Localized Large Language Models Performance Evaluation in Low-Resource Languages Based on Real Governmental and Industrial Text Corpora

Large Language Models (LLMs) have demonstrated remarkable capabilities across a wide range of natural language processing tasks; however, their performance remains uneven across languages, particularly in low-resource settings. Most existing LLMs are trained predominantly on high-resource languages, which raises concerns regarding their effectiveness, fairness, and reliability when applied to localized contexts such as governmental and industrial environments in low-resource languages. This study presents a systematic and empirical evaluation of localized LLMs with a specific focus on their performance across diverse low-resource language tasks using real-world governmental and industrial text corpora. The research adopts a multi-dimensional evaluation framework that assesses linguistic accuracy, semantic consistency, robustness to domain-specific terminology, and computational efficiency. Multiple transformer-based language models, including multilingual and language-specific variants, are evaluated across representative tasks such as text classification, named entity recognition, and semantic similarity analysis. The evaluation framework is designed to reflect practical deployment conditions, emphasizing domain adaptation and contextual relevance rather than benchmark-centric performance alone. A comprehensive experimental pipeline is proposed, integrating tokenizer-level analysis, transfer learning effectiveness, and adaptation strategies suitable for limited-data scenarios. Quantitative results are presented through detailed comparative tables and multi-parameter visualizations, highlighting performance disparities between multilingual general-purpose models and localized models fine-tuned on domain-specific corpora. The findings reveal that localized models consistently outperform generic multilingual counterparts in domain-sensitive tasks, while also exhibiting improved stability and interpretability. Beyond performance metrics, the study provides analytical insights into structural and linguistic factors influencing model behavior in low-resource environments. The results underscore the importance of localized data representations, adaptive model architectures, and evaluation protocols tailored to real institutional use cases. This research contributes a reproducible evaluation framework and empirical evidence that can inform future development, policy-making, and deployment of LLMs in governmental and industrial contexts where linguistic resources are limited.

کلیدواژه ها:

Large Language Models ، Low-Resource Languages ، Model Localization ، Governmental and Industrial Text Analysis ، Performance Evaluation