Tracing Source Language Interference in Translation with Graph-Isomorphism Measures

Proceedings of Recent Advances in Natural Language Processing, pages 375–385 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_043 375 Tracing Source Language Interference in Translation with Graph-Isomorphism Measures Koel Dutta Chowdhury Saarland Informatics Campus Saarland University koeldc@lst.uni-saarland.de Cristina Espa˜na-Bonet DFKI GmbH, Saarbr¨ucken cristinae@dfki.de Josef van Genabith Saarland University DFKI GmbH, Saarbr¨ucken josef.van genabith@dfki.de Abstract Previous research has used linguistic features to show that translations exhibit traces of source language interference and that phylo- genetic trees between languages can be re- constructed from the results of translations into the same language. Recent research has shown that instances of translationese (source language interference) can even be detected in embedding spaces, comparing embeddings spaces of original language data with embed- ding spaces resulting from translations into the same language, using a simple Eigenvector- based divergence from isomorphism measure. To date, it remains an open question whether alternative graph-isomorphism measures can produce better results. In this paper, we (i) ex- plore Gromov-Hausdorff distance, (ii) present a novel spectral version of the Eigenvector- based method, and (iii) evaluate all ap- proaches against a broad linguistic typologi- cal database (URIEL). We show that language distances resulting from our spectral isomor- phism approaches can reproduce genetic trees on a par with previous work without requiring any explicit linguistic information and that the results can be extended to non-Indo-European languages. Finally, we show that the methods are robust under a variety of modeling condi- tions. 1 Introduction The study of cross-linguistic variation has been a key focus of linguistics for genetic or typological classification of languages. Historical comparative linguistic methods determine genetic relationships between languages using concept lists of words with a common origin in multiple languages that share similar meaning and pronunciation (Swadesh, 1952; Dyen et al., 1992). Linguistic typology stud- ies how distinct languages are, and what gener- alizations can be made regarding cross-linguistic variation on different levels of linguistic analysis and representations (Trask, 2000). Comrie (1989), for example, studies language variance in terms of their functional processes, whereas Cysouw (2013) measures language distance using structural fea- tures. More recent research indicates that semantic similarity between languages can serve as a quan- titative means to determine cross-linguistic varia- tion across languages. Seminal work of Eger et al. (2016) provides evidence that semantic alignment between languages can be explained by geographi- cal factors. Likewise, Thompson et al. (2018) find that differences correlate with cultural distances among societies speaking the languages. Conversely, it has also been shown that language differences are so profound that the structure of a language is approximately preserved even when translated into another language. This is often re- ferred to as source language interference (Toury, 2012). Rabinovich et al. (2017) show that source languages of translations into the same target lan- guage can be clustered solely based on interfer- ence phenomena in the translations in the target language using simple linguistic features and that these clusters correspond with genetic distance. In a similar vein, Bjerva et al. (2019) find that com- parable results can be established by clustering …

Content truncated to 500 words