The reviewed record of science sign in
Pith

arxiv: 2604.26456 · v1 · submitted 2026-04-29 · cs.CL · cs.AI

Naamah: A Large Scale Synthetic Sanskrit NER Corpus via DBpedia Seeding and LLM Generation

Akhil Rajeev P , Annarao Kulkarni This is my paper

Reviewed by Pith2026-05-07 11:19 UTCgrok-4.3open to challenge →

classification cs.CL cs.AI
keywords SanskritNamed Entity RecognitionSynthetic DatasetDBpediaLarge Language ModelsNER CorpusClassical Language Processing
0
0 comments X

The pith

A synthetic dataset of 102,942 Sanskrit sentences is generated for named entity recognition by seeding DBpedia entities and prompting a large language model.

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

The paper seeks to overcome the lack of annotated data that slows digitisation of classical Sanskrit texts. It does so by extracting relevant entities from DBpedia and using a 24B-parameter hybrid reasoning model to embed those entities into full, grammatically natural sentences. The resulting silver-standard corpus supplies training material for transformer-based NER models. A reader would care because manual annotation at this scale is impractical, while synthetic data offers a scalable alternative for low-resource classical languages.

Core claim

We introduce Naamah, a high quality silver standard Sanskrit NER dataset comprising 102,942 sentences. We propose a methodology that combines entity extraction from DBpedia with the generative capabilities of a 24B parameter hybrid reasoning model to create grammatically natural and synthetically diverse training data. We utilize this dataset to benchmark two transformer architectures: the massive multilingual XLM RoBERTa and the parameter efficient IndicBERTv2.

What carries the argument

The DBpedia-seeded LLM generation pipeline that extracts Sanskrit entities and produces annotated sentences around them.

Load-bearing premise

DBpedia supplies accurate and relevant Sanskrit entities and the 24B-parameter model produces grammatically correct, diverse sentences without systematic errors or hallucinations that would degrade NER training quality.

What would settle it

Expert review of a random sample of generated sentences that finds frequent grammatical errors or incorrect entity labels would show the dataset does not meet the high-quality standard claimed.

read the original abstract

The digitisation of classical Sanskrit literature is impeded by a scarcity of annotated resources, particularly for Named Entity Recognition. While recent methodologies utilise generic Large Language Models (LLMs) for data augmentation, these approaches remain prone to error and often lack the reasoning depth required for classical grammar. In this work, we introduce Naamah, a high quality silver standard Sanskrit NER dataset comprising 102,942 sentences. We propose a methodology that combines entity extraction from DBpedia with the generative capabilities of a 24B parameter hybrid reasoning model to create grammatically natural and synthetically diverse training data. We utilize this dataset to benchmark two transformer architectures: the massive multilingual XLM RoBERTa and the parameter efficient IndicBERTv2.

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

2 major / 2 minor

Summary. The paper introduces Naamah, a synthetic Sanskrit NER dataset of 102,942 sentences generated by seeding entities from DBpedia and prompting a 24B-parameter hybrid reasoning LLM to produce grammatically natural sentences. It benchmarks the resulting silver-standard data on XLM-RoBERTa and IndicBERTv2.

Significance. A validated large-scale Sanskrit NER resource would address a genuine gap in annotated data for classical Indic languages and support downstream digitization efforts. The DBpedia-seeding plus LLM-generation pipeline is a plausible scalable approach for silver data creation in low-resource settings, but its utility hinges on unverified label accuracy and sentence quality.

major comments (2)
  1. [Abstract] Abstract: the central claim that Naamah constitutes a 'high quality silver standard' is unsupported because the manuscript supplies no human evaluation, inter-annotator agreement, error analysis, or quantitative comparison against any existing Sanskrit gold NER data. This directly undermines the reported benchmarking results on XLM-RoBERTa and IndicBERTv2.
  2. [Methodology] Methodology description: the assumption that DBpedia entities are relevant and accurate for classical Sanskrit and that the 24B LLM produces both grammatically correct sentences and correctly aligned NER tags is stated without any fidelity checks, hallucination analysis, or Sanskrit-specific grammatical validation. Systematic errors here would propagate directly into degraded training utility.
minor comments (2)
  1. The paper should report the exact prompting template, temperature settings, and any post-generation filtering or label-alignment heuristics used with the 24B model.
  2. Add a limitations paragraph discussing DBpedia coverage gaps for classical Sanskrit texts and potential domain shift between generated sentences and authentic literature.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments, which identify key areas where the validation of our synthetic dataset can be strengthened. We address each major comment below and indicate the revisions we will incorporate.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that Naamah constitutes a 'high quality silver standard' is unsupported because the manuscript supplies no human evaluation, inter-annotator agreement, error analysis, or quantitative comparison against any existing Sanskrit gold NER data. This directly undermines the reported benchmarking results on XLM-RoBERTa and IndicBERTv2.

    Authors: We agree that the phrasing 'high quality silver standard' in the abstract is not supported by human evaluation, IAA, or direct comparison to gold data. No large-scale gold-standard Sanskrit NER corpus exists for quantitative comparison, which reflects the data scarcity the paper addresses. We will revise the abstract to describe Naamah as a 'large-scale silver-standard' dataset generated via DBpedia seeding and LLM prompting, remove the 'high quality' qualifier, and add a limitations section that explicitly discusses the absence of human validation, potential error sources, and the preliminary nature of the benchmarking results on XLM-RoBERTa and IndicBERTv2. revision: yes

  2. Referee: [Methodology] Methodology description: the assumption that DBpedia entities are relevant and accurate for classical Sanskrit and that the 24B LLM produces both grammatically correct sentences and correctly aligned NER tags is stated without any fidelity checks, hallucination analysis, or Sanskrit-specific grammatical validation. Systematic errors here would propagate directly into degraded training utility.

    Authors: We acknowledge that the methodology section does not include explicit fidelity checks, hallucination analysis, or Sanskrit-specific grammatical validation. We will expand this section to state the core assumptions (DBpedia entity relevance for classical texts and LLM-generated sentence/tag alignment), discuss potential limitations such as temporal bias in DBpedia and possible LLM hallucinations or tag misalignments, and include qualitative examples of generated sentences. These additions will clarify the silver-standard nature of the data without claiming unverified accuracy. revision: yes

Circularity Check

0 steps flagged

No significant circularity; data-generation pipeline is self-contained

full rationale

The paper presents a methodological pipeline for generating a synthetic Sanskrit NER dataset by seeding entities from DBpedia and prompting an LLM for sentence creation, followed by benchmarking on standard transformer models. No equations, fitted parameters, or predictive claims exist that could reduce to inputs by construction. No self-citations invoke uniqueness theorems, ansatzes, or load-bearing premises. The work contains no derivations that equate outputs to their own definitions or prior author results, making the central contribution independent of circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim depends on two unverified domain assumptions about external resources rather than any fitted parameters or newly invented entities.

axioms (2)
  • domain assumption DBpedia contains sufficient accurate named entities relevant to classical Sanskrit literature
    Invoked as the seeding source for entity extraction
  • domain assumption A 24B-parameter hybrid reasoning LLM can reliably produce grammatically natural and diverse Sanskrit sentences from entity seeds
    Core mechanism for creating the synthetic sentences

pith-pipeline@v0.9.0 · 5423 in / 1361 out tokens · 69127 ms · 2026-05-07T11:19:31.042452+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

9 extracted references · 9 canonical work pages

  1. [1]

    Bhargava and P

    R. Bhargava and P. Sharma. Named entity recognition for Sanskrit using conditional random fields. InProceedings of the International Conference on Natural Language Processing, 2016

    work page 2016

  2. [2]

    IndoWordNet

    Pushpak Bhattacharyya. IndoWordNet. InProceedings of the Seventh International Confer- ence on Language Resources and Evaluation, Valletta, Malta. European Language Resources Association, 2010

    work page 2010

  3. [3]

    DAGA: Data augmentation with a generation approach for low-resource tagging tasks

    Bosheng Ding, Bill Yuchen Lin, Zhou Zhou, Zhefeng Chen, Bown Ren, and Yikang Zheng. DAGA: Data augmentation with a generation approach for low-resource tagging tasks. InPro- ceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 6045–6057. Association for Computational Linguistics, 2020

    work page 2020

  4. [4]

    Design and analysis of a Sanskrit sandhi splitter

    Pawan Goyal and Gerard Huet. Design and analysis of a Sanskrit sandhi splitter. InProceedings of the 26th International Conference on Computational Linguistics, pages 1382–1392, 2016

    work page 2016

  5. [5]

    DCS - the Digital Corpus of Sanskrit

    Oliver Hellwig. DCS - the Digital Corpus of Sanskrit. InLinguistics, Archaeology and the Human Past, Occasional Paper 9, Kyoto, Japan. Research Institute for Humanity and Nature, 2010

    work page 2010

  6. [6]

    Sanskrit segmentation with LSTM networks

    Oliver Hellwig and Sebastian Nehrdich. Sanskrit segmentation with LSTM networks. InProceed- ings of the 12th Language Resources and Evaluation Conference, pages 5324–5332. European Language Resources Association, 2020

    work page 2020

  7. [7]

    Arnav Mhaske, Harshit Kedia, Sumanth Doddapaneni, and Mitesh M. Khapra. Naamapadam: A large-scale named entity recognition dataset for Indian languages. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics, pages 5393–5414. Association for Computational Linguistics, 2023

    work page 2023

  8. [8]

    H. A. Murthy et al. Rule-based named entity recognition for Indian languages. InProceedings of the IJCNLP Workshop on NER for South and South East Asian Languages, 2008

    work page 2008

  9. [9]

    GPT - NER : Named Entity Recognition via Large Language Models

    Shuhe Wang, Xiaofei Sun, and Jiwei Li. GPT-NER: Named entity recognition via large language models.arXiv preprint arXiv:2304.10428, 2023. 8

    work page arXiv 2023