Pith. sign in

REVIEW 3 major objections 5 minor 27 references

Large Language Models for Low-Resource Languages: A Conceptual Framework for an Electronic Explanatory Dictionary of the Tajik Language

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read This paper proposes the first holistic conceptual architecture for an electronic explanatory dictionary of Tajik, integrating morphological analysis, semantic clustering, LLM-based entry generation, and quality assessment in one pipeline.

desk verdict A solid, well-written blueprint that assembles existing Tajik NLP resources for a dictionary no one has built; the main hole is that it assumes an electronic dictionary that Section 1 admits does not exist. read the letter →

arxiv 2608.04186 v2 pith:MHTEFLO7 submitted 2026-08-04 cs.CL

Mullosharaf K. Arabov , S. S. Pirov , B. Sultonov This is my paper · ORCID
classification cs.CL
keywords largelanguagemodelselectronicexplanatorydictionaryTajiklow-resourcelanguagesparameter-efficientfine-tuningsubwordtokenizationmorphologicalanalysissemanticclustering
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 argues that Tajik, while still lacking a comprehensive digital explanatory dictionary, has accumulated enough corpus data and morphological descriptions to support building one with modern large language models. It proposes a multi-stage architecture whose modules perform morphological analysis and lemmatization, semantic clustering, LLM-based entry generation, and quality assessment. The stated intent is to unify classical lexicography, statistical text analysis, and generative capabilities into a single system. The paper is explicitly a first conceptual stage: it does not yet report a running prototype or experimental evaluation of generated entries.

What carries the argument

The central object is a four-module pipeline. The morphological analysis and lemmatization module uses the morpheme database of 81 prefixes, 76,539 roots, and 128,760 postfixes together with word-formation classifications to reduce word forms to lemmas. The semantic clustering module maps lemmas to vectors with pretrained Word2Vec and FastText embeddings from the TajikNLP toolkit and groups them into semantic fields. The LLM-based generation module builds structured prompts from the lemma, grammatical labels, and semantic cluster, and fine-tunes an open-weight model with LoRA or QLoRA. The quality assessment module scores generated entries with BLEU, ROUGE, METEOR, and BERTScore, then routes them through expert validation. The architecture also relies on the Tajik Web Corpus of 168.5 million words and the Tajik National Corpus for context and verification, and it justifies subword tokenization and parameter-efficient fine-tuning by the agglutinative nature and high morphological variability of Tajik.

What would settle it

Attempt to build the proposed prototype against the cited resources and evaluate it on a validation set of 100 to 200 dictionary entries using BERTScore and expert linguists. The central claim would be falsified if the morpheme database cannot be obtained or reproduced, if the reported corpus and toolkit counts differ materially, or if the generated entries consistently fail the expert check despite acceptable automatic metric scores.

Watch

Extended reading notes

Core claim

The paper's central claim is that a complete electronic explanatory dictionary of Tajik is architecturally feasible today, and it specifies how to build one. The load-bearing design choice is to anchor LLM generation on a formal morphological model, with a database of 81 prefixes, 76,539 roots, and 128,760 postfixes, rather than letting the model process Tajik's agglutinative word forms from raw text. Generation is assigned to an open-weight LLM adapted with LoRA or QLoRA, with Mistral 7B plus QLoRA at rank 16 identified as the current benchmark leader at perplexity 5.03 and the Gemma-based Soro models as the alternative; entries are then checked by automatic metrics and expert linguists in an iterative loop. The paper illustrates the intended flow on the word form китобҳоямро, segmented as китоб + -ҳо + -ям + -ро, and shows how a full entry with definition, examples, synonyms, and thematic group would emerge. The authors state plainly that the architecture awaits prototype implementation and experimental comparison.

Load-bearing premise

The architecture assumes that the cited Tajik language resources, the morpheme database of 81 prefixes, 76,539 roots, and 128,760 postfixes, the Tajik Web Corpus, the TajikNLP toolkit, and the PEFT benchmark results, are real, complete, and accessible as described; if any of them are unavailable or not independently reproducible, the data foundation for the dictionary collapses.

Editorial extensions

If this is right

  • If implemented as described, Tajik would gain its first comprehensive electronic explanatory dictionary, built largely from existing corpus and morphological resources.
  • The dictionary would serve as a foundational resource for machine translation, automatic summarization, sentiment analysis, and question-answering systems in Tajik.
  • The PEFT strategy should permit the generation module to be trained with limited annotated word-entry pairs, supplemented by synthetic pairs and transfer from Persian parallels.
  • The choice between the Gemma-based Soro models and Mistral 7B would be settled experimentally on a validation set of 100 to 200 entries, judged by BERTScore and expert evaluation.
  • Automatic metrics alone would not be trusted: expert linguists remain part of the quality loop, so the resulting entries could be vetted lexicographically before publication.

Reading between the lines

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

  • Because the entire data foundation comes from resources attributed to the authors' own prior work, the blueprint stands or falls on whether those resources are publicly available and exactly as described; a reader should treat the architecture as speculative until a prototype runs.
  • If the pipeline works, it would give a reusable template for other low-resource agglutinative languages that have strong morphological descriptions but no digital dictionary.
  • The reported script barrier, in which multilingual LLMs degrade sharply on Tajik Cyrillic, suggests that the morphological analysis and tokenizer components, not model scale, may carry most of the performance in this design.
  • A direct test of the design would be to compare generated entries against a sample of the printed Explanatory Dictionary of the Tajik Language using the same evaluation metrics, to see whether the LLM output is usable or merely resembles lexicography.
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.

Referee Report

3 major / 5 minor

Summary. The paper proposes a conceptual architecture for an electronic explanatory dictionary of Tajik, integrating morphological analysis and lemmatization, semantic clustering, LLM-based dictionary entry generation, and quality assessment. It surveys existing Tajik linguistic and corpus resources, justifies subword tokenization and parameter-efficient fine-tuning, and presents an end-to-end illustrative example of processing the word form китобҳоямро. The stated contribution is a holistic blueprint that can ground the first comprehensive digital lexicographic resource for Tajik; no implementation, prototype, or evaluation of generated entries is reported.

Significance. If the proposed architecture were implemented and validated, it could provide a systematic template for LLM-based lexicography in Tajik and, by analogy, other low-resource agglutinative languages. The paper's strengths are its broad and systematic literature survey, a clear modular decomposition, a sensible choice of subword tokenization and PEFT methods given Tajik's morphology and scarce annotated data, and a candid discussion of limitations. The manuscript is also honest that this is a conceptual first stage rather than a finished system. However, the central claim of grounding in existing resources is only partially supported, and the evaluation design presupposes a machine-readable lexicographic substrate that the paper itself says is missing.

major comments (3)
  1. [Sections 1, 3.3, 3.4] The architecture presupposes machine-readable explanatory dictionaries: Section 3.3 says the training set will use existing explanatory Tajik dictionaries in electronic form, and Section 3.4 says the same dictionaries will supply reference entries for BLEU, ROUGE, METEOR, and BERTScore. Yet Section 1 states that existing Tajik explanatory dictionaries are available in printed form or as limited electronic versions that do not support dynamic updating or integration with automatic text processing. The paper never specifies a digitization, OCR, or expert-revision step to convert these dictionaries into the required electronic format. This is a load-bearing gap: without a concrete plan for obtaining the machine-readable lexicographic substrate, the training and quality-assessment modules cannot be realized as described.
  2. [Section 3.6] The limitations discussion admits that labeled word-dictionary-entry pairs for Tajik are not available in sufficient volume and proposes synthetic generation of training pairs. However, synthetic data cannot anchor the automatic metrics in Section 3.4, which require reference dictionary entries to compute BLEU, ROUGE, METEOR, and BERTScore. The paper does not specify how a gold-standard reference set will be built or expert-validated, or how synthetic pairs will be kept distinct from reference entries in evaluation. Without such a protocol, the quality-assessment module has no ground truth against which generated entries can be measured.
  3. [Sections 2.3 and 3.5] The feasibility argument depends heavily on a set of author-maintained or co-authored resources (Tajik Web Corpus, TajikNLP, the PEFT benchmark in [Arabov 2026c], and the Soro models) that are cited as arXiv preprints or HuggingFace datasets but are not independently verified in this manuscript. The architecture's modules rely on the existence, completeness, and accessibility of these resources; in particular, the morpheme database (81 prefixes, 76,539 roots, 128,760 postfixes) is a quantitative load-bearing input to the morphological analysis module. The paper should provide a minimal verification plan: exact dataset identifiers, access conditions, and a reproducibility audit of at least the morpheme counts and corpus size, or an explicit statement that these numbers are taken from the cited sources without independent verification.
minor comments (5)
  1. [Section 3.3] The generated dictionary entry for китобҳоямро is presented as if it demonstrates the pipeline's output, but it appears to be an illustrative hand-crafted example. This should be stated explicitly, for instance by labeling it 'illustrative output, not produced by the implemented system', to avoid overstating the current level of feasibility.
  2. [Section 3.4] The claim that BERTScore 'better correlates with expert evaluation' for lexicographic tasks is made without a citation or a planned experiment. Either add a supporting reference or frame this as a hypothesis to be tested during the validation stage.
  3. [Section 2.4] The reported POS-tagging benchmark result (weighted F1 = 0.62) is modest, yet the proposed architecture's morphological module feeds directly into entry generation. The paper should discuss the implications of this accuracy level for the downstream dictionary-entry quality and, if relevant, how morphological disambiguation errors will be handled.
  4. [References and data citations] Many citations are to 2026 preprints and HuggingFace datasets. Before publication, please verify that all cited arXiv identifiers and dataset URLs resolve and that the associated numbers (corpus size, number of aligned sentences, etc.) match the cited versions, since the paper's feasibility claims rely on these resources.
  5. [Section 3.2] The semantic clustering module proposes K-means or hierarchical clustering without specifying the number of clusters or a validation criterion. For a conceptual framework this is acceptable, but adding a sentence on how cluster granularity will be chosen (for example, by intrinsic clustering metrics or expert review) would improve reproducibility.

Circularity Check

1 steps flagged · score 2.0 of 10

Mostly non-circular architecture proposal; one methodological circularity in the planned training/evaluation design.

  1. fitted input called prediction [Section 3.3 (LLM-Based Dictionary Entry Generation Module) and Section 3.4 (Quality Assessment Module)]
    "For the training set, it is planned to use existing explanatory dictionaries of Tajik [Shukurov et al., 2008] in electronic form, parallel corpora and lexical resources [Arabov, 2026d,g], as well as synthetically generated 'word – dictionary entry' pairs based on existing lexicographic data. ... As reference entries, existing explanatory dictionaries of Tajik [Shukurov et al., 2008] are used, as well as, where available, entries from Persian lexicographic sources adapted to the Tajik context."

    The quality-assessment module compares generated entries against reference entries taken from the same 2008 explanatory dictionary that Section 3.3 designates as the training source for the LLM generator. No held-out split is specified. Under these conditions, BLEU, ROUGE, METEOR, and BERTScore measure the model's ability to reproduce training material, not independent lexicographic correctness; Section 3.6 even concedes that labeled word-dictionary-entry pairs 'are not available in sufficient volume,' so a clean train/reference separation is not guaranteed. The evaluation is therefore circular by construction rather than an external check.

full rationale

The paper is a conceptual architecture, not a derived quantitative result, and its central proposal—integrating morphological analysis, semantic clustering, LLM generation, and quality assessment—does not reduce to its inputs. The heavy reliance on author-cited resources (Tajik Web Corpus, TajikNLP, PEFT benchmark) is not circular because those resources are openly available and independently checkable. The one genuine circularity is the planned use of the same 2008 dictionary for both training and evaluation, which makes the proposed quality metrics measure fit to training data rather than external lexicographic quality. Separately, Section 1 describes the existing dictionary as available in printed form or limited electronic versions, while Sections 3.3 and 3.4 assume it is usable in electronic form; this is an unresolved feasibility gap, not a circular derivation. Section 3.6 acknowledges the scarcity of labeled pairs, reinforcing that the evaluation design needs an explicit held-out reference set to avoid circularity.

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

The central claim rests on the availability and correctness of prior Tajik language resources, most of which are self-cited. No free parameters are fitted because the paper stops at design. No new entities are postulated. The axioms listed are the load-bearing premises about the substrate.

assumptions (5)
  • domain assumption The Tajik morpheme database contains 81 prefixes, 76,539 roots, and 128,760 postfixes and is complete enough for lemmatization.
    Section 3.1 relies on this database from Dovudov (2018) and Usmanov & Dovudov (2010, 2015) to perform morphological analysis and lemmatization. If the database is incomplete or unavailable, the module's core function fails.
  • domain assumption The Tajik Web Corpus (1.1B characters), NKTJ (58.4M words), Tajik-Persian parallel corpus, and TajPersLexon exist and are accessible as described.
    Section 3.5 lists these as the data foundation for fine-tuning, context extraction, and evaluation. Much of this evidence comes from self-citations (Arabov 2026a, 2026d, 2026g) and is not independently verified in this paper.
  • domain assumption Subword tokenization with BPE or Unigram can effectively handle Tajik agglutinative morphology.
    Section 3.6 states this based on Arabov & Khaibullina (2026), a self-cited study. The architecture's tokenization choice depends on this result.
  • domain assumption Multilingual LLMs degrade catastrophically on Tajik Cyrillic script, motivating specialized models or fine-tuning.
    Section 3.6 relies on Arnob & Mahi (2026), an external study, to justify the need for Tajik-specialized models.
  • domain assumption Standard lexicographic metrics (BLEU, ROUGE, METEOR, BERTScore) are valid for evaluating generated Tajik dictionary entries.
    Section 3.4 proposes these metrics without empirical calibration for Tajik; BERTScore is asserted to correlate better with expert evaluation, but no local evidence is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Large Language Models for Low-Resource Languages: A Conceptual Framework for an Electronic Explanatory Dictionary of the Tajik Language." pith.science (2026). https://pith.science/paper/MHTEFLO7

@misc{pith2026260804186,
  author       = {Pith},
  title        = {Pith review of: Large Language Models for Low-Resource Languages: A Conceptual Framework for an Electronic Explanatory Dictionary of the Tajik Language},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MHTEFLO7}},
  note         = {Machine review of arXiv:2608.04186}
}
read the original abstract

This paper presents a conceptual framework for developing an electronic explanatory dictionary of the Tajik language using large language models (LLMs). The relevance of the work stems from the absence of a comprehensive digital lexicographic resource for Tajik that is comparable in functionality to dictionaries for high-resource languages, and from the limited adaptation of modern natural language processing technologies to low-resource language systems. Based on a systematic survey of existing linguistic, statistical, and corpus resources, we propose a dictionary architecture that integrates modules for morphological analysis, lemmatization, semantic clustering, and dictionary entry generation using LLMs. The choice of subword tokenization is justified by the agglutinative nature of Tajik morphology and its high morphological variability, along with a parameter-efficient fine-tuning (PEFT) strategy suitable for limited annotated data. The novelty of the work lies in proposing the first holistic conceptual architecture of an explanatory dictionary for Tajik that unifies classical lexicographic methods, language statistics, and generative capabilities of LLMs into a single system. The practical significance of the study is the formation of a methodological foundation for developing a full-featured electronic dictionary that can serve both as a lexicographic tool and as a core resource for machine translation, automatic summarization, sentiment analysis, and other applied NLP tasks. The paper is intended for specialists in computational linguistics, lexicography, and developers of natural language processing systems working with low-resource languages.

Figures

Figures reproduced from arXiv: 2608.04186 by the authors.

Figure 1
Figure 1. Generalized architecture of the LLM-based electronic explanatory dictionary of Tajik [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Sequence diagram of operations for dictionary entry generation [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Cycle of evaluation and quality improvement of dictionary entries [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 22 canonical work pages

  1. [2]

    Benchmarking Parameter-Efficient Fine-Tuning of Large Language Models for Low-Resource Tajik Text Generation with the Tajik Web Corpus

    doi:10.26907/1562- 5419-2026-29-2-546-564. EDN: KPZEZE. M. K. Arabov. Benchmarking parameter-efficient fine-tuning of large language models for low-resource tajik text generation with the tajik web corpus, 2026c. URLhttps://arxiv.org/abs/2605.03742. arXiv:2605.03742. Accessed: 2026-08-01. S.Liashkov,H.SaezdeOcarizBorde,A.Azimi,K.Shoymardonov,S.Khalilbekov...

    work page Pith review arXiv 2026

  2. [3]

    Soro: A Lightweight Foundation Model and Chatbot for Tajik

    URLhttps://arxiv.org/abs/2605.27379. arXiv:2605.27379. Accessed: 2026-08-01. Z. D. Usmanov and G. M. Dovudov.Morphological Analysis of Tajik Word Forms. Donish, Dushanbe,

    work page Pith review arXiv 2026

  3. [6]

    EDN: WZYMGP. M. K. Arabov and TajikNLPWorld. Tajik unified morphological corpus, 2026a. URLhttps://huggingface.co/ datasets/TajikNLPWorld/TajikUnifiedCorpus. Hugging Face. Accessed: 2026-08-01. M. K. Arabov and TajikNLPWorld. Tajik pos corpus,

    work page 2026

  4. [7]

    Hugging Face

    URL https://huggingface.co/datasets/ TajikNLPWorld/tajik-pos-corpus. Hugging Face. Accessed: 2026-08-01. M. K. Arabov and TajikNLPWorld. Tajpersparallelcorpusfull, 2026b. URLhttps://huggingface.co/datasets/ TajikNLPWorld/TajPersParallelCorpusFull. Hugging Face. Accessed: 2026-08-01. M. K. Arabov and TajikNLPWorld. Tajiknamesdataset, 2026c. URL https://hug...

    arXiv 2026

  5. [8]

    doi:10.18653/v1/2026.abjadnlp-1.10. M. K. Arabov. Tajperslexon: A tajik–persian lexical resource and hybrid model for cross-script low-resource nlp. In ProceedingsoftheFirstWorkshoponNLPandLLMsfortheIranianLanguageFamily,pages29–37,Rabat,Morocco, 2026g. doi:10.18653/v1/2026.silkroadnlp-1.4. N. M. Arnob and A. B. Mahi. One language, three of its voices: Ev...

    doi:10.18653/v1/2026.abjadnlp-1.10 2026

  6. [9]

    Natural Language Processing: A Comprehensive Practical Guide from Tokenisation to RLHF

    doi:10.18653/v1/2026.silkroadnlp-1.10. 14 LLMs for Tajik DictionaryA Preprint M. K. Arabov. Natural language processing: A comprehensive practical guide from tokenisation to rlhf, 2026h. URL https://arxiv.org/abs/2605.03799. arXiv:2605.03799. 136 pages. Accessed: 2026-08-01. M. K. Arabov and V. V. Sedykh. Comparative analysis of methods for modeling seman...

    work page Pith review arXiv 2026

  7. [10]

    EDN: ZHBKFG. D. Bear and P. Cook. Cross-lingual wolastoqey-english definition modelling. InProceedings of the International Conference on Recent Advances in Natural Language Processing (RANLP 2021), pages 138–146. INCOMA Ltd.,

    work page 2021

  8. [11]

    Jakubicek and M

    M. Jakubicek and M. Rundell. The end of lexicography? can chatgpt outperform current tools for post-editing lexicography. InElectronic Lexicography in the 21st Century (eLex 2023): Invisible Lexicography, pages 518–533,

    work page 2023

Show all 27 references
  1. [13]

    arXiv:2305.06530

    URL https://arxiv.org/abs/2305.06530. arXiv:2305.06530. Accessed: 2026-08-01. J.Janetzki,G.DeMelo,J.Nemecek,andD.Whitenack. Guide: Creatingsemanticdomaindictionariesforlow-resource languages. InProceedings of the 6th Workshop on Research in Computational Linguistic Typology an...

    2026 arXiv

  2. [15]

    doi:10.18653/v1/2024.acl-demos.28. I. Adebara, A. Elmadany, and M. Abdul-Mageed. Cheetah: Natural language generation for 517 african languages. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12798–12823. ACL,

    2024 doi

  3. [16]

    H.Yu,T.Xu,M.A.Hedderich,etal

    doi:10.18653/v1/2024.acl-long.691. H.Yu,T.Xu,M.A.Hedderich,etal. Afriquellm: Howdatamixingandmodelarchitectureimpactcontinuedpre-training for african languages. InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p...

    2024 doi

  4. [17]

    doi:10.18653/v1/2026.acl-long.267. A. Maheshwari, A. K. Singh, N. J. Karthika, et al. Lexgen: Domain-aware multilingual lexicon generation. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7364–7375. ACL,

    2026 doi

  5. [18]

    doi:10.18653/v1/2025.acl-long.365. M. Jumashev, A. Kasieva, G. Dzhumalieva, et al. Structured data from dictionary text: Applying llms for low-resource cross-lingualinformationextraction. InAnalysisofImages,SocialNetworksandTexts: 12thInternationalConference, AIST 2024, Revise...

    2025 doi

  6. [19]

    doi:10.1007/978-3-031-97019-1_2. R. Merx, E. Vylomova, and K. Kurniawan. Generating bilingual example sentences with large language models as lexicography assistants. InProceedings of the 22nd Annual Workshop of the Australasian Language Technology Association, pages 64–74. ACL,

    doi

  7. [20]

    Stockle, D

    P. Stockle, D. Elsner, W. Koppensteiner, and K. Korecky-Kroll. Llm-assisted dialect lexicography: Challenges and opportunities in processing historical bavarian dialects. InElectronic Lexicography in the 21st Century (eLex 2025): Intelligent Lexicography, pages 457–479. Lexica...

    2025

  8. [21]

    Apipelineforautomateddictionarycreationwithoptionalhumanintervention

    T.Widmann. Apipelineforautomateddictionarycreationwithoptionalhumanintervention. InElectronicLexicography in the 21st Century (eLex 2025): Intelligent Lexicography, pages 33–43. Lexical Computing CZ s.r.o.,

    2025

  9. [22]

    M. Rabe, M. J. Puttkammer, and G. B. van Huyssteen. Compiling a candidate list of taboo constructions for an under-resourced language. InElectronic Lexicography in the 21st Century (eLex 2025): Intelligent Lexicography, pages 749–765. Lexical Computing CZ s.r.o.,

    2025

  10. [23]

    L. Lugli. The mangalam dictionary of buddhist sanskrit: Automating lexicographic data with generative llms. In Electronic Lexicography in the 21st Century (eLex 2025): Intelligent Lexicography, pages 766–782. Lexical Computing CZ s.r.o.,

    2025

  11. [24]

    Marrese-Taylor, E

    E. Marrese-Taylor, E. K. Shimomoto, A. Solano, and E. Reid. Multilingual definition modeling. InFindings of the Asso- ciation for Computational Linguistics: ACL 2025, pages 25888–25906. ACL,

    2025

  12. [25]

    doi:10.18653/v1/2025.findings- acl.1328. H. Lu, Z. L. Li, and W. Lam. Dictionary insertion prompting for multilingual reasoning on multilingual large language models. InProceedings of the 15th Joint Conference on Lexical and Computational Semantics (*SEM 2026), pages 529–544. ACL,

    2025 doi

  13. [26]

    15 LLMs for Tajik DictionaryA Preprint K

    doi:10.18653/v1/2026.starsem-conference.36. 15 LLMs for Tajik DictionaryA Preprint K. Kassab, N. Teslya, and E. Vozhik. Automated dataset-creation and evaluation pipeline for ner in russian literary heritage.Applied Sciences, 15(4):2072,

    2026 doi

  14. [27]

    doi:10.3390/app15042072. 16

    doi

  15. [2021]

    EDN: XYSGOO

    doi:10.17212/2782-2001-2021-2-83-94. EDN: XYSGOO. Kh. A. Khudoyberdiev. On the problems of literary translation and its relationship with machine translation using the example of the tajik language.Bulletin of the Technological University of Tajikistan, 4(47):169–175,

    2001 doi

  16. [2023]

    doi:10.1093/ijl/ecad025. J. Ojo and K. Ogueji. How good are commercial large language models on african languages?,

    doi

  17. [2024]

    doi:10.18653/v1/2024.sigtyp-1.2. X.-P. Nguyen, W. Zhang, X. Li, et al. Seallms - large language models for southeast asia. InProceedings of the 62nd AnnualMeetingoftheAssociationforComputationalLinguistics(Volume3: SystemDemonstrations),pages294–304. ACL,

    2024 doi

  18. [2025]

    EDN: SKZDMO. M. K. Arabov. Tajik–persian parallel corpus: 328,253 aligned sentences, 2026d. URLhttps://huggingface.co/ datasets/arabov/tajik-persian-parallel. Hugging Face. Accessed: 2026-08-01. M. K. Arabov, Kh. S. Makhmadaliev, and K. Kh. Khabibullozoda. Creating a multiform...

    2026

  19. [2026]

    arXiv:2605.04583

    URL https://arxiv.org/abs/2605.04583. arXiv:2605.04583. Accessed: 2026-08-01. 12 LLMs for Tajik DictionaryA Preprint M. K. Arabov. Tajik web corpus: 319,298 documents, 168.5m words, 1.11b characters, 2026a. URLhttps: //huggingface.co/datasets/arabov/tajik-web-corpus. HuggingFa...

    2026 arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.