RusLTC
popular-scientific description of RFBR project No. 17-06-00107 (2017–2019)
Linguistic properties of translations: descriptive norms and translation quality estimation (based on English-to-Russian mass-media texts)
This research project aims to produce a list of translationese indicators for English-Russian language pair, i.e. to establish a set of human-interpretable linguistic features that reflect the ontological status of any translation, their property of “being a translation”. These properties are revealed through comparison between translations and originally authored texts in the target language (“non-translations”). This comparison is implemented as a binary text classification task in the machine learning approach. The results of the research are largely contingent on the quality and comparability of the underlying corpora. It is not unexpected because (1) specific text categories (registers, genres, functional styles) are likely to demonstrate individual translational tendencies; (2) it makes sense to estimate the approximation of translation (as a measure of fluency) to the expected target language norm regarding the register and genre, rather than to a collection of genre-incompatible or under-defined collection of non-translations.
One of the major hypotheses put forward in this work is the correlation between the degree of translationese effect (i) and the quality of translation, (ii) and the professional varieties of translations (professional and student translations). Besides, we tested whether the same translationese indicators can be used to capture translationese in another language pair (English-German). Given that many of our indicators come from the intersection of universal text-internal properties used in cross-linguistic register studies (Biber, 1988), this assumption makes sense.
The project results include a system of significant corpus resources designed to study translation quality and translation varieties. The core component of this system – Russian Learner Translator Corpus (RusLTC) is publicly available.
Theoretically, an aspect of corpus comparability that is important for the study of translationese and translation varieties is the functional text comparability. Functional properties of texts reflected in the genre classification. However, discrete genre labels do not account for hybrid texts and it is not clear how similar are the texts from the text categories of the same name annotated within different projects. To address these issues, we used the approach and data sets provided in (Sharoff, 2018) and developed a method to identify functionally similar subsets of texts in heterogeneous collections for both monolingual and cross-linguistic setup. Our results are verified on the genre annotation provided by the British National Corpus and Russian National Corpus as well as within two annotation experiments carried out by the project research group (Kunilovskaya & Sharoff, 2019; Kunilovskaya et al., 2019).
Our major research direction is associated with discovering translationese indicators. Translationese indicators are linguistic properties of translations that distinguish them from the genre-comparable non-translations in the target language (within this project – translated and non-translated mass-media texts). A few examples of clear translationese indicators include frequencies of possessive pronouns, relative clauses, deverbal process nouns, subordinating conjunctions. The resulting list of 45 indicators is based on the previous research in translationese studies, didactics of translation, contrastive stylistics and especially in corpus-based translation studies. The extraction of the values for these features is based on automatic morphosyntactic annotation (Straka & Straková, 2017) and lists of lexical items (mostly, function words and discourse markers) verified by experts.
Generally, the intuitively reasonable assumption that translationese is more characteristic for student translations has been confirmed by our results for both English-German and English-Russian data. However, the distinctions revealed by our (admittedly limited) range of features are not significant enough for automatic classification of student and professional translations.
Despite the suggested feature set reliably captures translationese (the accuracy of the classification into translations and non-translations reaches 83% for German and 92% for Russian), it cannot be used as is to predict translation quality. We have seen only weak correlation between the translationese effects and quality (annotated as binary contrastive classes of good and bad translations based on human expert judgments). It is clear that our quality annotation (if it is to be trusted) reflects many other and more important aspects of translation quality that run beyond morphosyntactic acceptability.