Identifying Cognates by Phonetic and Semantic Similarity

Identifying Cognates by Phonetic and Semantic Similarity Grzegorz Kondrak Department of Computer Science University of Toronto Toronto, Ontario, Canada M5S 3G4 kondrak@cs.toronto.edu Abstract I present a method of identifying cognates in the vo- cabularies of related languages. I show that a mea- sure of phonetic similarity based on multivalued fea- tures performs better than “orthographic” measures, such as the Longest Common Subsequence Ratio (LCSR) or Dice’s coefficient. I introduce a proce- dure for estimating semantic similarity of glosses that employs keyword selection and WordNet. Tests performed on vocabularies of four Algonquian lan- guages indicate that the method is capable of discov- ering on average nearly 75% percent of cognates at 50% precision. 1 Introduction In the narrow sense used in historical linguistics, cognates are words in related languages that have developed from the same ancestor word. An ex- ample of a cognate pair is French lait and Span- ish leche, both of which come from Latin lacte. In other contexts, including this paper, the term is often used more loosely, denoting words in different lan- guages that are similar in form and meaning, without making a distinction between borrowed and genet- ically related words; for example, English sprint and the Japanese borrowing supurinto are consid- ered cognate, even though these two languages are unrelated. In historical linguistics, the identification of cog- nates is a component of two principal tasks of the field: establishing the relatedness of languages and reconstructing the histories of language families. In corpus linguistics, cognates have been used for bi- text alignment (Simard et al., 1992; Church, 1993; McEnery and Oakes, 1996; Melamed, 1999), and for extracting lexicographically interesting word- pairs from multilingual corpora (Brew and Mc- Kelvie, 1996). The task addressed in this paper can be formu- lated in two ways. On the word level, given two words (lexemes) from different languages, the goal is to compute a value that reflects the likelihood of the pair being cognate. I assume that each lexeme is given in a phonetic notation, and that it is accom- panied by one or more glosses that specify its mean- ing in a metalanguage for which a lexical resource is available (for example, English). On the language level, given two vocabulary lists representing two languages, the goal is to single out all pairs that ap- pear to be cognate. Tables 1 and 2 show sample en- tries from two typical vocabulary lists. Such vocab- ulary lists are sometimes the only data available for lesser-studied languages. In general, deciding whether two words are ge- netically related requires expert knowledge of the history of the languages in question. With time, words in all languages change their form and mean- ing. After several millennia, cognates often acquire very different phonetic shapes. For example, En- glish hundred, French cent, and Polish sto are all descendants of Proto-Indo-European *kmtom. The semantic change can be no less dramatic; for ex- ample, English guest and Latin hostis ‘enemy’ are cognates even though their meanings are diametri- cally different. On …

Content truncated to 500 words