A Classification-Based Approach to Cognate Detection Combining Orthographic and Semantic Similarity Information

Proceedings of Recent Advances in Natural Language Processing, pages 602–610, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_071 602 A Classification-Based Approach to Cognate Detection Combining Orthographic and Semantic Similarity Information Sofie Labat and Els Lefever LT3, Language and Translation Technology Team, Ghent University Groot-Brittanni¨elaan 45, 9000 Ghent, Belgium sofie.labat@gmail.com, els.lefever@ugent.be Abstract This paper presents proof-of-concept ex- periments for combining orthographic and semantic information to distinguish cog- nates from non-cognates. To this end, a context-independent gold standard is de- veloped by manually labelling English- Dutch pairs of cognates and false friends in bilingual term lists. These annotated cognate pairs are then used to train and evaluate a supervised binary classification system for the automatic detection of cog- nates. Two types of information sources are incorporated in the classifier: fifteen string similarity metrics capture form sim- ilarity between source and target words, while word embeddings model semantic similarity between the words. The ex- perimental results show that even though the system already achieves good results by only incorporating orthographic infor- mation, the performance further improves by including semantic information in the form of embeddings. 1 Introduction In general linguistics, the term cognate is defined as a “language or a linguistic form which is his- torically derived from the same source as another language/form” (Crystal, 2008, page 83). The assumption of common etymology is, however, often disregarded in the literature, because cer- tain research areas such as psycho-linguistics or natural language processing (NLP) tend to shift their focus from diachronic to perceptual related- ness (Shlesinger and Malkiel, 2005; Mitkov et al., 2007; Schepens et al., 2013; Hansen-Schirra et al., 2017). We follow this second strand of research in that we define cognates as words with high formal and semantic cross-lingual similarity. Conversely, false friends are words which have similar forms, but which differ in their meaning. The ability to distinguish cognates from non- cognates (and especially) false friends is an im- portant skill for second language learners. Simi- larly, source language interference is a problem of- ten experienced by translators that is partly caused by the influence of cognates and false friends. Research in natural language processing can ad- dress these bottlenecks by, for instance, develop- ing computer tools that aid second language users. Nevertheless, most studies have mainly focused on the detection of cognates (Bergsma and Kon- drak, 2007; Hauer and Kondrak, 2011; Ciobanu and Dinu, 2014; Rama, 2016), while relatively little attention has been devoted to false friends (Frunza and Inkpen, 2007; Mitkov et al., 2007; Ljubeˇsi´c and Fiˇser, 2013; Castro et al., 2018). Mitkov (2007) explains that the main goal of in- vestigation is often the cross-lingual identification of equivalent lexical items, as such knowledge can be integrated in other applications. Automatic cognate detection has indeed proven very useful for NLP, e.g. to boost the performance of auto- matic alignment between related languages or to compile bilingual lexicons (Smith et al., 2017). The aim of this research is twofold: (1) we introduce a context-independent gold standard which can be used to …

Content truncated to 500 words