Performance Limitations of Large Language Models in Processing Low-Resource Regional Languages: Empirical Evidence and Pre-Training Optimization Strategies

Abstract

While Large Language Models (LLMs) such as ChatGPT and Gemini have transformed information processing in resource-rich languages, their effectiveness significantly diminishes when applied to low-resource regional languages (e.g., Sundanese and Javanese). This paper investigates the performance bottlenecks experienced by users, focusing on latency and output fluency. Through a quantitative survey (N = 50), we argue that these limitations stem from underrepresentation during the model pre-training phase and inefficient tokenization. Consequently, we propose prioritizing vocabulary expansion and open-source regional corpus collection during the pre-training stage to enhance AI inclusivity and fluency.