MDL-based Models for Alignment of Etymological Data

Proceedings of Recent Advances in Natural Language Processing, pages 111–117, Hissar, Bulgaria, 12-14 September 2011. MDL-based Models for Alignment of Etymological Data Hannes Wettig, Suvi Hiltunen, Roman Yangarber Department of Computer Science University of Helsinki, Finland First.Last@cs.helsinki.fi Abstract We introduce several models for alignment of etymological data, that is, for finding the best alignment, given a set of etymological data, at the sound or symbol level. This is intended to obtain a means of measuring the quality of the etymological data sets, in terms of their in- ternal consistency. One of our main goals is to devise automatic methods for aligning the data that are as objective as possible, the models make no a priori assumptions—e.g., no prefer- ence for vowel-vowel or consonant-consonant alignments. We present a baseline model and several successive improvements, using data from the Uralic language family. 1 Introduction We present work on induction of alignment rules for etymological data, in a project that studies genetic re- lationships among the Uralic language family. This is a continuation of previous work, reported in (Wettig and Yangarber, 2011), where the methods were intro- duced. In this paper, we extend the models reported earlier and give a more comprehensive evaluation of results. In addition to the attempt to induce alignment rules, we aim to derive measures of quality of data sets in terms of their internal consistency. More consis- tent dataset should receive a higher score in the evalu- ations. Currently our goal is to analyze given, existing etymological datasets, rather than to construct cognate sets from raw linguistic data. The question to be an- swered is whether a complete description of the corre- spondence rules can be discovered automatically. Can they be found directly from raw etymological data— sets of cognate words from languages within the lan- guage family? Are the alignment rules are “inherently encoded” in a dataset (the corpus) itself? We aim to develop methods that are as objective as possible, that rely only on the data, rather than on any prior assump- tions about the data, the possible rules and alignments. Computational etymology encompasses several problem areas, including: discovery of sets of genet- ically related words—cognates; determination of ge- netic relations among groups of languages, from raw or organized linguistic data; discovering regular sound correspondences across languages in a given language family; and reconstruction, either diachronic—i.e., re- construction of proto-forms for a hypothetical par- ent language, from which the word-forms found in the daughter languages derive, or synchronic—i.e., of word forms that are missing from existing languages. Several approaches to etymological alignment have emerged over the last decade. The problem of discov- ering cognates is addressed, e.g., in, e.g., (Bouchard- Cˆot´e et al., 2007; Kondrak, 2004; Kessler, 2001). In our work, we do not attempt to find cognate sets, but begin with given sets of etymological data for a lan- guage family, possibly different or even conflicting. We use the principle of recurrent sound correspon- dence, as in much of the literature, including the men- tioned …

Content truncated to 500 words