Computational analysis of different translations: by professionals, students and machines

Computational analysis of different translations: by professionals, students and machines Maja Popovi´c1, Ekaterina Lapshinova-Koltunski2, Maarit Koponen3 1 ADAPT Centre, School of Computing, Dublin City University, Ireland maja.popovic@adaptcentre.ie 2 Linguistics and Information Sciences, University of Hildesheim, Germany lapshinovakoltun@uni-hildesheim.de 3 School of Humanities, University of Eastern Finland maarit.koponen@uef.fi Abstract In this work, we analyse translated texts in terms of various features. We compare two types of human translations, profes- sional and students’, and machine trans- lation (MT) outputs in terms of lexical and grammatical variety, sentence length, as well as frequencies of different part-of- speech (POS) tags and POS-trigrams. Our analyses are carried out on parallel trans- lations into Croatian, Finnish and Russian, all originating from the same source En- glish texts. Our results indicate that ma- chine translations are the closest to the source text, followed by student transla- tions. Also, student translations are some- times more similar to MT than to profes- sional translations. Furthermore, we iden- tify sets of features distinctive for machine translations. 1 Introduction It is well-known that there is generally more than one way to translate any given source text (seg- ment) and that versions created by different hu- man translators therefore can vary from each other. Variation between different translators has been observed in terms of various linguistic features from lexis to syntax (see Section 2). As there is usually no single correct translation, these diverg- ing versions may be equally good despite their dif- ferences. On the other hand, it has also been ob- served that machine translations differ from human translations in ways that might contain errors. Dis- tinguishing genuine variation in choices of lexical © 2023 The authors. This article is licensed under a Creative Commons 4.0 licence, no derivative works, attribution, CC- BY-ND. or grammatical expressions from the kind of di- vergence that indicates actual errors or other qual- ity issues would be important for example for ma- chine translatio evaluation. Separating these two would require clearer understanding of how di- verging translation versions in fact differ from each other. So, we analyse differences between texts trans- lated by MT systems and those translated by two groups of human translators: professionals and students. Although previous studies (see Section 2 below) already compared such translation variants, they focused on one language pair and different genres, and did not consider neural MT. We also want to compare translations with their sources, as close resemblance to the source text could indicate more literal translations, which may be less than optimal in terms of fluency and style, even if the meaning is correct. Besides that, we investigate in which aspects in terms of linguistic features trans- lations resemble each other. Thus, the main goals of this work are: RG1 to re-examine linguistic features from previ- ous work on a parallel data set and three target languages; RG2 to automatically distinguish between source texts, professional, student and machine translations; RG3 to further explore linguistic features in terms of distinctiveness for every translation variant under analysis. 2 …

Content truncated to 500 words