Using context and phonetic features in models of etymological sound change

Proceedings of the EACL 2012 Joint Workshop of LINGVIS & UNCLH, pages 108–116, Avignon, France, April 23 - 24 2012. c⃝2012 Association for Computational Linguistics Using context and phonetic features in models of etymological sound change Hannes Wettig1, Kirill Reshetnikov2 and Roman Yangarber1 1Department of Computer Science 2Institute of Linguistics University of Helsinki, Finland Academy of Sciences First.Last@cs.helsinki.fi Moscow, Russia Abstract This paper presents a novel method for aligning etymological data, which mod- els context-sensitive rules governing sound change, and utilizes phonetic features of the sounds. The goal is, for a given corpus of cognate sets, to find the best alignment at the sound level. We introduce an imputa- tion procedure to compare the goodness of the resulting models, as well as the good- ness of the data sets. We present evalu- ations to demonstrate that the new model yields improvements in performance, com- pared to previously reported models. 1 Introduction This paper introduces a context-sensitive model for alignment and analysis of etymological data. Given a raw collection of etymological data (the corpus)—we first aim to find the “best” alignment at the sound or symbol level. We take the corpus (or possibly several different corpora) for a lan- guage family as given; different data sets are typ- ically conflicting, which creates the need to deter- mine which is more correct. Etymological data sets are found in digital etymological databases, such as ones we use for the Uralic language fam- ily. A database is typically organized into cog- nate sets; all elements within a cognate set are posited (by the database creators) to be derived from a common origin, which is a word-form in the ancestral proto-language. Etymology encompasses several problems, including: discovery of sets of cognates— genetically related words; determination of ge- netic relations among groups of languages, based on linguistic data; discovering regular sound cor- respondences across languages in a given lan- guage family; and reconstruction of forms in the proto-languages. Computational methods can provide valuable tools for the etymological community. The meth- ods can be judged by how well they model certain aspects of etymology, and by whether the auto- matic analysis produces results that match theo- ries established by manual analysis. In this work, we allow all the data—and only the data—to determine what rules underly it, rather than relying on external (and possibly bi- ased) rules that try to explain the data. This ap- proach will provide a means of measuring the quality of the etymological data sets in terms of their internal consistency—a dataset that is more consistent should receive a higher score. We seek methods that analyze the data automatically, in an unsupervised fashion, to determine whether a complete description of the correspondences can be discovered automatically, directly from raw etymological data—cognate sets within the lan- guage family. Another way to state the question is: what alignment rules are “inherently encoded” in the given corpus itself. At present, our aim is to analyze given etymo- logical datasets, rather than to construct new ones from …

Content truncated to 500 words