Cross-Family Similarity Learning for Cognate Identification in Low-Resource Languages

Proceedings of Recent Advances in Natural Language Processing, pages 1121–1130, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_129 1121 Cross-Family Similarity Learning for Cognate Identification in Low-Resource Languages Eliel Soisalon-Soininen Department of Computer Science University of Helsinki eliel.soisalon-soininen@helsinki.fi Mark Granroth-Wilding Department of Computer Science University of Helsinki mark.granroth-wilding@helsinki.fi Abstract We address the problem of cognate iden- tification across vocabularies of any pair of languages. In particular, we focus on the case where the examined languages are low-resource, to the extent that no train- ing data whatsoever in these languages, or even closely related ones, is available for the task. We investigate the extent to which training data from another, un- related language family can be used in- stead. Our approach consists of learning a similarity metric from example cognates in Indo-European languages and applying it to low-resource Sami languages of the Uralic family. We apply two models, fol- lowing previous work: a Siamese convolu- tional neural network (S-CNN) and a sup- port vector machine (SVM), and compare them with a Levenshtein distance base- line. We test performance on three Sami languages and find that the S-CNN out- performs the other approaches, suggesting that it is better able to learn such general characteristics of cognateness that carry over across language families. We also experiment with fine-tuning the S-CNN model with data from within the language family in order to quantify how well this model can make use of a small amount of target-domain data to adapt. 1 Introduction Cognate identification is a core task in the com- parative method, a collection of techniques used in historical linguistics for the inference of lan- guage family trees, reconstruction of protolan- guages, and other areas of study related to lan- guage history (List, 2013). Cognate informa- tion can also be used to improve natural language processing (NLP) applications, such as machine translation (Gr¨onroos et al., 2018). In addition, knowledge of cognates can be useful for second- language learning (Beinborn et al., 2014). For a subset of the world’s languages, such as Indo-European, language-family trees, pro- tolanguages, and etymological databases are well- established. However, the majority of languages have only few speakers, and such resources are scarce. Since the tasks of the comparative method are laborious to do manually, computational ap- proaches have been taken to automatize these tasks. In addition to cognate identification, pre- vious work addresses phonetic alignment (Kon- drak, 2000; Proki´c et al., 2009; List, 2013), in- ference of family trees (Chang et al., 2015; J¨ager, 2014; Bouckaert et al., 2012), and reconstruction of proto-words (Bouchard-Cˆot´e et al., 2013). Ideally, computational approaches to historical linguistics should be applicable to any language, even in the absence of hand-crafted resources and analyses. Recent work addressing cognate iden- tification for low-resource languages assumes ei- ther the existence of high-resource relatives, to be used as training data (McCoy and Frank, 2018), or the availability of detailed dictionary definitions (St Arnaud et al., 2017). In this paper, we address cognate identifica- tion in a scenario where we are only …

Content truncated to 500 words