A Transformer-Based Language Model for Nyishi, a Low-Resource Language of Northeast India

A Transformer-Based Language Model for Nyishi, a Low-Resource Language of Northeast India | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article A Transformer-Based Language Model for Nyishi, a Low-Resource Language of Northeast India Badal Nyalang This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8544228/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract NyishiBERT is a foundational transformer-based language model specifically developed for Nyishi (njz-Latn), a Sino-Tibetan language of Northeast India. Utilizing the ModernBERT-Base architecture, the model was trained on a severely low-resource corpus of 55,870 sentences sourced from the WMT25 shared task, achieving a test perplexity of 20.78. Downstream performance was evaluated using a sentiment classification task constructed via label projection and high-confidence filtering. Results demonstrate that the learned representations effectively support classification even under weak supervision. The model, training code, and evaluation datasets are publicly released to provide a foundational baseline for future research in the Tani language subgroup. Artificial Intelligence and Machine Learning Nyishi low-resource languages Nyishi language modeling transformer models Arunachal Pradesh masked language modeling Northeast India Introduction Northeast India is home to over 220 indigenous languages representing multiple language families including Sino-Tibetan, Austroasiatic, Indo-Aryan, and Tai-Kadai. Despite this linguistic richness, these languages remain critically underserved by modern natural language processing technologies. Most lack basic computational resources such as corpora, lexicons, or language models. Nyishi (ISO 639-3: njz) is a populous ethnic group in Arunachal Pradesh, primarily inhabiting the central districts of the state and parts of Assam (Census of India, 2011 ). Linguistically, it belongs to the Western Tani branch of the Sino-Tibetan family (Post, 2015 ). Despite having approximately 300,000 speakers, the language is classified as 'Vulnerable' by UNESCO due to an increasing shift toward dominant lingua francas (UNESCO, 2017 ). Computational resources remain scarce, as the language lacks the large-scale digital presence required for standard LLM pre-training. While Nyishi has more data than ultra-low-resource settings (< 10k sentences), it remains orders of magnitude smaller than the corpora used to train most modern language models In this work, we present NyishiBERT, a monolingual masked language model for Nyishi trained on data from the Tenth Conference on Machine Translation (WMT25). Our contributions are: The first publicly available foundational language model for Nyishi Evaluation on both intrinsic (perplexity) and extrinsic (sentiment classification) metrics Open release of model, code, and evaluation data Demonstration that effective language models can be trained with limited resources (55K sentences, 3.5 GPU-hours) Related Work Multilingual Language Models Multilingual BERT and XLM-R have enabled cross-lingual transfer for many languages. However, these models face capacity constraints when scaling to hundreds of languages, with low-resource languages often receiving insufficient training exposure. For Indian languages, IndicBERT covers 12 major languages but excludes all Northeast Indian languages. Our work demonstrates that for low-resource languages, dedicated monolingual models remain viable and effective. Low-Resource Language Technology Recent work has explored various strategies for low-resource NLP including transfer learning, data augmentation, and cross-lingual pretraining. However, most studies focus on languages with at least hundreds of thousands of sentences. Data Training Corpus Our Nyishi corpus consists of 55,870 unique sentences (approximately 680,000 tokens) sourced from the WMT EMNLP 2025 shared task data. All text uses Roman orthography (njz-Latn), which has emerged as the primary script for digital Nyishi communication. We split the data as follows: Training : 44,696 sentences (80%) Validation : 5,587 sentences (10%) Test : 5,587 sentences (10%) The corpus represents mixed-domain text reflecting contemporary digital Nyishi usage. We performed minimal preprocessing: Unicode normalization (NFC) and whitespace normalization. Sentiment Dataset For downstream evaluation, we construct a weakly supervised Nyishi sentiment dataset via label projection from English-Nyishi parallel sentences. Sentiment labels are inferred on the English side using a pretrained English sentiment classifier and transferred to the corresponding Nyishi sentences. To reduce noise, we retain only examples with high prediction confidence (≥0.99), yielding 1,000 Nyishi sentences: Positive : 515 sentences Negative : 485 sentences The dataset was split 80% training (800 sentences) and 20% validation (200 sentences). Sentences cover diverse topics including social interactions, opinions, and personal experiences. Tokenization Rather than training a corpus-specific tokenizer, we leverage the existing tokenizer from NE-BERT, a 50,368-token SentencePiece Unigram model trained on 16 Northeast Indian languages. This choice provides better coverage for Nyishi morphology than a small corpus-specific tokenizer while enabling cross-model compatibility. Model Architecture and Training Architecture NyishiBERT uses the ModernBERT-Base architecture, a recent BERT variant with improved efficiency: Parameters : 149 million Layers : 22 transformer layers Hidden size : 768 Attention heads : 12 Context window : 1024 tokens Positional embeddings : Rotary Position Embeddings (RoPE) Normalization : Pre-LayerNorm All model weights are randomly initialized; we do not use any pretrained checkpoints. Training Configuration We train using masked language modeling (MLM) with 15% masking probability. Training hyperparameters: Optimizer : AdamW ( , , ) Learning rate : with linear warmup (10% of steps) Weight decay : 0.01 Batch size : 16 (effective, with gradient accumulation) Gradient clipping : 1.0 Training epochs : 10 Total steps : 27,940 Precision : bfloat16 Training was conducted on a single NVIDIA A40 GPU (48GB) and completed in approximately 1.7 hours, demonstrating the feasibility of developing language models for low-resource languages with minimal computational resources. Experiments Intrinsic Evaluation We evaluate NyishiBERT on the held-out test set using standard language modeling metrics: Table 1. Intrinsic evaluation results on Nyishi test set (5,587 sentences). Metric Value Test Loss 3.03 Perplexity 20.78 The perplexity of 20.78 indicates that NyishiBERT has learned meaningful representations of Nyishi language despite the limited training data. Baseline Comparison Table 2. Comparative Intrinsic Evaluation on Nyishi Test Set. Model Pseudo-Perplexity (PPPL) ↓ mBERT (base-multilingual-cased) 6.25 X 10 8 XLM-RoBERTa (base) 6.29 X10 8 NyishiBERT (Ours) 97.34 To contextualize NyishiBERT’s performance, we conduct a comparative analysis against two widely used multilingual baselines: mBERT and XLM-RoBERTa (base). We calculate word-normalized Pseudo-Perplexity (PPPL) on the same held-out test set to ensure a fair comparison across differing tokenizers. As shown in Table 2, both mBERT and XLM-R exhibit catastrophic failure in representing Nyishi, with PPPL values exceeding 10 8 . This failure is largely attributed to 'vocabulary dilution' and the models' lack of exposure to the unique morphological and syntactic structures of the Tani language group during their massive multilingual pre-training. Conversely, NyishiBERT achieves a functional PPPL of 97.34. This million-fold improvement empirically validates the necessity of dedicated monolingual resources for severely low-resource languages like Nyishi. Sentiment Classification To evaluate NyishiBERT’s utility for downstream tasks, we fine-tune the model for binary sentiment classification on our weakly supervised annotated dataset. Fine-tuning Configuration : Learning rate: Batch size: 16 Epochs: 5 Classification head: Single linear layer on [CLS] token Results : Table 3. Sentiment classification results on Nyishi validation set (200 sentences). Random baseline would achieve 50% accuracy and 50% F1. Model Accuracy F1 (Macro) NyishiBERT 60.5% 37.7% While the sentiment results are modest, they demonstrate that NyishiBERT captures linguistic patterns useful for downstream classification. The moderate F1 score (37.7%) likely reflects the limited size of the sentiment training set (800 examples) rather than fundamental model limitations. This represents a baseline for future work on Nyishi sentiment analysis. Discussion Model Performance NyishiBERT achieves reasonable perplexity given the extremely limited training data (55K sentences). For context, typical BERT models are trained on billions of tokens; NyishiBERT uses less than 1 million tokens yet learns meaningful language representations. The sentiment classification results, while modest, exceed random baseline performance and demonstrate that the learned representations capture linguistic patterns useful for downstream classification under weak supervision. With more labeled data, we expect significant performance improvements. We note that this work focuses on monolingual model construction; controlled analyses of data upsampling and sampling strategies are explored in separate work. Resource Efficiency A key finding is the computational efficiency of NyishiBERT training: Training time : 1.7 GPU-hours Hardware : Single A40 GPU Cost : Approximately $15 on commercial cloud platforms This demonstrates that high-quality language models for low-resource languages can be developed with minimal computational resources, making community-driven language technology development economically feasible. Limitations We acknowledge several limitations: Orthographic variation : Nyishi lacks standardized orthography. Our model reflects spelling conventions in the WMT25 corpus, which may vary from other writing practices. Domain coverage : Training data represents contemporary digital usage. Performance on specialized domains (legal, medical, traditional literature) may vary. Evaluation scope : Our downstream evaluation uses only sentiment classification. Development of additional task-specific datasets (NER, POS tagging, parsing) would enable more comprehensive evaluation. Script limitation : The model is trained exclusively on Roman script and will not work with other writing systems. Broader Impact NyishiBERT represents the first step toward comprehensive NLP support for Nyishi. Potential applications include: Educational technology for Nyishi language learning Machine translation between Nyishi and other languages Digital content creation and curation Language documentation and preservation More broadly, this work provides a template for developing language models for the 200+ other indigenous languages of Northeast India, many of which have similar or smaller digital corpora. Release and Reproducibility We publicly release: Model : Available on HuggingFace at MWireLabs/nyishibert Training code : Complete training scripts with hyperparameters Evaluation data : Sentiment dataset and evaluation scripts Documentation : Detailed model card and usage examples All resources are released under CC-BY-4.0 license to maximize accessibility and enable future research. Conclusion We present NyishiBERT, the first transformer-based language model for Nyishi, trained on 55,870 sentences from WMT25. The model achieves test perplexity of 20.78 and demonstrates utility for downstream sentiment classification despite limited training data. Our work demonstrates that effective language models for low-resource languages can be developed with minimal computational resources (3.5 GPU-hours, < $ 15). We hope this work inspires similar efforts for the hundreds of other indigenous languages of Northeast India and serves as a foundation for future Nyishi NLP research. Declarations Acknowledgments We thank the WMT25 organizers for making Nyishi data available. We acknowledge the Nyishi language community for their linguistic heritage that makes this work possible. References Census of India. (2011). Arunachal Pradesh: Primary Census Abstract. Office of the Registrar General & Census Commissioner. Post, M. W. (2015). The Tani Languages. In P. N. Thurgood & R. J. LaPolla (Eds.), The Sino-Tibetan Languages. Routledge. UNESCO. (2017). Atlas of the World’s Languages in Danger. UNESCO Publishing. Devlin, J., Chang, M. W., Lee, K., & Toutanova, K. (2019). BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT (pp. 4171-4186). Conneau, A., Khandelwal, K., Goyal, N., Chaudhary, V., Wenzek, G., Guzmán, F., ... & Stoyanov, V. (2020). Unsupervised cross-lingual representation learning at scale. In Proceedings of ACL (pp. 8440-8451). Kakwani, D., Kunchukuttan, A., Golla, S., Gokul, N. C., Bhattacharyya, A., Khapra, M. M., & Kumar, P. (2020). IndicNLPSuite: Monolingual corpora, evaluation benchmarks and pre-trained multilingual language models for Indian languages. In Findings of EMNLP (pp. 4948-4961). Portes, J., Gross, A., Havens, J., Patel, D., Kaplan, J., & Arora, S. (2024). ModernBERT: A modern bidirectional encoder trained with Flash Attention. arXiv preprint arXiv:2412.13663 . Zoph, B., Yuret, D., May, J., & Knight, K. (2016). Transfer learning for low-resource neural machine translation. In Proceedings of EMNLP (pp. 1568-1575). Fadaee, M., Bisazza, A., & Monz, C. (2017). Data augmentation for low-resource neural machine translation. In Proceedings of ACL (pp. 567-573). Artetxe, M., Ruder, S., & Yogatama, D. (2020). On the cross-lingual transferability of monolingual representations. In Proceedings of ACL (pp. 4623-4637). Singh, T. D., Ekbal, A., & Bandyopadhyay, S. (2018). Building computational infrastructure for Manipuri: ILCI corpus. In Proceedings of LREC . Pathak, A., Pakray, P., Bentham, J., & Jha, R. R. (2018). Neural machine translation for Indian languages. Journal of Intelligent Systems , 28(3), 465-477. Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8544228","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":571022968,"identity":"219c2edc-a0ae-4082-93f8-27aa1fb73290","order_by":0,"name":"Badal Nyalang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA7klEQVRIiWNgGAWjYHACxgNAQg5INzAwFDAkEKUHpMUYosWABC2JDWAmMVrMJZIfHPi5ozZ97Yzktg8fDOzy5BuYHz66gUeL5Yw0g4O9Z47nbruR2DxzhkFyscEBNmPjHDxaDG4kGBzgbTuWu+12YjMzjwFz4gYGHjZp/FrSPxz823Ys3Qyk5Y9BfeL8BoJacgwO87bVJIC1MBgcTmw4QEjLmTcFh2XbDhhuu/+wmbHH4HixwWFCfjmevvHh27Y6ebMzxx8z/KiozpNvb374GJ8WBoEEEHkYSYQZn3IQ4D8AIusIKRsFo2AUjIKRDAB3NFbrQMKkvwAAAABJRU5ErkJggg==","orcid":"https://orcid.org/0009-0006-6923-0797","institution":"MWire Labs","correspondingAuthor":true,"prefix":"","firstName":"Badal","middleName":"","lastName":"Nyalang","suffix":""}],"badges":[],"createdAt":"2026-01-07 17:39:06","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-8544228/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8544228/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":99863252,"identity":"4a9ed096-29b5-4752-ae5b-4e0b30f2b680","added_by":"auto","created_at":"2026-01-09 07:29:13","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":29690,"visible":true,"origin":"","legend":"","description":"","filename":"nyishibertpreprint.docx","url":"https://assets-eu.researchsquare.com/files/rs-8544228/v1/d722c12d0c04fcf25e48be6a.docx"},{"id":99863251,"identity":"b2ae30cd-ad6a-47ea-a7e7-d092e1c3d3b0","added_by":"auto","created_at":"2026-01-09 07:29:13","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":342,"visible":true,"origin":"","legend":"","description":"","filename":"rs8544228.json","url":"https://assets-eu.researchsquare.com/files/rs-8544228/v1/d458bfd4648a59e51f8e9dcf.json"},{"id":99863254,"identity":"8dc48f3b-a793-402a-b6a6-db146a2d5435","added_by":"auto","created_at":"2026-01-09 07:29:14","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":37029,"visible":true,"origin":"","legend":"","description":"","filename":"rs85442280enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8544228/v1/6988a8df993963ab85386ee6.xml"},{"id":99863255,"identity":"01721bf0-5ae0-4433-84a3-5b32661731b9","added_by":"auto","created_at":"2026-01-09 07:29:15","extension":"xml","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":35042,"visible":true,"origin":"","legend":"","description":"","filename":"rs85442280structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8544228/v1/86b6bac0488c3834efb7530b.xml"},{"id":99863253,"identity":"01daea0c-95a0-4fa3-a151-9e09dc1401e8","added_by":"auto","created_at":"2026-01-09 07:29:14","extension":"html","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":43083,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8544228/v1/50811d07cb50b12c349a940c.html"},{"id":100357287,"identity":"16e1ae50-0db5-4196-bc2a-c02415181b19","added_by":"auto","created_at":"2026-01-16 07:19:36","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":665601,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8544228/v1/b006369a-fa08-4004-808f-a9a3167244fc.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003eA Transformer-Based Language Model for Nyishi, a Low-Resource Language of Northeast India\u003c/p\u003e","fulltext":[{"header":"Introduction","content":"\u003cp\u003eNortheast India is home to over 220 indigenous languages representing multiple language families including Sino-Tibetan, Austroasiatic, Indo-Aryan, and Tai-Kadai. Despite this linguistic richness, these languages remain critically underserved by modern natural language processing technologies. Most lack basic computational resources such as corpora, lexicons, or language models.\u003c/p\u003e \u003cp\u003eNyishi (ISO 639-3: njz) is a populous ethnic group in Arunachal Pradesh, primarily inhabiting the central districts of the state and parts of Assam (Census of India, \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2011\u003c/span\u003e). Linguistically, it belongs to the Western Tani branch of the Sino-Tibetan family (Post, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2015\u003c/span\u003e). Despite having approximately 300,000 speakers, the language is classified as 'Vulnerable' by UNESCO due to an increasing shift toward dominant lingua francas (UNESCO, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). Computational resources remain scarce, as the language lacks the large-scale digital presence required for standard LLM pre-training.\u003c/p\u003e \u003cp\u003eWhile Nyishi has more data than ultra-low-resource settings (\u0026lt;\u0026thinsp;10k sentences), it remains orders of magnitude smaller than the corpora used to train most modern language models\u003c/p\u003e \u003cp\u003eIn this work, we present NyishiBERT, a monolingual masked language model for Nyishi trained on data from the Tenth Conference on Machine Translation (WMT25). Our contributions are:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eThe first publicly available foundational language model for Nyishi\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eEvaluation on both intrinsic (perplexity) and extrinsic (sentiment classification) metrics\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eOpen release of model, code, and evaluation data\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eDemonstration that effective language models can be trained with limited resources (55K sentences, 3.5 GPU-hours)\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e"},{"header":"Related Work","content":"\u003ch2\u003eMultilingual Language Models\u003c/h2\u003e\n\u003cp\u003eMultilingual BERT and XLM-R have enabled cross-lingual transfer for many languages. However, these models face capacity constraints when scaling to hundreds of languages, with low-resource languages often receiving insufficient training exposure.\u003c/p\u003e\n\u003cp\u003eFor Indian languages, IndicBERT covers 12 major languages but excludes all Northeast Indian languages. Our work demonstrates that for low-resource languages, dedicated monolingual models remain viable and effective.\u003c/p\u003e\n\u003ch2\u003eLow-Resource Language Technology\u003c/h2\u003e\n\u003cp\u003eRecent work has explored various strategies for low-resource NLP including transfer learning, data augmentation, and cross-lingual pretraining. However, most studies focus on languages with at least hundreds of thousands of sentences.\u003c/p\u003e"},{"header":"Data","content":"\u003ch2\u003eTraining Corpus\u003c/h2\u003e\n\u003cp\u003eOur Nyishi corpus consists of 55,870 unique sentences (approximately 680,000 tokens) sourced from the WMT EMNLP 2025 shared task data. All text uses Roman orthography (njz-Latn), which has emerged as the primary script for digital Nyishi communication.\u003c/p\u003e\n\u003cp\u003eWe split the data as follows:\u003c/p\u003e\n\u003cul\u003e\n \u003cli\u003e\u003cstrong\u003eTraining\u003c/strong\u003e: 44,696 sentences (80%)\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eValidation\u003c/strong\u003e: 5,587 sentences (10%)\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eTest\u003c/strong\u003e: 5,587 sentences (10%)\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eThe corpus represents mixed-domain text reflecting contemporary digital Nyishi usage. We performed minimal preprocessing: Unicode normalization (NFC) and whitespace normalization.\u003c/p\u003e\n\u003ch2\u003eSentiment Dataset\u003c/h2\u003e\n\u003cp\u003eFor downstream evaluation, we construct a weakly supervised Nyishi sentiment dataset via label projection from English-Nyishi parallel sentences. Sentiment labels are inferred on the English side using a pretrained English sentiment classifier and transferred to the corresponding Nyishi sentences. To reduce noise, we retain only examples with high prediction confidence (\u0026ge;0.99), yielding 1,000 Nyishi sentences:\u003c/p\u003e\n\u003cul\u003e\n \u003cli\u003e\u003cstrong\u003ePositive\u003c/strong\u003e: 515 sentences\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eNegative\u003c/strong\u003e: 485 sentences\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eThe dataset was split 80% training (800 sentences) and 20% validation (200 sentences). Sentences cover diverse topics including social interactions, opinions, and personal experiences.\u003c/p\u003e\n\u003ch2\u003eTokenization\u003c/h2\u003e\n\u003cp\u003eRather than training a corpus-specific tokenizer, we leverage the existing tokenizer from NE-BERT, a 50,368-token SentencePiece Unigram model trained on 16 Northeast Indian languages. This choice provides better coverage for Nyishi morphology than a small corpus-specific tokenizer while enabling cross-model compatibility.\u003c/p\u003e"},{"header":"Model Architecture and Training","content":"\u003ch2\u003eArchitecture\u003c/h2\u003e\n\u003cp\u003eNyishiBERT uses the ModernBERT-Base architecture, a recent BERT variant with improved efficiency:\u003c/p\u003e\n\u003cul\u003e\n \u003cli\u003e\u003cstrong\u003eParameters\u003c/strong\u003e: 149 million\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eLayers\u003c/strong\u003e: 22 transformer layers\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eHidden size\u003c/strong\u003e: 768\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eAttention heads\u003c/strong\u003e: 12\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eContext window\u003c/strong\u003e: 1024 tokens\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003ePositional embeddings\u003c/strong\u003e: Rotary Position Embeddings (RoPE)\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eNormalization\u003c/strong\u003e: Pre-LayerNorm\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eAll model weights are randomly initialized; we do not use any pretrained checkpoints.\u003c/p\u003e\n\u003ch2\u003eTraining Configuration\u003c/h2\u003e\n\u003cp\u003eWe train using masked language modeling (MLM) with 15% masking probability. Training hyperparameters:\u003c/p\u003e\n\u003cul class=\"decimal_type\"\u003e\n \u003cli\u003e\u003cstrong\u003eOptimizer\u003c/strong\u003e: AdamW (\u003cimg width=\"57\" height=\"19\" src=\"https://myfiles.space/user_files/58895_8739fc6c57c1c19a/58895_custom_files/img176794348719.png\" alt=\"image\"\u003e, \u003cimg width=\"75\" height=\"19\" src=\"https://myfiles.space/user_files/58895_8739fc6c57c1c19a/58895_custom_files/img176794348784.png\" alt=\"image\"\u003e, \u003cimg width=\"62\" height=\"19\" src=\"https://myfiles.space/user_files/58895_8739fc6c57c1c19a/58895_custom_files/img1767943487.png\" alt=\"image\"\u003e)\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eLearning rate\u003c/strong\u003e: \u003cimg width=\"61\" height=\"19\" src=\"https://myfiles.space/user_files/58895_8739fc6c57c1c19a/58895_custom_files/img176794348749.png\" alt=\"image\"\u003e\u0026nbsp;with linear warmup (10% of steps)\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eWeight decay\u003c/strong\u003e: 0.01\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eBatch size\u003c/strong\u003e: 16 (effective, with gradient accumulation)\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eGradient clipping\u003c/strong\u003e: 1.0\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eTraining epochs\u003c/strong\u003e: 10\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eTotal steps\u003c/strong\u003e: 27,940\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003ePrecision\u003c/strong\u003e: bfloat16\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eTraining was conducted on a single NVIDIA A40 GPU (48GB) and completed in approximately 1.7 hours, demonstrating the feasibility of developing language models for low-resource languages with minimal computational resources.\u003c/p\u003e"},{"header":"Experiments","content":"\u003ch2\u003eIntrinsic Evaluation\u003c/h2\u003e\n\u003cp\u003eWe evaluate NyishiBERT on the held-out test set using standard language modeling metrics:\u003c/p\u003e\n\u003cp\u003eTable 1. Intrinsic evaluation results on Nyishi test set (5,587 sentences).\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"601\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e\u003cstrong\u003eMetric\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e\u003cstrong\u003eValue\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\u0026nbsp;\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTest Loss\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e3.03\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u0026nbsp;\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePerplexity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e20.78\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u0026nbsp;\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eThe perplexity of 20.78 indicates that NyishiBERT has learned meaningful representations of Nyishi language despite the limited training data.\u003c/p\u003e\n\u003ch2\u003eBaseline Comparison\u003c/h2\u003e\n\u003cp\u003eTable 2. Comparative Intrinsic Evaluation on Nyishi Test Set.\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"601\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e\u003cstrong\u003eModel\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e\u003cstrong\u003ePseudo-Perplexity (PPPL) \u0026darr;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003emBERT (base-multilingual-cased)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e6.25 X 10\u003csup\u003e8\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eXLM-RoBERTa (base)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e6.29 X10\u003csup\u003e8\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eNyishiBERT (Ours)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e97.34\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eTo contextualize NyishiBERT\u0026rsquo;s performance, we conduct a comparative analysis against two widely used multilingual baselines: mBERT and XLM-RoBERTa (base). We calculate word-normalized Pseudo-Perplexity (PPPL) on the same held-out test set to ensure a fair comparison across differing tokenizers.\u003c/p\u003e\n\u003cp\u003eAs shown in Table 2, both mBERT and XLM-R exhibit catastrophic failure in representing Nyishi, with PPPL values exceeding 10\u003csup\u003e8\u003c/sup\u003e. This failure is largely attributed to \u0026apos;vocabulary dilution\u0026apos; and the models\u0026apos; lack of exposure to the unique morphological and syntactic structures of the Tani language group during their massive multilingual pre-training. Conversely, NyishiBERT achieves a functional PPPL of 97.34. This million-fold improvement empirically validates the necessity of dedicated monolingual resources for severely low-resource languages like Nyishi.\u003c/p\u003e\n\u003ch2\u003eSentiment Classification\u003c/h2\u003e\n\u003cp\u003eTo evaluate NyishiBERT\u0026rsquo;s utility for downstream tasks, we fine-tune the model for binary sentiment classification on our weakly supervised annotated dataset.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFine-tuning Configuration\u003c/strong\u003e:\u003c/p\u003e\n\u003cul\u003e\n \u003cli\u003eLearning rate: \u003cimg width=\"61\" height=\"19\" src=\"https://myfiles.space/user_files/58895_8739fc6c57c1c19a/58895_custom_files/img1767943536.png\" alt=\"image\"\u003e\u003c/li\u003e\n \u003cli\u003eBatch size: 16\u003c/li\u003e\n \u003cli\u003eEpochs: 5\u003c/li\u003e\n \u003cli\u003eClassification head: Single linear layer on [CLS] token\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003e\u003cstrong\u003eResults\u003c/strong\u003e:\u003c/p\u003e\n\u003cp\u003eTable 3. Sentiment classification results on Nyishi validation set (200 sentences). Random baseline would achieve 50% accuracy and 50% F1.\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"624\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e\u003cstrong\u003eModel\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e\u003cstrong\u003eAccuracy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e\u003cstrong\u003eF1 (Macro)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eNyishiBERT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e60.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e37.7%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eWhile the sentiment results are modest, they demonstrate that NyishiBERT captures linguistic patterns useful for downstream classification. The moderate F1 score (37.7%) likely reflects the limited size of the sentiment training set (800 examples) rather than fundamental model limitations. This represents a baseline for future work on Nyishi sentiment analysis.\u003c/p\u003e"},{"header":"Discussion","content":"\u003ch2\u003eModel Performance\u003c/h2\u003e\n\u003cp\u003eNyishiBERT achieves reasonable perplexity given the extremely limited training data (55K sentences). For context, typical BERT models are trained on billions of tokens; NyishiBERT uses less than 1 million tokens yet learns meaningful language representations.\u003c/p\u003e\n\u003cp\u003eThe sentiment classification results, while modest, exceed random baseline performance and demonstrate that the learned representations capture linguistic patterns useful for downstream classification under weak supervision. With more labeled data, we expect significant performance improvements. We note that this work focuses on monolingual model construction; controlled analyses of data upsampling and sampling strategies are explored in separate work.\u003c/p\u003e\n\u003ch2\u003eResource Efficiency\u003c/h2\u003e\n\u003cp\u003eA key finding is the computational efficiency of NyishiBERT training:\u003c/p\u003e\n\u003cul\u003e\n \u003cli\u003e\u003cstrong\u003eTraining time\u003c/strong\u003e: 1.7 GPU-hours\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eHardware\u003c/strong\u003e: Single A40 GPU\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eCost\u003c/strong\u003e: Approximately $15 on commercial cloud platforms\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eThis demonstrates that high-quality language models for low-resource languages can be developed with minimal computational resources, making community-driven language technology development economically feasible.\u003c/p\u003e\n\u003ch2\u003eLimitations\u003c/h2\u003e\n\u003cp\u003eWe acknowledge several limitations:\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eOrthographic variation\u003c/strong\u003e: Nyishi lacks standardized orthography. Our model reflects spelling conventions in the WMT25 corpus, which may vary from other writing practices.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDomain coverage\u003c/strong\u003e: Training data represents contemporary digital usage. Performance on specialized domains (legal, medical, traditional literature) may vary.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEvaluation scope\u003c/strong\u003e: Our downstream evaluation uses only sentiment classification. Development of additional task-specific datasets (NER, POS tagging, parsing) would enable more comprehensive evaluation.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eScript limitation\u003c/strong\u003e: The model is trained exclusively on Roman script and will not work with other writing systems.\u003c/p\u003e\n\u003ch2\u003eBroader Impact\u003c/h2\u003e\n\u003cp\u003eNyishiBERT represents the first step toward comprehensive NLP support for Nyishi. Potential applications include:\u003c/p\u003e\n\u003cul\u003e\n \u003cli\u003eEducational technology for Nyishi language learning\u003c/li\u003e\n \u003cli\u003eMachine translation between Nyishi and other languages\u003c/li\u003e\n \u003cli\u003eDigital content creation and curation\u003c/li\u003e\n \u003cli\u003eLanguage documentation and preservation\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eMore broadly, this work provides a template for developing language models for the 200+ other indigenous languages of Northeast India, many of which have similar or smaller digital corpora.\u003c/p\u003e\n\u003ch2\u003eRelease and Reproducibility\u003c/h2\u003e\n\u003cp\u003eWe publicly release:\u003c/p\u003e\n\u003cul class=\"decimal_type\"\u003e\n \u003cli\u003e\u003cstrong\u003eModel\u003c/strong\u003e: Available on HuggingFace at MWireLabs/nyishibert\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eTraining code\u003c/strong\u003e: Complete training scripts with hyperparameters\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eEvaluation data\u003c/strong\u003e: Sentiment dataset and evaluation scripts\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eDocumentation\u003c/strong\u003e: Detailed model card and usage examples\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eAll resources are released under CC-BY-4.0 license to maximize accessibility and enable future research.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eWe present NyishiBERT, the first transformer-based language model for Nyishi, trained on 55,870 sentences from WMT25. The model achieves test perplexity of 20.78 and demonstrates utility for downstream sentiment classification despite limited training data.\u003c/p\u003e \u003cp\u003eOur work demonstrates that effective language models for low-resource languages can be developed with minimal computational resources (3.5 GPU-hours, \u0026lt;\u003cspan\u003e$\u003c/span\u003e15). We hope this work inspires similar efforts for the hundreds of other indigenous languages of Northeast India and serves as a foundation for future Nyishi NLP research.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAcknowledgments\u003c/h2\u003e \u003cp\u003eWe thank the WMT25 organizers for making Nyishi data available. We acknowledge the Nyishi language community for their linguistic heritage that makes this work possible.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eCensus of India. (2011). Arunachal Pradesh: Primary Census Abstract. Office of the Registrar General \u0026amp; Census Commissioner.\u003c/li\u003e\n \u003cli\u003ePost, M. W. (2015). The Tani Languages. In P. N. Thurgood \u0026amp; R. J. LaPolla (Eds.), The Sino-Tibetan Languages. Routledge.\u003c/li\u003e\n \u003cli\u003eUNESCO. (2017). Atlas of the World\u0026rsquo;s Languages in Danger. UNESCO Publishing.\u003c/li\u003e\n \u003cli\u003eDevlin, J., Chang, M. W., Lee, K., \u0026amp; Toutanova, K. (2019). BERT: Pre-training of deep bidirectional transformers for language understanding. In \u003cem\u003eProceedings of NAACL-HLT\u003c/em\u003e (pp. 4171-4186).\u003c/li\u003e\n \u003cli\u003eConneau, A., Khandelwal, K., Goyal, N., Chaudhary, V., Wenzek, G., Guzm\u0026aacute;n, F., ... \u0026amp; Stoyanov, V. (2020). Unsupervised cross-lingual representation learning at scale. In \u003cem\u003eProceedings of ACL\u003c/em\u003e (pp. 8440-8451).\u003c/li\u003e\n \u003cli\u003eKakwani, D., Kunchukuttan, A., Golla, S., Gokul, N. C., Bhattacharyya, A., Khapra, M. M., \u0026amp; Kumar, P. (2020). IndicNLPSuite: Monolingual corpora, evaluation benchmarks and pre-trained multilingual language models for Indian languages. In \u003cem\u003eFindings of EMNLP\u003c/em\u003e (pp. 4948-4961).\u003c/li\u003e\n \u003cli\u003ePortes, J., Gross, A., Havens, J., Patel, D., Kaplan, J., \u0026amp; Arora, S. (2024). ModernBERT: A modern bidirectional encoder trained with Flash Attention. \u003cem\u003earXiv preprint arXiv:2412.13663\u003c/em\u003e.\u003c/li\u003e\n \u003cli\u003eZoph, B., Yuret, D., May, J., \u0026amp; Knight, K. (2016). Transfer learning for low-resource neural machine translation. In \u003cem\u003eProceedings of EMNLP\u003c/em\u003e (pp. 1568-1575).\u003c/li\u003e\n \u003cli\u003eFadaee, M., Bisazza, A., \u0026amp; Monz, C. (2017). Data augmentation for low-resource neural machine translation. In \u003cem\u003eProceedings of ACL\u003c/em\u003e (pp. 567-573).\u003c/li\u003e\n \u003cli\u003eArtetxe, M., Ruder, S., \u0026amp; Yogatama, D. (2020). On the cross-lingual transferability of monolingual representations. In \u003cem\u003eProceedings of ACL\u003c/em\u003e (pp. 4623-4637).\u003c/li\u003e\n \u003cli\u003eSingh, T. D., Ekbal, A., \u0026amp; Bandyopadhyay, S. (2018). Building computational infrastructure for Manipuri: ILCI corpus. In \u003cem\u003eProceedings of LREC\u003c/em\u003e.\u003c/li\u003e\n \u003cli\u003ePathak, A., Pakray, P., Bentham, J., \u0026amp; Jha, R. R. (2018). Neural machine translation for Indian languages. \u003cem\u003eJournal of Intelligent Systems\u003c/em\u003e, 28(3), 465-477.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"MWire Labs, Shillong, Meghalaya","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Nyishi, low-resource languages, Nyishi language modeling, transformer models Arunachal Pradesh, masked language modeling, Northeast India","lastPublishedDoi":"10.21203/rs.3.rs-8544228/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8544228/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eNyishiBERT is a foundational transformer-based language model specifically developed for Nyishi (njz-Latn), a Sino-Tibetan language of Northeast India. Utilizing the ModernBERT-Base architecture, the model was trained on a severely low-resource corpus of 55,870 sentences sourced from the WMT25 shared task, achieving a test perplexity of 20.78. Downstream performance was evaluated using a sentiment classification task constructed via label projection and high-confidence filtering. Results demonstrate that the learned representations effectively support classification even under weak supervision. The model, training code, and evaluation datasets are publicly released to provide a foundational baseline for future research in the Tani language subgroup.\u003c/p\u003e","manuscriptTitle":"A Transformer-Based Language Model for Nyishi, a Low-Resource Language of Northeast India","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-01-09 07:29:07","doi":"10.21203/rs.3.rs-8544228/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"56d2833e-7421-4441-b9f7-177b1b9b702a","owner":[],"postedDate":"January 9th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":60766126,"name":"Artificial Intelligence and Machine Learning"}],"tags":[],"updatedAt":"2026-01-09T07:29:07+00:00","versionOfRecord":[],"versionCreatedAt":"2026-01-09 07:29:07","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8544228","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8544228","identity":"rs-8544228","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}