Crowdsourcing Regional Variation Data and Automatic Geolocalisation of Speakers of European French

Crowdsourcing Regional Variation Data and Automatic Geolocalisation of Speakers of European French Jean-Philippe Goldman1, Yves Scherrer2, Julie Glikman3, Mathieu Avanzi4, Christophe Benzitoun5, Philippe Boula de Mareüil6 1 LATL-CUI University of Geneva, 2 University of Helsinki, 3 LILPA University of Strasbourg, 4 FNRS/Université catholique de Louvain, 5 ATILF Université de Lorraine, 6 LIMSI-CNRS jeanphilippegoldman@gmail.com Abstract We present the crowdsourcing platform Donnez Votre Français à la Science (DFS, or “Give your French to Science”), which aims to collect linguistic data and document language use, with a special focus on regional variation in European French. The activities not only gather data that is useful for scientific studies, but they also provide feedback to the general public; this is important in order to reward participants, to encourage them to follow future surveys, and to foster interaction with the scientific community. The two main activities described here are 1) a linguistic survey on lexical variation with immediate feedback and 2) a speaker geolocalisation system; i.e., a quiz that guesses the linguistic origin of the participant by comparing their answers with previously gathered linguistic data. For the geolocalisation activity, we set up a simulation framework to optimise predictions. Three classification algorithms are compared: the first one uses clustering and shibboleth detection, whereas the other two rely on feature elimination techniques with Support Vector Machines and Maximum Entropy models as underlying base classifiers. The best-performing system uses a selection of 17 questions and reaches a localisation accuracy of 66%, extending the prediction from the one-best area (one among 109 base areas) to its first-order and second-order neighbouring areas. Keywords: language variation, regionalism, crowdsourcing, geolocalisation, linguistic geography, cartography 1. Introduction Linguistic surveys and experiments can now easily reach thousands of people through the internet and smartphones. This ‘crowdsourcing’ methodology allows researchers to collect linguistic resources on a large scale (Cook et al. 2013), helping to describe linguistic phenomena (regional variation, non-normative forms, among others) that are underrepresented in traditional corpora. As a quick and easily reproducible way of collecting data, one can even conceive of using crowdsourcing over several years in order to compare successive snapshots of linguistic variables and thus describe variation in time. In recent years, various crowdsourcing projects have been set up for collecting linguistic data through web applications, such as Français de nos régions (Avanzi et al. 2016) for European and Canadian French, Atlas der deutschen Alltagssprache (Möller & Elspaß 2015) for regional varieties of German, the Harvard Dialect Survey (Vaux & Bert 2013) for regional variation in the USA, VerbaAlpina (Krefeld and Lücke 2014) for dialects spoken in the Alps, and more recently, the twin websites tonaccent/dindialäkt focusing on the perception of Swiss French accents and Swiss German dialects (Goldman et al. 2018). Other applications have been explicitly framed as geolocalisation games, aimed at predicting the provenance of the user on the basis of their naming of a set of objects. Whereas some projects have been distributed as smartphone apps, e.g. to document English and German dialectal variation (Dialäkt Äpp, Leemann et al. …

Content truncated to 500 words