Identifying Cognate Sets Across Dictionaries of Related Languages

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2519–2528 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Identifying Cognate Sets Across Dictionaries of Related Languages Adam St Arnaud Dept of Computing Science University of Alberta ajstarna@ualberta.ca David Beck Dept of Linguistics University of Alberta dbeck@ualberta.ca Grzegorz Kondrak Dept of Computing Science University of Alberta gkondrak@ualberta.ca Abstract We present a system for identifying cog- nate sets across dictionaries of related lan- guages. The likelihood of a cognate re- lationship is calculated on the basis of a rich set of features that capture both pho- netic and semantic similarity, as well as the presence of regular sound correspon- dences. The similarity scores are used to cluster words from different languages that may originate from a common proto- word. When tested on the Algonquian lan- guage family, our system detects 63% of cognate sets while maintaining cluster pu- rity of 70%. 1 Introduction Cognates are words in related languages that have originated from the same word in an ancestor lan- guage; for example English earth and German Erde. On average, cognates display higher pho- netic and semantic similarity than random word pairs between languages that are indisputably re- lated (Kondrak, 2013). The term cognate is some- times used within computational linguistics to de- note orthographically similar words that have the same meaning (Nakov and Tiedemann, 2012). In this work, however, we adhere to the strict linguis- tic definition of cognates and aim to distinguish them from lexical borrowings by detecting regular sound correspondences. Cognate information between languages is crit- ical to the field of historical and comparative lin- guistics, where it plays a central role in determin- ing the relations and structures of language fami- lies (Trask, 1996). Automated phylogenetic recon- structions often rely on cognate information as in- put (Bouchard-Cˆot´e et al., 2013). The percentage of shared cognates can also be used to estimate the time of pre-historic language splits (Dyen et al., 1992). While cognates are valuable to linguists, their identification is a time-consuming process, even for experts, who have to sift through hun- dreds or even thousands of words in related lan- guages. The languages that are the least well stud- ied, and therefore the ones in which historical lin- guists are most interested, often lack cognate in- formation. A number of computational methods have been proposed to automate the process of cognate iden- tification. Many of the systems focus on iden- tifying cognates within classes of semantically equivalent words, such as Swadesh lists of basic concepts. Those systems, which typically con- sider only the phonetic or orthographic forms of words, can be further divided into the ones that operate on language pairs (pairwise) vs. multilin- gual approaches. However, because of seman- tic drift, many cognates are no longer exact syn- onyms, which severely limits the effectiveness of such systems. For example, a cognate pair like En- glish bite and French fendre “to split” cannot be detected because these words are listed under …

Content truncated to 500 words