Detecting Shibboleths

Proceedings of the EACL 2012 Joint Workshop of LINGVIS & UNCLH, pages 72–80, Avignon, France, April 23 - 24 2012. c⃝2012 Association for Computational Linguistics Detecting Shibboleths Jelena Proki´c Ludwig-Maximilians-Universit¨at j.prokic@lmu.de C¸ a˘grı C¸ ¨oltekin University of Groningen c.coltekin@rug.nl John Nerbonne University of Groningen j.nerbonne@rug.nl Abstract A SHIBBOLETH is a pronunciation, or, more generally, a variant of speech that betrays where a speaker is from (Judges 12:6). We propose a generalization of the well-known precision and recall scores to deal with the case of detecting distinctive, characteristic variants when the analysis is based on numerical difference scores. We also compare our proposal to Fisher’s linear discriminant, and we demonstrate its effec- tiveness on Dutch and German dialect data. It is a general method that can be applied both in synchronic and diachronic linguis- tics that involve automatic classification of linguistic entities. 1 Introduction and Background The background of this contribution is the line of work known as DIALECTOMETRY (S´eguy, 1973; Goebl, 1982), which has made computational work popular in dialectology. The basic idea of dialectometry is simple: one acquires large sam- ples of corresponding material (e.g., a list of lex- ical choices, such as the word for carbonated soft drink, which might be ‘soda’, ‘pop’, ‘tonic’ etc.) from different sites within a language area, and then, for each pair of samples, one counts (or more generally measures) the difference at each point of correspondence. The differences are summed, and, given representative and suffi- ciently large samples, the results characterizes the degree to which one site differs from another. Earlier work in dialectology mapped the dis- tributions of individual items, recording lines of division on maps, so-called ISOGLOSSES, and then sought bundles of these as tell-tale indica- tors of important divisions between DIALECT AR- EAS. But as Chambers & Trudgill (1998) note, the earlier methodology is fraught with prob- lems, many of which stem from the freedom of choice with respect to isoglosses, and their (nor- mal) failure to ‘bundle’ neatly. Nerbonne (2009) notes that dialectometry improves on the tradi- tional techniques in many ways, most of which stem from the fact that it shifts focus to AGGRE- GATE LEVEL of differences. Dialectometry uses large amounts of material; it reduces the sub- jectivity inherent in choosing isoglosses; it fre- quently analyzes material in ways unintended by those who designed dialect data collection efforts, including more sources of differences; and finally it replaces search for categorical overlap by a sta- tistical analysis of differences. Dialectometry does not enjoy overwhelming popularity in dialectology, however, and one of the reasons is simply that dialectologists, but also laymen, are interested not only in the aggregate relations among sites, or even the determination of dialect areas (or the structure of other geo- graphic influence on language variation, such as dialect continua), but are quite enamored of the details involved. Dialectology scholars, but also laymen, wish to now where ‘coffee’ is ordered (in English) with a labialized /k/ sound ([kwOfi]) or where in Germany one is likely …

Content truncated to 500 words