Fluency and Faithfulness in Human and Machine Literary Translation

文学翻译中人类与机器的流畅度与忠实度研究

Abstract: Literary translation requires balancing target-language fluency with faithfulness to the source. Recent large language models (LLMs) often produce fluent translations, but it remains unclear whether fluency corresponds to semantic preservation in literary text.

摘要： 文学翻译需要在目标语言的流畅度与对源文本的忠实度之间取得平衡。近期的大型语言模型（LLMs）往往能生成流畅的译文，但目前尚不清楚这种流畅度是否等同于文学文本中语义的完整保留。

We examine this relationship using 130,486 translated paragraphs from 106 novels in 16 source languages, including human, Google Translate, and TranslateGemma translations. Fluency is measured as original-likeness with a translationese classifier trained on paragraph part-of-speech n-grams, and faithfulness with the automatic translation evaluation metric COMET-KIWI.

我们通过分析来自 16 种源语言、106 部小说的 130,486 个翻译段落，探讨了这一关系，其中包括人工翻译、谷歌翻译（Google Translate）以及 TranslateGemma 的翻译结果。我们使用基于段落词性 n-gram 训练的“翻译腔”分类器来衡量流畅度（即与原文的相似度），并使用自动翻译评估指标 COMET-KIWI 来衡量忠实度。

We control for paragraph length and find a consistent negative correlation between fluency and faithfulness. The pattern appears for both human and Google Translate, but is weaker and often non-significant for TranslateGemma. These results show that segment length matters for automatic evaluation and suggest a tradeoff between fluency and faithfulness in literary translation.

在控制了段落长度变量后，我们发现流畅度与忠实度之间存在持续的负相关关系。这种模式在人工翻译和谷歌翻译中均有体现，但在 TranslateGemma 中表现较弱，且往往不显著。这些结果表明，片段长度对自动评估至关重要，并暗示了文学翻译中流畅度与忠实度之间存在权衡。