跨語言遷移學習應用菲律賓語於達悟語—中文機器翻譯之研究

Cross-Lingual Transfer Learning from Filipino to Tao–Chinese Machine Translation

本研究以達悟語（雅美語）—中文機器翻譯為研究對象，探討跨語言遷移學習在低資源且方言與拼寫變體並存情境下之可行策略與成效。前期實驗顯示，受限於平行語料規模有限及語料異質性，自零訓練之基線模型易出現泛化不足、語意偏離，以及語法結構、詞彙辨識與疑問句處理不穩定等問題。

為改善上述限制，本研究選用 Meta AI 所開發之多語言預訓練模型 NLLB-200-distilled-1.3B，採「代理來源語設定」（proxy source setting），以菲律賓他加祿語（Filipino/Tagalog）作為跨語遷移之代理來源語，並藉由達悟語—中文平行語料進行監督式微調，同時實施資料清理、句對品質檢核、變體整理與資料增強，以降低資料噪音並提升對齊一致性。研究結果顯示，遷移微調模型之整體翻譯品質優於基線模型；在達悟語→中文方向，模型於較長或結構較複雜句型（如疑問句詞序與否定結構）、詞彙對應，以及方言與拼寫變體處理的容錯性均有明顯改善。

結果亦指出，語料品質與多樣性仍為影響低資源翻譯效能的關鍵因素，而前處理工作對模型穩健性與譯文品質具有重要影響。中文→達悟語方向之補充性人工評估亦顯示模型表現已有初步改善，惟在低頻詞、文化詞彙對應、較複雜句型與特定語用條件下，仍有優化空間。最後，本研究提出一套可複製之低資源雙語資源建置與遷移學習流程，可供其他低資源語言翻譯研究與資源開發參考。

 

This study investigated Tao (Yami)–Chinese machine translation to examine strategies and the effectiveness of cross-lingual transfer learning under low-resource conditions with co-existing dialectal and orthographic variation. Preliminary testing suggested that, given the limited size and heterogeneity of the parallel corpus, a from-scratch baseline model exhibited weak generalization and semantic drift, with unstable performance in syntactic parsing, lexical identification, and the handling of interrogatives.

To address these limitations, we built on Meta AI’s multilingual pretrained model NLLB-200-distilled-1.3B and adopted a proxy source setting, with Filipino/Tagalog serving as the proxy source language for cross-lingual transfer during supervised fine-tuning on Tao–Chinese parallel data. We also applied data-level procedures including data cleaning, parallel sentence-pair quality checks, and variant consolidation and augmentation to reduce noise and improve cross-lingual correspondence.

Results suggested that the fine-tuned model achieved higher translation quality than the baseline, with clear improvements in long and complex sentence patterns, e.g., word order in interrogatives and negation, lexical correspondence, and better handling of orthographic variation in the Tao-to-Chinese direction. The findings further indicated that corpus quality and diversity remain key determinants of model performance in low-resource settings. A supplementary manual evaluation of Chinese-to-Tao outputs also indicated preliminary improvement, although challenges remained in low-frequency words, cultural vocabulary, complex sentences, and specific pragmatic conditions. Overall, this study provided a reproducible workflow for bilingual resource construction and cross-lingual transfer in low-resource settings, serving as a reference for resource development and system building for other under-resourced languages.