臺語中之不可譯項：AI翻譯中語意流失的語料庫研究

Untranslatables in Taiwanese: A Corpus-Based Study of Semantic Leakage in AI Translations

本研究提出一項多維度指標—語意流失指數（Semantic Leakage Indicator, SLI），用以量化臺語翻譯為華語時，人工智慧譯文中之語意流失。研究使用之語料庫共計26,046組臺語句子，涵蓋多元的口語與書面語境。分析聚焦於五項高度易受語意侵蝕之形態語用特徵：體標記、句尾助詞、重疊形式、親屬稱謂詞，以及動向複合詞。SLI結合語意熵減、回譯差異及語意嵌入位移三項演算法，以識別那些表層流暢度指標難以偵測之「不可譯」結構。研究結果顯示，即便譯文表面上存在詞彙對應，高SLI值案例仍常伴隨語用扭曲與情感不協調。群聚分析則勾勒出不同語法範疇中的語意流失樣態；另輔以質性實例，呈現各文體類型所固有之翻譯脆弱性。此研究為翻譯學提供一套可擴充、具語言敏感度之診斷工具，用以評估人工智慧在低資源、形態複雜語言上的表現，並於「不可譯性」、語用意涵與語意保存等議題上，貢獻具體實證價值，對演算法設計、語言保存與語料本位評估等均具實質啟示意義。

This study proposes a multidimensional metric—Semantic Leakage Indicator (SLI)—to quantify meaning loss in AI-generated translations of Taiwanese into Mandarin. Drawing on a corpus of 26,046 Taiwanese sentences sourced from diverse spoken and written contexts, the analysis identifies five morpho-pragmatic features highly susceptible to semantic erosion: aspect markers, sentence-final particles, reduplication, kinship terms, and directional compounds. The SLI integrates entropy reduction, back-translation divergence, and semantic embedding shifts to detect untranslatable constructions that evade surface-level fluency metrics. Results reveal that even when lexical equivalents are present, high-SLI instances exhibit pragmatic distortion and affective dissonance. Cluster analysis delineates patterns of leakage across grammatical categories, while qualitative examples illustrate genre-specific vulnerabilities. The study advances translation studies by offering a scalable, language-sensitive diagnostic tool for evaluating AI performance on lowresource, morphologically rich languages. It contributes empirically to discussions on untranslatability, pragmatics, and meaning preservation in AI translation, with direct implications for algorithm design, language preservation, and corpus-informed evaluation in underrepresented languages.