Automatic Detection of Translation Direction

Proceedings of Recent Advances in Natural Language Processing, pages 1131–1140, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_130 1131 Automatic Detection of Translation Direction Ilia Sominsky and Shuly Wintner Department of Computer Science University of Haifa 31905 Haifa, Israel Abstract Parallel corpora are crucial resources for NLP applications, most notably for ma- chine translation. The direction of the (hu- man) translation of parallel corpora has been shown to have significant implica- tions for the quality of statistical machine translation systems that are trained with such corpora. We describe a method for determining the direction of the (man- ual) translation of parallel corpora at the sentence-pair level. Using several linguistically-motivated features, coupled with a neural network model, we obtain high accuracy on several language pairs. Furthermore, we demonstrate that the ac- curacy is correlated with the (typological) distance between the two languages. 1 Introduction Parallel corpora are used for various purposes, including for training and evaluation of statisti- cal machine translation (SMT) systems (Koehn, 2010). While traditional SMT systems are agnos- tic with respect to the direction in which the par- allel corpora they are trained on were (manually) translated, several studies have shown that tak- ing directionality into account when training SMT systems has a significant effect on the quality of the translation (Kurokawa et al., 2009; Lembersky et al., 2012, 2013; Twitto-Shmuel et al., 2015). In this paper we show the same effect also holds for neural machine translation (NMT) systems. We address the task of determining the direc- tion of translation given a parallel text; this is cast as a binary classification task. To strain the classi- fier, we focus on retaining high accuracy when the size of text chunks to be classified is minimal: sin- gle sentence pairs. This is an extremely difficult task for humans, in most cases: a single sentence pair often does not reveal any obvious signal of which of the two sentences is the original. It is also a highly challenging task for machines: Fig- ure 1 depicts a few English-French examples of sentence pairs whose translation direction none of our classifiers predicted correctly. We define sets of features that reflect insights drawn from Translation Studies regarding the spe- cial properties of translated texts, and in particular the asymmetric nature of translation (Toury, 1980, 1995; Baker, 1993). These include the tendency of translated texts to be simpler (Blum-Kulka and Levenston, 1983; Vanderauwerea, 1985; Baker, 1993; Laviosa, 1998, 2002); the tendency of trans- lators to explicate the source text (Blum-Kulka, 1986; Baker, 1993); the different distributions of various statistical phenomena (e.g., the fre- quencies of function words or certain syntactic structures) between the source and the translation (Gellerstam, 1986; Blum-Kulka, 1986; Øver˚as, 1998; Koppel and Ordan, 2011); and interference of language constructions from the source to the target (Toury, 1979; Teich, 2003). The contribution of this paper is manifold. (1) First and foremost, we introduce a method for accurately determining the translation direction of sentence pairs in parallel corpora; the method is based on the introduction …

Content truncated to 500 words