文体学视角下文学文本的机器翻译风格保留研究——以莫言《红高粱家族》为例

A Study on the Preservation of Stylistic Features in Machine Translation of Literary Texts from the Perspective of Stylistics—A Case Study of Mo Yan’s Red Sorghum

摘要: 在人工智能技术推动下，神经机器翻译(NMT)在通用文本处理中成效显著，但其对文学文本独特风格的保留能力仍有待检验。本文以文体学理论为指导，以莫言《红高粱家族》为对象，对比葛浩文英译本与有道、DeepL等主流NMT系统译文，从词汇、句法、修辞、叙事及语境五个维度，考察NMT的风格传递表现。研究发现，NMT可处理部分显性风格特征(如常规比喻与句法)，但面对通感修辞、文化负载词、情感语调及叙事整体性等深层要素时，呈现明显的机械复制与扁平化倾向，其“创造性”本质为概率生成，并非有意识的艺术选择。据此，本文认为当前NMT尚无法独立完成文学风格的创造性翻译；未来文学翻译的理想路径在于人机协作，即NMT负责语义转换与初步风格模仿，人类译者负责在机器译文基础上打磨风格、融入文化内涵。本研究为界定机器翻译的文学应用边界、优化协作模式提供了理论与实证参考。

Abstract: Driven by advances in artificial intelligence, neural machine translation (NMT) has achieved notable success in general text processing, but its ability to preserve the unique style of literary texts remains to be tested. Drawing on stylistics, this paper examines Mo Yan’s Red Sorghum, comparing Howard Goldblatt’s English translation with translations produced by mainstream NMT systems such as Youdao and DeepL. It assesses NMT’s performance in conveying style across five dimensions: vocabulary, syntax, rhetoric, narrative, and context. The study finds that NMT can handle some explicit stylistic features (e.g., conventional metaphors and syntax), but when faced with deeper elements such as synesthetic rhetoric, culture-loaded words, emotional tone, and narrative coherence, it shows a clear tendency toward mechanical replication and flattening. Its “creativity” is essentially probabilistic generation rather than conscious artistic choice. Accordingly, this paper argues that current NMT cannot independently achieve creative translation of literary style. The ideal path for future literary translation lies in human-machine collaboration, with NMT handling semantic conversion and preliminary stylistic imitation, while human translators refine the style and incorporate cultural nuances based on the machine output. This study provides theoretical and empirical references for defining the boundaries of machine translation in literary applications and optimizing collaborative models.