This paper applies classification and clustering methods in machine learning studies, builds a parallel corpus, and examines the translation styles of the three versions of Hardy's masterpiece Tess of the D'Urberville. From a total of 68 features,15 significant ones are selected and quantitatively synthesized with examples for detailed explanation. The results show that these salient features can effectively distinguish the stylistic differences among the three translations, with both classifying and clustering experiments achieving an average accuracy rate of about 97%. The study found that at the document-level, each translation shows different style features at the vocabulary, syntax, and discourse aspects; in terms of the keyword level, the frequency differences of certain keywords also present the translator's personal preferences. The article provides data support and fine-grained analysis for previous qualitative research, while also proposing some corrective conclusions, such as the higher lexical density, extremely lower proportion of passive bei sentence, and similar number of idioms in Zhang's translation compared to the others. Eventually we attempt to provide some improvements and supplements to the research methodology in translation style and translator's style studies.

machine learning , translation style , parrel corpus , Tess of the D'Urbervilles