Looking under the hood: Which linguistic features contribute to the source language classification of direct and indirect translations into Finnish, and why is that?

Abstract The study of features that affect the linguistic form of translated texts has been one of the central questions within the field of corpus-based translation studies. In the partially overlapping field of computational linguistics, previous studies have shown that source languages of individual texts can be detected automatically in direct translations and indirect translations (i.e., translations done from translations). However, computationally oriented approaches have paid limited attention to what specific linguistic features make successful classification possible. Consequently, the types of linguistic phenomena characterizing translations and the kinds of linguistic interference that can be detected in them remain underexplored. In this study, we study the linguistic features that contribute to the identification of the source language of direct translations from English, French, German, Greek, and Swedish, as well as indirect translations from Greek into Finnish, with English, French, German, and Swedish as mediating languages. Theoretically, this study builds on Halverson's (2017) gravitational pull model to explain the mechanisms behind our findings in a theoretically sound fashion and to generate theoretically motivated, specific hypotheses to be tested by future research. The analysis makes use of keyness analysis as a supervised machine learning technique, as well as exploratory factor analysis (EFA) as an unsupervised machine learning technique. The results indicate that sentence length, sentence-initial adverbs and sentence-final specification are the linguistic features that set the different types of translations apart from each other. Furthermore, the salient features of the ultimate source language outweigh those of the mediating languages in indirect translations or the entrenched parallels between specific language pairs.