combined via majority voting, evaluated against a single GujaratiBERT baseline and six heterogeneous alternatives, including combinations with MuRIL-base, MuRIL-large, IndicBERT, mBERT, BiLSTM, CRF, and a stacked BiLSTM-CRF-GujaratiBERT architecture. All eight models were trained under a consistent budget and evaluated using entity-level F1 on the Naamapadam Gujarati test split. HomoEnsNER achieved the highest F1 (0.8442), surpassing the baseline (0.8347) and every heterogeneous alternative (lowest: 0.7855), indicating that language alignment is a more effective, budget-conscious ensembling strategy than architectural complexity for low-resource Indian language NER.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 16 canonical work pages

[1]

SampurNER: Fine-Grained Named Entity Recognition Dataset for 22 Indian Languages,

P. Kaushik and A. Anand, "SampurNER: Fine-Grained Named Entity Recognition Dataset for 22 Indian Languages," inProceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 37, pp. 31410–31418, 2026. DOI:https://doi.org/10.1609/aaai.v40i37.40405

work page doi:10.1609/aaai.v40i37.40405 2026
[2]

State-of-the-Art Approach for Indian Language-Based Named Entity Recognition: A Comprehensive Review,

P. Pandey and B. Nathani, "State-of-the-Art Approach for Indian Language-Based Named Entity Recognition: A Comprehensive Review," Research Square, preprint, Jan. 2024. DOI:https://doi.org/10.21203/ rs.3.rs-3827718/v1

work page 2024
[3]

Recent Developments in Named Entity Recognition Techniques for Indian Languages: A Com- 17 prehensive Review,

P. Sarmah, M. P. Lahkar, S. Kalita, and U. Sharma, "Recent Developments in Named Entity Recognition Techniques for Indian Languages: A Com- 17 prehensive Review,"Natural Language Processing Journal, vol. 14, Article 100199, 2026. DOI:https://doi.org/10.1016/j.nlp.2026.100199

work page arXiv 2026
[4]

A Multiobjective Simulated Annealing Approach for Classifier Ensemble: Named Entity Recognition in Indian Languages as Case Studies,

A. Ekbal and S. Saha, "A Multiobjective Simulated Annealing Approach for Classifier Ensemble: Named Entity Recognition in Indian Languages as Case Studies,"Expert Systems with Applications, vol. 38, no. 12, pp. 14760– 14772, 2011. DOI:https://doi.org/10.1016/j.eswa.2011.05.004

work page doi:10.1016/j.eswa.2011.05.004 2011
[5]

Systematic Literature Review on Named Entity Recognition: Approach, Method, and Application,

Warto, S. Rustad, G. F. Shidik, E. Noersasongko, Purwanto, Muljono, and D. R. I. M. Setiadi, "Systematic Literature Review on Named Entity Recognition: Approach, Method, and Application,"Statistics, Optimiza- tion & Information Computing, vol. 12, no. 4, pp. 907–942, 2024. DOI: https://doi.org/10.19139/soic-2310-5070-1631

work page doi:10.19139/soic-2310-5070-1631 2024
[6]

Attention Is All You Need,

A. Vaswani et al., "Attention Is All You Need," inAdvances in Neural Information Processing Systems 30(NeurIPS 2017), pp. 5998–6008, 2017, arXiv preprint.https://arxiv.org/pdf/1706.03762

arXiv 2017
[7]

BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding,

J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding," inPro- ceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technolo- gies (NAACL-HLT), Minneapolis, MN, USA, pp. 4171–4186, 2019. DOI: https://doi.o...

doi:10.18653/v1/n19-1423 2019
[8]

EL4NER: Ensemble Learning for Named Entity Recognition via Multiple Small-Parameter Large Language Models

Y. Xiao, J. Song, Y. Xu, R. Zhang, Y. Xiao, X. Lu, R. Zhu, B. Jiang, and J. Zhao, "EL4NER: Ensemble Learning for Named Entity Recognition via Multiple Small-Parameter Large Language Models," 2025, arXiv preprint. https://arxiv.org/pdf/2505.23038

work page Pith review arXiv 2025

Show all 22 references

[9]

L3Cube-MahaCorpus and MahaBERT: Marathi Monolingual Corpus, MarathiBERT Language Models, and Resources,

R. Joshi, "L3Cube-MahaCorpus and MahaBERT: Marathi Monolingual Corpus, MarathiBERT Language Models, and Resources," inProceedings of the WILDRE-6 Workshop within the 13th Language Resources and Eval- uation Conference (LREC 2022), Marseille, France, pp. 97–101, 2022. https://a...

2022
[10]

Unsu- pervised Cross-Lingual Representation Learning at Scale,

A. Conneau, K. Khandelwal, N. Goyal, V. Chaudhary, G. Wenzek, F. Guzmán, É. Grave, M. Ott, L. Zettlemoyer, and V. Stoyanov, "Unsu- pervised Cross-Lingual Representation Learning at Scale," 2019, arXiv preprint.https://arxiv.org/pdf/1911.02116

2019 arXiv
[11]

IndicNLPSuite: MonolingualCorpora, Evalua- tion Benchmarks and Pre-trained Multilingual Language Models for Indian Languages,

D. Kakwani, A. Kunchukuttan, S. Golla, N. Gokul, A. Bhattacharyya, M. M.Khapra, andP.Kumar, "IndicNLPSuite: MonolingualCorpora, Evalua- tion Benchmarks and Pre-trained Multilingual Language Models for Indian Languages," inFindings of the Association for Computational Linguisti...

2020
[12]

MuRIL: Multilingual Representations for Indian Lan- guages,

S. Khanuja et al., "MuRIL: Multilingual Representations for Indian Lan- guages," 2021, arXiv preprint.https://arxiv.org/pdf/2103.10730

2021 arXiv
[13]

L3Cube-HindBERT and DevBERT: Pre-trained BERT Trans- former Models for Devanagari-Based Hindi and Marathi Languages,

R. Joshi, "L3Cube-HindBERT and DevBERT: Pre-trained BERT Trans- former Models for Devanagari-Based Hindi and Marathi Languages," 2023, arXiv preprint.https://arxiv.org/pdf/2211.11418

2023 arXiv
[14]

Naamapadam: A Large-Scale Named Entity Annotated Data for Indic Languages,

A. Mhaske, H. Kedia, S. Doddapaneni, M. M. Khapra, P. Kumar, R. Murthy, and A. Kunchukuttan, "Naamapadam: A Large-Scale Named Entity Annotated Data for Indic Languages," inProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Vol- ume 1: Long...

2023 arXiv
[15]

Efficient Deep Learning Ensemble Approach for Indic Named Entity Recognition in Natural Language Processing,

R. Patel and B. Prajapati, "Efficient Deep Learning Ensemble Approach for Indic Named Entity Recognition in Natural Language Processing,"Indian Journal of Science and Technology, vol. 18, no. 38, pp. 3083–3090, 2025. DOI:https://doi.org/10.17485/IJST/v18i38.1602

2025 doi
[16]

Named Entity Recognition for Gujarati: A CRF-Based Approach,

V. Garg, N. Saraf, and P. Majumder, "Named Entity Recognition for Gujarati: A CRF-Based Approach," inMining Intelligence and Knowledge Exploration, Lecture Notes in Computer Science, vol. 8284, Cham, Switzerland: Springer, 2013. DOI:https://doi.org/10.1007/ 978-3-319-03844-5_74

2013
[17]

Named Entity Recognition from Gujarati Text Using Rule-Based Approach,

D. N. Shah and H. B. Bhadka, "Named Entity Recognition from Gujarati Text Using Rule-Based Approach," inAdvances in Intelligent Systems and Computing, vol. 736, Cham, Switzerland: Springer, 2018. DOI:https: //doi.org/10.1007/978-3-319-76348-4_76

2018 doi
[18]

Custom Named Entity Recognition for Gujarati Text Using spaCy,

K. B. Vora, A. R. Vasant, and S. Shah, "Custom Named Entity Recognition for Gujarati Text Using spaCy,"Mathematical Statistician and Engineering Applications, vol. 71, no. 3, pp. 1483–1495, 2022.https://www.philstat. org/index.php/MSEA/article/view/502/260

2022
[19]

Fine-Tuning Pre-trained Named Entity Recognition Models for Indian Languages,

S. Bahad, P. Mishra, P. Krishnamurthy, and D. Sharma, "Fine-Tuning Pre-trained Named Entity Recognition Models for Indian Languages," in Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technolo- ...

2024 doi
[20]

Role of Ensemble Learning in Identifying Hindi Names,

S. Srivastava, S. Abraham, M. Sanglikar, and D. C. Kothari, "Role of Ensemble Learning in Identifying Hindi Names,"Interna- tional Journal of Computer Science and Applications, vol. 8, no. 2, 2011.https://www.cscjournals.org/manuscript/Journals/IJCL/ Volume2/Issue1/IJCL-19.pdf 19

2011
[21]

Enhancing Hindi Named Entity Recognition Through Ensemble Learning,

A. Singh, S. Singh, and U. S. Tiwary, "Enhancing Hindi Named Entity Recognition Through Ensemble Learning," inProceedings of the 2023 10th IEEE Uttar Pradesh Section International Conference on Electrical, Elec- tronics and Computer Engineering (UPCON), Gautam Buddha Nagar, In...

2023
[22]

HybridNER: A Multi-Model En- semble Framework for Robust Named Entity Recognition—From General Domains to Adversarial GNSS Scenarios,

Y. Liu, J. Zhang, R. Luan, and X. Yu, "HybridNER: A Multi-Model En- semble Framework for Robust Named Entity Recognition—From General Domains to Adversarial GNSS Scenarios,"Sensors, vol. 26, no. 5, Article 1553, 2026. DOI:https://doi.org/10.3390/s26051553 20

2026 doi

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

SampurNER: Fine-Grained Named Entity Recognition Dataset for 22 Indian Languages,

P. Kaushik and A. Anand, "SampurNER: Fine-Grained Named Entity Recognition Dataset for 22 Indian Languages," inProceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 37, pp. 31410–31418, 2026. DOI:https://doi.org/10.1609/aaai.v40i37.40405

work page doi:10.1609/aaai.v40i37.40405 2026

[2] [2]

State-of-the-Art Approach for Indian Language-Based Named Entity Recognition: A Comprehensive Review,

P. Pandey and B. Nathani, "State-of-the-Art Approach for Indian Language-Based Named Entity Recognition: A Comprehensive Review," Research Square, preprint, Jan. 2024. DOI:https://doi.org/10.21203/ rs.3.rs-3827718/v1

work page 2024

[3] [3]

Recent Developments in Named Entity Recognition Techniques for Indian Languages: A Com- 17 prehensive Review,

P. Sarmah, M. P. Lahkar, S. Kalita, and U. Sharma, "Recent Developments in Named Entity Recognition Techniques for Indian Languages: A Com- 17 prehensive Review,"Natural Language Processing Journal, vol. 14, Article 100199, 2026. DOI:https://doi.org/10.1016/j.nlp.2026.100199

work page arXiv 2026

[4] [4]

A Multiobjective Simulated Annealing Approach for Classifier Ensemble: Named Entity Recognition in Indian Languages as Case Studies,

A. Ekbal and S. Saha, "A Multiobjective Simulated Annealing Approach for Classifier Ensemble: Named Entity Recognition in Indian Languages as Case Studies,"Expert Systems with Applications, vol. 38, no. 12, pp. 14760– 14772, 2011. DOI:https://doi.org/10.1016/j.eswa.2011.05.004

work page doi:10.1016/j.eswa.2011.05.004 2011

[5] [5]

Systematic Literature Review on Named Entity Recognition: Approach, Method, and Application,

Warto, S. Rustad, G. F. Shidik, E. Noersasongko, Purwanto, Muljono, and D. R. I. M. Setiadi, "Systematic Literature Review on Named Entity Recognition: Approach, Method, and Application,"Statistics, Optimiza- tion & Information Computing, vol. 12, no. 4, pp. 907–942, 2024. DOI: https://doi.org/10.19139/soic-2310-5070-1631

work page doi:10.19139/soic-2310-5070-1631 2024

[6] [6]

Attention Is All You Need,

A. Vaswani et al., "Attention Is All You Need," inAdvances in Neural Information Processing Systems 30(NeurIPS 2017), pp. 5998–6008, 2017, arXiv preprint.https://arxiv.org/pdf/1706.03762

arXiv 2017

[7] [7]

BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding,

J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding," inPro- ceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technolo- gies (NAACL-HLT), Minneapolis, MN, USA, pp. 4171–4186, 2019. DOI: https://doi.o...

doi:10.18653/v1/n19-1423 2019

[8] [8]

EL4NER: Ensemble Learning for Named Entity Recognition via Multiple Small-Parameter Large Language Models

Y. Xiao, J. Song, Y. Xu, R. Zhang, Y. Xiao, X. Lu, R. Zhu, B. Jiang, and J. Zhao, "EL4NER: Ensemble Learning for Named Entity Recognition via Multiple Small-Parameter Large Language Models," 2025, arXiv preprint. https://arxiv.org/pdf/2505.23038

work page Pith review arXiv 2025