同步口譯之英語關係子句處理：人類與人工智慧輸出之語料庫研究

Abstract

本研究以語料庫為基礎，探討人類口譯員與人工智慧系統於英進中同步口譯處理英語關係子句之處理方式差異。研究分析取自政治辯論及科技演講的真實人類口譯資料，以及Claude 3.5 Sonnet V2與Wordly.ai的機器翻譯，指出人類與人工智慧系統在語法與語意處理方式上的明顯差異。研究結果顯示，人類口譯員傾向運用較多元的語法處理方式，包括順譯（40.97%）、重組（22.47%）、先行詞刪除（11.89%）、省略（8.37%）、關係子句刪除（8.37%）及改述（7.93%）。人類口譯員亦常採用語意簡化（28.19%），通常似乎是為了解決時間壓力，並優先排序主要訊息。此外，研究亦發現人類口譯員會使用增譯（11.89%），可能是為了釐清講者訊息或增強語句銜接，並可能有助於促進聽眾參與。相較之下，人工智慧系統主要採用重組（48.76%）與順譯（44.49%），經常高度保留原意，但較少採用增譯（7.64%）或簡化（0.45%）。人工智慧系統產出似乎缺乏釐清講者溝通目的或為聽眾調整訊息的溝通彈性，且似乎在結構分析與語境理解上遇到挑戰。此研究補足英進中口譯關係子句實證研究之不足，並為口譯教學、人工智慧系統優化及未來人機協作提供參考建議。

This corpus-based study investigated how human interpreters and artificial intelligence (AI) systems manage English relative clauses during English-to-Chinese simultaneous interpretation. By analyzing authentic human interpretation data and machine-generated translations from Claude 3.5 Sonnet V2 and Wordly.ai, sourced from a political debate and a technical speech, the study identified notable differences in syntactic and semantic approaches between humans and AI systems. The findings suggested that human interpreters tend to employ a broader range of syntactic approaches —including Syntactic Linearity (40.97%), Restructuring (22.47%), Deletion of the Antecedent (11.89%), Omission (8.37%), Deletion of the Relative Clause (8.37%), and Paraphrasing (7.93%). Humans also frequently use semantic Simplification (28.19%), which often appears to be a response to time constraints. Addition (11.89%) is also observed, which may serve to clarify the speaker’s message or enhance cohesion, and could potentially facilitate audience engagement. In contrast, AI outputs predominantly exhibit Restructuring (48.76%) and Syntactic Linearity (44.49%), largely retaining the original message while showing limited instances of Addition (7.64%) or Simplification (0.45%). AI outputs appear to lack the communicative flexibility to clarify the speaker’s intent or adapt to the audience and seem to encounter challenges in structural parsing and contextual understanding. These findings contribute to the limited empirical research on relative clauses in English-to-Chinese interpretation and provide insights for interpreter training, AI system improvement, and future human–machine collaboration.