Unsupervised Identification of Translationese

Unsupervised Identification of Translationese Ella Rabinovich Shuly Wintner Department of Computer Science Department of Computer Science University of Haifa University of Haifa ellarabi@csweb.haifa.ac.il shuly@cs.haifa.ac.il Abstract Translated texts are distinctively different from original ones, to the extent that su- pervised text classification methods can dis- tinguish between them with high accuracy. These differences were proven useful for sta- tistical machine translation. However, it has been suggested that the accuracy of transla- tion detection deteriorates when the classifier is evaluated outside the domain it was trained on. We show that this is indeed the case, in a variety of evaluation scenarios. We then show that unsupervised classification is highly ac- curate on this task. We suggest a method for determining the correct labels of the clustering outcomes, and then use the labels for voting, improving the accuracy even further. More- over, we suggest a simple method for cluster- ing in the challenging case of mixed-domain datasets, in spite of the dominance of domain- related features over translation-related ones. The result is an effective, fully-unsupervised method for distinguishing between original and translated texts that can be applied to new domains with reasonable accuracy. 1 Introduction Human-translated texts (in any language) have dis- tinct features that distinguish them from original, non-translated texts. These differences stem ei- ther from the effect of the translation process on the translated outcomes, or from “fingerprints” of the source language on the target language product. The term translationese was coined to indicate the unique properties of translations. Awareness to translationese can improve statis- tical machine translation (SMT). First, for training translation models, parallel texts that were translated in the direction of the SMT task are preferable to texts translated in the opposite direction; second, for training language models, monolingual corpora of translated texts are better than original texts. It is possible to automatically distinguish between original (O) and translated (T) texts, with very high accuracy, by employing text classification methods. Existing approaches, however, only employ super- vised machine-learning; they therefore suffer from two main drawbacks: (i) they inherently depend on data annotated with the translation direction, and (ii) they may not be generalized to unseen (related or unrelated) domains.1 These shortcomings under- mine the usability of supervised methods for trans- lationese identification in a typical real-life scenario, where no labelled in-domain data are available. In this work we explore unsupervised techniques for reliable discrimination of original and translated texts. More precisely, we apply dimension reduction and centroid-based clustering methods (enhanced by internal clustering evaluation), for telling O from T in an unsupervised scenario. Furthermore, we in- troduce a robust methodology for labelling the ob- tained clusters, i.e., annotating them as “original” or “translated”, by inspecting similarities between the clustering outcomes and O and T prototypical ex- amples. Rigorous experiments with four diverse cor- pora demonstrate that clustering of in-domain texts using lexical, content-independent features system- atically yields very high accuracy, only 10 percent points lower than the performance of supervised classification on the same data (in most cases). Ac- …

Content truncated to 500 words