Similarity Patterns in Words (Invited talk)

Proceedings of the EACL 2012 Joint Workshop of LINGVIS & UNCLH, pages 49–53, Avignon, France, April 23 - 24 2012. c⃝2012 Association for Computational Linguistics Similarity Patterns in Words Grzegorz Kondrak Department of Computing Science University of Alberta Edmonton, Alberta, Canada, T6G 2E8 gkondrak@ualberta.ca Abstract Words are important both in historical lin- guistics and natural language processing. They are not indivisible abstract atoms; much can be gained by considering smaller units such as morphemes, phonemes, syl- lables, and letters. In this presentation, I attempt to sketch the similarity pat- terns among a number of diverse research projects in which I participated. 1 Introduction Languages are made up of words, which con- tinuously change their form and meaning. Lan- guages that are related contain cognates — re- flexes of proto-words that survive in some form in the daughter languages. Sets of cognates reg- ularly exhibit recurrent sound correspondences. Together, cognates and recurrent sound corre- spondences provide evidence of a common origin of languages. Although I consider myself more a computer scientist than a linguist, I am deeply interested in words. Even though many NLP algorithms treat words as indivisible abstract atoms, I think that much can be gained by considering smaller units: morphemes, phonemes, syllables, and let- ters. Words that are similar at the sub-word level often exhibit similarities on the syntactic and se- mantic level as well. Even more important, as we move beyond written text towards speech and pro- nunciation, the make-up of words cannot be ig- nored anymore. I commenced my NLP research by investigat- ing ways of developing computer programs for various stages of the language reconstruction pro- cess (Kondrak, 2002a). From the very start, I aimed at proposing language-independent solu- tions grounded in the current advances in NLP, bioinformatics, and computer science in general. The algorithms were evaluated on authentic lin- guistic data and compared quantitatively to pre- vious proposals. The projects directly related to language histories still form an important part of my research. In Section 2, I refer to several of my publications on the subject, while in Section 3, I focus on other NLP applications contributions that originate from my research on diachronic lin- guistics. 2 Diachronic NLP The comparative method is the technique applied by linguists for reconstructing proto-languages. It consists of several stages, which include the iden- tification of cognates by semantic and phonetic similarity, the alignment of cognates, the deter- mination of recurrent sound correspondences, and finally the reconstruction of the proto-forms. The results of later steps are used to refine the judg- ments made in earlier ones. The comparative method is not an algorithm, but rather a collection of heuristics, which involve intuitive criteria and broad domain knowledge. As such, it is a very time-consuming process that has yet to be accom- plished for many language families. Since the comparative method involves detec- tion of regularities in large amounts of data, it is natural to investigate whether it can be performed by a computer program. In this section, I …

Content truncated to 500 words