[4I4-OS-17a] OS-17

1:30 PM - 1:45 PM JST(4:30 AM - 4:45 AM UTC)

[4I4-OS-17a-01] Attempt at Neural Machine Translation for Endangered Languages Using LLM-RAGMulti-Metric Evaluation of Coptic, Northern Ryukyuan Yoron, Truku, and Ainu and Impact of Language Resource Availability

〇So Miyagawa¹ (1. University of Tsukuba)

Keywords:

Endangered Language,Artificial Intelligence (AI),Neural Machine Translation (NMT),Retrieval-Augumented Generation (RAG),Large Language Model (LLM)

This study explores machine translation for four endangered languages from distinct language families: Coptic [cop] (Afro-Asiatic, Egyptian), Northern Ryukyuan Yoron [yox] (Japonic, Ryukyuan), Truku [trv] (Austronesian, Formosan), and Ainu [ain] (isolate). We construct RAG (retrieval-augmented generation)-enhanced systems for Coptic and Yoron, a fine-tuned NMT (neural machine translation) system for Ainu, and survey resources for a Truku system under development. Translation quality is evaluated using BLEU, METEOR, chrF, TER, and ROUGE-L metrics. For Coptic and Yoron, RAG improves BLEU by 9.2 and 9.9 points over base LLMs (large language models), respectively; for Ainu, corpus-driven fine-tuning achieves SacreBLEU 32.90 (Japanese→Ainu). We analyze the relationship between digital resource availability and translation quality, finding that resource structure and typological distance, rather than quantity alone, influence NMT performance.