A Review of the Marathi Natural Language Processing

Asang Dani; Shailesh R Sathe

REVIEW 5 major objections 6 minor 1 cited by

A Review of the Marathi Natural Language Processing

T0 review · 5 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read This survey maps the evolution of Marathi NLP from early rule-based tools to modern multilingual transformer models and identifies the resources and gaps that define the field.

desk verdict A useful but uneven survey of Marathi NLP; the resource coverage is broad, but the 'state-of-the-art' claim is under-supported and several technical details are wrong. read the letter →

arxiv 2412.15471 v2 pith:NJQEKCXJ submitted 2024-12-20 cs.CL

Asang Dani , Shailesh R Sathe This is my paper · ORCID

classification cs.CL

keywords MarathiNLPIndiclanguageslanguageresourcesneuralmachinetranslationmultilingualmodelsevaluationmetricscode-mixingsurvey

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

This paper is a survey of Marathi natural language processing. It argues that NLP advances did not reach Marathi quickly because of script diversity, scarce public resources, and the language's rich morphology, but that the past decade of Indic-language initiatives has changed the situation. The survey's central claim is that Marathi NLP now has a broad, if uneven, set of corpora, pretrained models, benchmarks, and evaluation metrics, and it maps those resources onto the stages of a neural NLP pipeline. A reader interested in building or evaluating Marathi NLP systems would use this as a starting map of what exists and where the gaps are.

What carries the argument

The organizing device is the nine-step neural NLP processing pipeline—data collection and preprocessing, tokenization, embedding creation, model training, evaluation, fine-tuning, inference, post-processing, and deployment—used as a checklist against which every Marathi resource and tool is placed. The paper pairs this pipeline with a taxonomy of corpora (monolingual versus parallel) and of evaluation metrics (string-based versus model-based). This is what lets the survey convert a list of resources into a statement about where Marathi NLP stands.

What would settle it

Compare the survey's resource inventory against a systematic, time-bounded search of Marathi NLP publications and public model hubs; if major recent datasets or models are absent, or if reported corpus sizes and benchmark numbers do not match the cited releases, the 'state-of-the-art' framing fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is a synthesis: Marathi has moved from being a low-resource language with early rule-based morphological analyzers and small corpora to one with large monolingual and parallel corpora (the paper cites, among others, a 142-million-word IndicNLP corpus and a 752-million-token MahaCorpus), multilingual and Marathi-specific transformer models (mBERT, XLM-R, mT5, IndicBERT, IndicBART, MahaBERT, and IndicTrans2, which covers all 22 scheduled Indian languages), and a set of evaluation practices that favor character-level and model-based metrics such as chrF++, BERTScore, and COMET over raw BLEU. It also identifies what remains missing: code-mixed Marathi-English and Marathi-Hindi data, annotated machine-reading-comprehension and summarization datasets, and robust handling of dialect variation in tasks like text-to-speech.

Load-bearing premise

The survey's overall picture of Marathi NLP depends on the assumption that the corpora, models, and tools it chose to include are representative of the field, since it does not state a search protocol, inclusion criterion, or cutoff date.

Editorial extensions

If this is right

Marathi NLP no longer lacks basic building blocks: large monolingual corpora and English-Marathi parallel corpora are publicly available, so new work can start from pretrained models rather than from data collection.
Because the paper reports that monolingual Marathi models (MahaBERT and related models) claim state-of-the-art results on sentiment analysis, NER, and text classification, a practitioner would reasonably start from these rather than from multilingual models for Marathi tasks.
Evaluation of Marathi generation should move beyond BLEU toward character-level metrics like chrF++ and model-based metrics like COMET, since the paper presents these as better correlated with human judgment for morphologically rich languages.
IndicTrans2, as presented, provides an accessible translation baseline for all 22 scheduled languages, with compact distilled variants that are deployable in limited-resource settings.
The gaps the paper names—code-mixed text, MRC, summarization, and TTS dialect variation—are the likely next bottlenecks for Marathi NLP.

Reading between the lines

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

If the survey's inventory is representative, the practical bottleneck in Marathi NLP has shifted from raw data scarcity to scarcity of high-quality annotated task data, which suggests annotation efforts may now have higher marginal value than further corpus crawling.
The same resource pattern likely holds for other moderately resourced Indo-Aryan languages, so the pipeline-and-corpus map in this survey could serve as a template for surveys of Hindi, Gujarati, or Bengali NLP.
A testable extension would be to track the growth of Marathi corpora by year and correlate corpus size with benchmark gains; the paper's data (from 2.2 million words in EMILLE to 142 million in IndicNLP to 752 million tokens in MahaCorpus) suggests a steep recent takeoff.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

Short version: if you need a quick map of Marathi NLP resources, this is a reasonable starting point. It has no new science, and the survey's own framing overreaches. But it does collect a solid set of references.

What's good: the paper touches the major corpora (EMILLE, IndicNLP, Samanantar, FLORES, OPUS), the main models (mBERT, XLM-R, IndicBERT, IndicBART, IndicTrans2, MahaBERT), and evaluation metrics. For a newcomer, that list alone is useful. The pipeline figure and the narrative from rule-based to neural are fine.

Where it gets shaky: the authors call it 'state-of-the-art' but give no search protocol, inclusion criteria, or cutoff date. You can't verify that anything is missing or that the SOTA claim is current. In a survey that's a real gap.

Also, there are checkable mistakes. Section 6.1 says BERT 'is also known as mBERT' — that's wrong; mBERT is the multilingual variant. The hyperparameter definitions are flipped: A is attention heads, not attention layers; H is hidden size, not hidden layers. Section 7.3 reports XLM-RoBERTa as '279M parameters' — the XLM-R paper lists 270M for base and 559M for large. These are the kind of errors that a reader will hit immediately. They are not fatal, but for a survey whose sole product is accurate summary, they matter.

I'm being asked whether this deserves serious refereeing. I'd say yes, if the venue publishes surveys. The topic is real and the consolidation has value. But it needs a real revision: state the scope and cutoff, fix the technical inaccuracies, and probably tone down the 'state-of-the-art' language. If you want to use it as a citation anchor, wait until it's cleaned up.

Bottom line: useful as a working document, not as a citable reference yet.

Referee Report

5 major / 6 minor

Summary. This manuscript is a survey of Marathi NLP covering the evolution of Indic NLP research, training corpora and benchmarks, tokenization methods, neural models (BERT, BART, mT5, MahaBERT, XLM-R, IndicBERT, IndicBART, IndicTrans2), and evaluation metrics (BLEU, chrF++, BERTScore, COMET, ROUGE, BLEURT). The abstract claims to provide a broad overview of the field and of state-of-the-art resources and tools for Marathi. The paper cites many primary sources and positions itself relative to the earlier survey by Lahoti et al. (2022), but it does not describe any search protocol, inclusion criteria, or cutoff date for the resources surveyed, and it contains several checkable technical inaccuracies in model descriptions.

Significance. If made accurate, this survey would be a useful entry point for researchers new to Marathi NLP: it assembles in one place the main corpora (EMILLE, IndicNLP, Samanantar, FLORES, OPUS/NLLB), the main models (MahaBERT family, mT5, XLM-R, IndicBERT, IndicBART, IndicTrans2), and the main evaluation metrics, with references to the primary literature. Its strengths are breadth of coverage and the inclusion of very recent resources such as IndicTrans2 and mahaNLP. However, because the paper's only product is a trustworthy summary of the field, its reliability is load-bearing: the lack of a documented selection methodology and the presence of concrete factual errors in resource descriptions currently undercut the 'state-of-the-art' claim. The paper has no machine-checked proofs or code, but that is not expected for a survey; the obligation is instead accuracy and transparency of coverage.

major comments (5)

[Abstract, Section 2] The central claim of a 'broad overview' and 'state-of-the-art resources and tools' is not verifiable because the manuscript never states how the resource inventory was assembled. Section 2 compares the paper with Lahoti et al. (2022) but provides no search protocol, no inclusion or exclusion criteria, and no cutoff date. Please add a methodology paragraph specifying the databases searched, the time window, and the criteria for including or excluding resources; this is essential for a survey whose value depends on representativeness and currency.
[Section 7.3] The reported parameter count for XLM-RoBERTa is incorrect. The text states 'Their best model XLM-RoBERTa (279M parameters)', but Conneau et al. (2020) report XLM-R Base at approximately 270M parameters and XLM-R Large at approximately 559M parameters. Since readers may rely on this survey to choose models, this factual error should be corrected with the specific variant and its actual parameter count.
[Section 6.1] The description of BERT conflates BERT with mBERT. The sentence 'It is also known as mBERT or multilingual-BERT' is inaccurate: BERT is the monolingual English model, while mBERT is the multilingual variant trained on 104 languages including Marathi. The distinction matters because Section 7.3 and Section 7.4 compare XLM-R and IndicBERT against mBERT. Please revise the text to distinguish the two models explicitly.
[Section 6.1] The BERT hyperparameter definitions are mislabeled. The manuscript says 'A (the number of Attention Layers), L (the number of Encoder Layers), and H (the number of Hidden Layers)', but in the BERT paper A is the number of attention heads, L is the number of transformer layers (encoder blocks), and H is the hidden size. This makes the subsequent values BER T BASE (L=12, H=768, A=12) confusing. Please correct these definitions.
[Section 7.6] The assertion 'Current SOTA (State-Of-The-Art) model is NLLB 54B MOE' is unsupported and time-dependent. No benchmark, task, or leaderboard is cited for this claim, and 'current' is meaningless without a cutoff date. Even if intended as 'state of the art for supervised machine translation at the time of writing', the claim needs a citation, a date, and a scope restriction; otherwise it should be removed or qualified.

minor comments (6)

[Section 7.6] The final paragraph contains typographical and grammatical errors: 'its too large to be to be deployed' should be 'it is too large to be deployed', and 'However, its too large' should be 'However, it is too large'.
[Section 6.1] The section heading 'BER T' has an unintended space and should be 'BERT'.
[Section 8.2] The sentence 'It uses sacreBLEU to compute the scores' is imprecise: chrF++ is an evaluation metric, and sacreBLEU is a software implementation that can compute it. Please rephrase to clarify that the authors refer to the reference implementation available through sacreBLEU.
[Section 7.2] Several performance claims are reported from the original papers, but the text sometimes says 'Authors claim' and sometimes states results as fact. Please consistently mark third-party claims as claims, since this survey does not independently evaluate them.
[References] There are minor reference formatting issues: 'Tomá Mikolov' should be 'Tomáš Mikolov', 'A vik Bhattacharyya' should be 'Avik Bhattacharyya', and 'Gokul N.C.' is inconsistently spaced. These should be cleaned up.
[Section 9] The conclusion states 'the morphological richness make it harder for the models to deal with dialect variation' and 'Cross-lingual Information Retrieval (IR) and Question Answering (QA) is quite limited'; these subject-verb agreement errors should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: this is a survey that summarizes external resources and models, with no fitted parameters, predictions, or derivation chain that reduces to its inputs.

full rationale

The paper is a literature and resource survey. Its load-bearing content is descriptive: it reports corpora sizes, model families, benchmarks, and evaluation metrics, all attributed to external sources (EMILLE, IndicNLP, Samanantar, FLORES-200, XLM-R, IndicBART, IndicTrans2, etc.). There is no derivation, no fitted parameter, no self-cited uniqueness or ansatz, and no prediction that is later validated against the same data used to construct it. The authors do not cite their own prior work at all, and no load-bearing argument is justified by a self-citation chain. The skeptical concerns about the survey—that its resource selection has no stated protocol or cutoff date, and that Section 7.3 reports an inaccurate XLM-R parameter count (279M instead of the published 270M/559M)—are accuracy and scope issues, not circularity. Similarly, the 'state-of-the-art' framing and the statement 'Current SOTA model is NLLB 54B MOE' are external factual claims whose support is not internal to the paper; they may be checkable or contestable, but they do not reduce to the paper's own definitions or fitted quantities. A survey can be incomplete or mistaken without being circular. No step in the paper's reasoning equates an output with an input by construction, so the circularity score is 0.

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

The paper introduces no free parameters, no mathematical axioms, and no invented entities. As a survey, it relies only on the standard assumption that the cited external resources exist and are described accurately.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Review of the Marathi Natural Language Processing." pith.science (2026). https://pith.science/paper/NJQEKCXJ

@misc{pith2026241215471,
  author       = {Pith},
  title        = {Pith review of: A Review of the Marathi Natural Language Processing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NJQEKCXJ}},
  note         = {Machine review of arXiv:2412.15471}
}

read the original abstract

Marathi is one of the most widely used languages in the world. One might expect that the latest advances in NLP research in languages like English reach such a large community. However, NLP advancements in English didn't immediately reach Indian languages like Marathi. There were several reasons for this. They included diversity of scripts used, lack of (publicly available) resources like tokenization strategies, high quality datasets \& benchmarks, and evaluation metrics. In addition to this, the morphologically rich nature of Marathi, made NLP tasks challenging. Advances in Neural Network (NN) based models and tools since the early 2000s helped improve this situation and make NLP research more accessible. In the past 10 years, significant efforts were made to improve language resources for all 22 scheduled languages of India. This paper presents a broad overview of evolution of NLP research in Indic languages with a focus on Marathi and state-of-the-art resources and tools available to the research community. It also provides an overview of tools \& techniques associated with Marathi NLP tasks.

Figures

Figures reproduced from arXiv: 2412.15471 by the authors.

**Figure 1.** shows the pipeline for typical NLP systems based on neural networks. Step 7 is task specific and will be different for NLP tasks like NMT, NER, Abstractive Summarization, etc. 1. Data Collection & Preprocessing 2. Tokenization & Normalization 3. Word Embedding Creation 4. Model Training 5. Evaluation & Validation 6. Fine-tuning 7. Inference & Task specific processing 8. Post-processing 9. Deployment & Monitoring [P… view at source ↗

Discussion (0). Continue with ORCID 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

MahaParaphrase: A Marathi Paraphrase Detection Corpus and BERT-based Models
cs.CL 2025-08 conditional novelty 6.0 of 10

A new human-corrected Marathi paraphrase detection corpus with 8,000 pairs in five difficulty buckets, benchmarked with BERT models, with MahaBERT reaching 88.7% F1.

Reference graph

Works this paper leans on

8 extracted references · 3 canonical work pages · cited by 1 Pith paper

[5]

Proceedings of the Annual Meeting of the Association for Com- putational Linguistics , pages 7871–7880

BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension . Proceedings of the Annual Meeting of the Association for Com- putational Linguistics , pages 7871–7880. Chin-Yew Lin. 2004. ROUGE: A package for auto- matic evaluation of summaries . In Text Summa- rization Branches Out , pages 74–81, Barcel...

work page 2004
[8]

CoRR, abs/1911.00359

Ccnet: Extracting high quality mono- lingual datasets from web crawl data . CoRR, abs/1911.00359. Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V. Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, Jeﬀ Klingner, Apurva Shah, Melvin Johnson, Xiaobing Liu, ukasz Kaiser, Stephan Gouws, Yoshikiyo Kato, Taku Kudo, Hideto ...

arXiv 1911
[2008]

In Pro- ceedings of the International Conference on Ma- chine Learning and Applications (ICMLA) , pages 318–323

Statistical machine translation for indian languages: Mission hindi to marathi. In Pro- ceedings of the International Conference on Ma- chine Learning and Applications (ICMLA) , pages 318–323. Piotr Bojanowski, Édouard Grave, Armand Joulin, and Tomá Mikolov. 2017. Enriching Word Vec- tors with Subword Information . Transactions of the Association for Comp...

arXiv 2017
[2016]

54th Annual Meeting of the Association for Computational Linguistics, ACL 2016 - Long Papers , 3:1715–1725

Neural machine translation of rare words with subword units . 54th Annual Meeting of the Association for Computational Linguistics, ACL 2016 - Long Papers , 3:1715–1725. Shashank Siripragada, Jerin Philip, Vinay P. Nam- boodiri, and C V Jawahar. 2020. A multilin- gual parallel corpora collection eﬀort for In- dian languages . In Proceedings of the Twelfth...

arXiv 2016
[2019]

8th In- ternational Conference on Learning Representa- tions, ICLR 2020

ALBERT: A Lite BERT for Self-supervised Learning of Language Representations . 8th In- ternational Conference on Learning Representa- tions, ICLR 2020 . Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer

work page 2020
[2020]

In Proceedings of the 58th Annual Meeting of the Association for Com- putational Linguistics , pages 8440–8451, Online

Unsupervised Cross-lingual Representa- tion Learning at Scale . In Proceedings of the 58th Annual Meeting of the Association for Com- putational Linguistics , pages 8440–8451, Online. Association for Computational Linguistics. Marta R Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heaﬁeld, Kevin Heﬀer- nan, Elahe Kalbassi, Janice Lam, Daniel...

arXiv 2022
[2022]

ACM Transactions on Asian and Low-Resource Language Information Processing , 22(2)

A Survey on NLP Resources, Tools, and Techniques for Marathi Language Processing . ACM Transactions on Asian and Low-Resource Language Information Processing , 22(2). Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut

work page
[2023]

mahaNLP: A Marathi Natural Language Processing Library . eprint. Tomas Mikolov, Kai Chen, Greg Corrado, and Jef- frey Dean. 2013a. Eﬃcient Estimation of Word Representations in Vector Space . 1st Interna- tional Conference on Learning Representations, ICLR 2013 - Workshop Track Proceedings . Tomas Mikolov, Wen-tau Yih, and Geoﬀrey Zweig. 2013b. Linguistic...

arXiv 2013

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [5]

Proceedings of the Annual Meeting of the Association for Com- putational Linguistics , pages 7871–7880

BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension . Proceedings of the Annual Meeting of the Association for Com- putational Linguistics , pages 7871–7880. Chin-Yew Lin. 2004. ROUGE: A package for auto- matic evaluation of summaries . In Text Summa- rization Branches Out , pages 74–81, Barcel...

work page 2004

[2] [8]

CoRR, abs/1911.00359

Ccnet: Extracting high quality mono- lingual datasets from web crawl data . CoRR, abs/1911.00359. Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V. Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, Jeﬀ Klingner, Apurva Shah, Melvin Johnson, Xiaobing Liu, ukasz Kaiser, Stephan Gouws, Yoshikiyo Kato, Taku Kudo, Hideto ...

arXiv 1911

[3] [2008]

In Pro- ceedings of the International Conference on Ma- chine Learning and Applications (ICMLA) , pages 318–323

Statistical machine translation for indian languages: Mission hindi to marathi. In Pro- ceedings of the International Conference on Ma- chine Learning and Applications (ICMLA) , pages 318–323. Piotr Bojanowski, Édouard Grave, Armand Joulin, and Tomá Mikolov. 2017. Enriching Word Vec- tors with Subword Information . Transactions of the Association for Comp...

arXiv 2017

[4] [2016]

54th Annual Meeting of the Association for Computational Linguistics, ACL 2016 - Long Papers , 3:1715–1725

Neural machine translation of rare words with subword units . 54th Annual Meeting of the Association for Computational Linguistics, ACL 2016 - Long Papers , 3:1715–1725. Shashank Siripragada, Jerin Philip, Vinay P. Nam- boodiri, and C V Jawahar. 2020. A multilin- gual parallel corpora collection eﬀort for In- dian languages . In Proceedings of the Twelfth...

arXiv 2016

[5] [2019]

8th In- ternational Conference on Learning Representa- tions, ICLR 2020

ALBERT: A Lite BERT for Self-supervised Learning of Language Representations . 8th In- ternational Conference on Learning Representa- tions, ICLR 2020 . Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer

work page 2020

[6] [2020]

In Proceedings of the 58th Annual Meeting of the Association for Com- putational Linguistics , pages 8440–8451, Online

Unsupervised Cross-lingual Representa- tion Learning at Scale . In Proceedings of the 58th Annual Meeting of the Association for Com- putational Linguistics , pages 8440–8451, Online. Association for Computational Linguistics. Marta R Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heaﬁeld, Kevin Heﬀer- nan, Elahe Kalbassi, Janice Lam, Daniel...

arXiv 2022

[7] [2022]

ACM Transactions on Asian and Low-Resource Language Information Processing , 22(2)

A Survey on NLP Resources, Tools, and Techniques for Marathi Language Processing . ACM Transactions on Asian and Low-Resource Language Information Processing , 22(2). Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut

work page

[8] [2023]

mahaNLP: A Marathi Natural Language Processing Library . eprint. Tomas Mikolov, Kai Chen, Greg Corrado, and Jef- frey Dean. 2013a. Eﬃcient Estimation of Word Representations in Vector Space . 1st Interna- tional Conference on Learning Representations, ICLR 2013 - Workshop Track Proceedings . Tomas Mikolov, Wen-tau Yih, and Geoﬀrey Zweig. 2013b. Linguistic...

arXiv 2013