An Appropriate-Technology Approach to On-Premise AI Multilingual Translation of Village Broadcasting

Abstract

Rural village broadcasting operates almost exclusively in Korean, leaving marriage immigrants and migrant workers without access to its predominantly disaster-safety content; multilingual support for this public channel is effectively 0%. Addressing this information-inclusivity gap, this study takes an appropriate-technology approach delivering the best quality realistically attainable within field — constraints rather than state-of-the-art performance and empirically examines how far an on-premise — lightweight AI translation system can expand information inclusivity. The system raises accessibility from 0% to 100% by providing translated text and attains 75.4% comprehensibility at the level of actionable understanding, on the basis of which policy recommendations for extending Article 11 of the Multicultural Family Support Act to village broadcasting are derived. First, through text mining of 86,413 village broadcast messages collected from 23 municipalities nationwide, this study systematically analyzed the characteristics of village broadcast content from the perspective of multilingual translation. The analysis revealed that 62.9% of unique messages are disaster-safety messages, of which 75.0% are classified as urgent, confirming that village broadcasting serves as a critical public information channel directly linked to life and safety. Additionally, extreme diversity in broadcast type composition across regions and stylistic variations attributable to operator tendencies were identified. Second, the performance of a two-stage pivot translation pipeline combining a Korean specialized small large language model （GECKO-7B） with a multilingual neural machine translation model （NLLB-200 1.3B） was validated. Operating as a self-contained on-premise deployment on a consumer-grade GPU with 8GB VRAM, the system achieved an adequate quality rate of 75.4% for translations into Vietnamese, Thai and Khmer, under an operational criterion of 3 or above on a 5-point LLM-as-a-Judge scale. Through a three-axis evaluation framework integrating LaBSE multilingual semantic similarity, IFS （Information Fidelity Score） for key information preservation, and LLM-as-a-Judge qualitative assessment, the key finding was derived that type-specific translation quality disparities are more pronounced in information delivery （IFS gap: 45.1%） than in semantic delivery （LaBSE gap: 12.6%）. Furthermore, a TF-IDF-based emoji sequence mapping system was designed as an independent visual safety net against translation errors, and its structural validity was presented. Third, policy recommendations were derived for extending the legislative intent of Article 11 of the Multicultural Family Support Act to village broadcasting. These recommendations were systematized along three axes legislative and institutional framework, technical standardization and ― governance, and diffusion strategy encompassing phased mandates based on foreign resident ratios, ― incorporation of multilingual requirements into smart village equipment specifications, a centralized development with distributed deployment model for standard translation governance, and a three-stage diffusion strategy prioritizing broadcast types and regions with higher adequate quality rates. This study offers four academic contributions: empirical analysis of rural village broadcast data, demonstration of the feasibility of on-premise distributed AI translation, a reference-free three-axis evaluation methodology for low-resource language environments, and the application of appropriate technology theory to AI system design.