A Supervised Learning Approach Towards Profiling the Preservation of Authorial Style in Literary Translations

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 376–386, Dublin, Ireland, August 23-29 2014. A Supervised Learning Approach Towards Profiling the Preservation of Authorial Style in Literary Translations Gerard Lynch Centre for Applied Data Analytics Research University College Dublin Ireland firstname.lastname@ucd.ie Abstract Recently there has been growing interest in the application of approaches from the text classi- fication literature to fine-grained problems of textual stylometry. This paper seeks to answer a question which has concerned the translation studies community: how does a literary transla- tor’s style vary across their translations of different authors? This study focuses on the works of Constance Garnett, one of the most prolific English-language translators of Russian literature, and uses supervised learning approaches to analyse her translations of three well-known Rus- sian authors, Ivan Turgenev, Fyodor Dosteyevsky and Anton Chekhov. This analysis seeks to identify common linguistic patterns which hold for all of the translations from the same author. Based on the experimental results, it is ascertained that both document-level metrics and n-gram features prove useful for distinguishing between authorial contributions in our translation corpus and their individual efficacy increases further when these two feature types are combined, result- ing in classification accuracy of greater than 90 % on the task of predicting the original author of a textual segment using a Support Vector Machine classifier. The ratio of nouns and pronouns to total tokens are identified as distinguishing features in the document metrics space, along with occurrences of common adverbs and reporting verbs from the collection of n-gram features. 1 Introduction The application of supervised learning technologies to textual data from the humanities in order to shed light on stylometric questions has become more popular of late. In particular, these approaches have been applied to questions from the field of translation studies, which concern the notion of translationese1 detection in Italian and other languages, (Baroni and Bernardini, 2006; Ilisei et al., 2010; Ilisei and Inkpen, 2011; Popescu, 2011; Koppel and Ordan, 2011; Lembersky et al., 2011). Work has also been carried out on source language detection from translation corpora, (van Halteren, 2008; Lynch and Vogel, 2012) and translation direction detection in parallel MT training corpora, (Kurokawa et al., 2009), which can have applications in the domain of machine translation where the direction of bilingual translation corpora has been shown to impact on the accuracy of automated translations using such corpora2. This work seeks to apply these methods to the task of identifying authorial style within a corpus of translations by the same translator. Venuti (1995) mentions the concept of the translator’s invisibility, that the measure of the best translator is that their style is not distinguishable in the translation, that their main concern and focus is to deliver the original text in a faithful manner. Of course, this task is often subject to their own vocabulary choices and as was often the case, cultural or personal bias of the translator or the regime or government in which they were …

Content truncated to 500 words