Can Peter Pan Survive MT? A Stylometric Study of LLMs, NMTs, and HTs in Children’s Literature Translation
Date:
This study evaluates the performance of machine translations (MTs) versus human translations (HTs) in English-to-Chinese children’s literature translation (CLT) from a stylometric perspective. A Peter Pan corpus was constructed, including 21 translations: 7 HTs, 7 large language model translations (LLMs), and 7 neural machine translation outputs (NMTs). The analysis uses both a generic feature set (lexical, syntactic, readability, n-gram features) and a creative text translation (CTT-specific) feature set (repetition, rhythm, translatability, miscellaneous), totaling 447 linguistic features.
Stylometric analysis using classification and clustering techniques reveals that, for generic features, HTs and MTs differ significantly in conjunction word distributions and the ratio of 1-word-gram-YiYang. NMTs and LLMs differ in descriptive word usage and adverb ratios. For CTT-specific features, LLMs outperform NMTs in distribution and align more closely with HTs in stylistic characteristics, highlighting the potential of LLMs in CLT.
Comments: 19 pages, 8 figures, 4 tables. Accepted at the 2nd Workshop on Creative-text Translation and Technology, co-located with MT Summit 2025. Official paper may later be accessed from ACL Anthology.
PPT: Link