arXiv:2511.08145v1 Announce Type: new 
Abstract: Large Language Models (LLMs) are increasingly treated as universal, general-purpose solutions across NLP tasks, particularly in English. But does this assumption hold for low-resource, morphologically rich languages such as Sanskrit? We address this question by comparing instruction-tuned and in-context-prompted LLMs with smaller task-specific encoder-decoder models on the Sanskrit poetry-to-prose conversion task. This task is intrinsically challenging: Sanskrit verse exhibits free word order combined with rigid metrical constraints, and its conversion to canonical prose (anvaya) requires multi-step reasoning involving compound segmentation, dependency resolution, and syntactic linearisation. This makes it an ideal testbed to evaluate whether LLMs can surpass specialised models. For LLMs, we apply instruction fine-tuning on general-purpose models and design in-context learning templates grounded in Paninian grammar and classical commentary heuristics. For task-specific modelling, we fully fine-tune a ByT5-Sanskrit Seq2Seq model. Our experiments show that domain-specific fine-tuning of ByT5-Sanskrit significantly outperforms all instruction-driven LLM approaches. Human evaluation strongly corroborates this result, with scores exhibiting high correlation with Kendall's Tau scores. Additionally, our prompting strategies provide an alternative to fine-tuning when domain-specific verse corpora are unavailable, and the task-specific Seq2Seq model demonstrates robust generalisation on out-of-domain evaluations.

تدرس دراسة حديثة مقارنة بين نماذج اللغة الكبيرة (LLMs) والنماذج الأصغر المتخصصة في مهمة تحويل الشعر السنسكريتي إلى نثر. تُظهر الأبحاث أن نماذج LLMs، التي تُعتبر غالبًا حلولًا عالمية، تواجه صعوبات مع اللغات ذات الموارد المحدودة مثل السنسكريتية، حيث تتفوق النماذج المتخصصة مثل ByT5-Sanskrit بشكل كبير. هذه النتائج مهمة لتقدم تقنيات معالجة اللغة الطبيعية في سياقات لغوية متنوعة.

Un estudio reciente compara los grandes modelos de lenguaje (LLMs) con modelos más pequeños y específicos para la tarea de conversión de poesía sánscrita a prosa. La investigación destaca que los LLMs, a menudo considerados soluciones universales, enfrentan dificultades con lenguas de bajos recursos como el sánscrito, donde modelos especializados como ByT5-Sanskrit los superan significativamente. Este hallazgo es crucial para avanzar en las tecnologías de procesamiento del lenguaje natural en contextos lingüísticos diversos.

Une étude récente compare les grands modèles de langage (LLMs) à des modèles plus petits et spécifiques à une tâche pour la conversion de la poésie sanskrite en prose. La recherche montre que les LLMs, souvent considérés comme des solutions universelles, rencontrent des difficultés avec des langues à faibles ressources comme le sanskrit, où des modèles spécialisés comme ByT5-Sanskrit les surpassent de manière significative. Cette découverte est cruciale pour faire progresser les technologies NLP dans des contextes linguistiques divers.

A recent study compares large language models (LLMs) with smaller, task-specific models in converting Sanskrit poetry to prose. The research highlights that LLMs, often seen as universal solutions, struggle with low-resource languages like Sanskrit, where specialized models like ByT5-Sanskrit significantly outperform them. This finding is crucial for advancing NLP technologies in diverse linguistic contexts.

Still Not There: Can LLMs Outperform Smaller Task-Specific Seq2Seq Models on the Poetry-to-Prose Conversion Task?

Was this article worth reading? Share it

Ready to build your own newsroom?