Toward a Lightweight Solution for Less-resourced Languages: Creating a POS Tagger for Alsatian Using Voluntary Crowdsourcing

Toward a Lightweight Solution for Less-resourced Languages: Creating a POS Tagger for Alsatian Using Voluntary Crowdsourcing Alice Millour, Kar¨en Fort STIH/Universit´e Paris-Sorbonne, France {alice.millour, karen.fort}@paris-sorbonne.fr Abstract We present here the results of an experiment aiming at crowdsourcing part-of-speech annotations for a less-resourced French regional language, Alsatian. We used for this purpose a specifically-developed slightly gamified platform, Bisame. It allowed us to gather annotations on a variety of corpora covering some of the language dialectal variations. The quality of the annotations, which reach an averaged F-measure of 93%, enabled us to train a first tagger for Alsatian that is nearly 84% accurate. The platform as well as the produced annotations and tagger are all freely available. The platform can easily be adapted to other languages, thus providing a solution to (some of) the less-resourced languages issue. Keywords: crowdsourcing, less-resourced languages, POS tagging, Alsatian 1. Introduction Despite the progress made in unsupervised learning, man- ually annotated corpora are still necessary both to develop and to evaluate natural language processing (NLP) tools. However, building such corpora is notoriously expensive (see, for example (B¨ohmov´a et al., 2001)). For less- resourced languages, the (lack of) availability of language experts represents yet another obstacle to overcome. We hypothesized that contributing to the creation of NLP tools for their language would be an incentive for non- expert speakers, especially for languages known for their speakers’ activism, to participate in such a project. We therefore developed a lightweight crowdsourcing platform, Bisame1, which enables participants to collaboratively produce part-of-speech (POS) annotations. We tested it on a French regional language with the appropriate specifici- ties, i.e. activism and easy access to the Internet: Alsatian. We present here the related work performed on Alsatian, POS tagging less-resourced languages and crowdsourcing linguistic annotations. We then describe the methodology we used and the results we obtained. Finally, we discuss the limits of this work. 2. Related Work 2.1. Alsatian Alsatian is a generic term for the continuum of Germanic dialects spoken in Alsace and part of Moselle, two diglossic regions where French and Alsatian dialects coexist.2 Be- yond this generic glotonym lies a set of dialectal subsys- tems (Malherbe, 1983). In fact, low Alemannic, the main variant of Alsatian, can itself be divided into two subsets: Northern (NV) and Southern (SV) low Alemannic. Besides, Strasbourg presents a slightly different variant of North- ern low Alemannic (STRV), tinted with Franconian, as evi- denced by the use of Dienschtd`aj for “Tuesday”, instead of the Alemannic Zischdig. Thus, the 7 (phonetically differ- ent) forms Baum, B`a`am, Bœm (SV); Bauim, B`auim, B¨aum 1See: http://bisame.paris-sorbonne.fr. 2In spite of the decrease of the family linguistic transmission, a study registered 550,000 Alsatian speakers (Barre and Vander- schelden, 2004). (NV); Baam (STRV) can be found for the word “tree”. Ad- ditionally to the dialectal variation and although the Orthal spelling guidelines (Cr´evenat-Werner and Zeidler, 2008) have existed since 2006, no orthographic convention is rec- ognized yet as the legitimate standard among the speakers. Consequently, elided forms such as m’r for mir …

Content truncated to 500 words