Collaborative web UI localization, or how to build feature-rich multilingual datasets

Collaborative Web UI Localization, or How to Build Feature-rich Multilingual Datasets Vicent Alabau PRHLT Research Center Universitat Polit`ecnica de Val`encia valabau@prhlt.upv.es Luis A. Leiva PRHLT Research Center Universitat Polit`ecnica de Val`encia luileito@prhlt.upv.es Abstract We present a method to generate feature- rich multilingual parallel datasets for ma- chine translation systems, including e.g. type of widget, user’s locale, or geoloca- tion. To support this argument, we have developed a bookmarklet that instruments arbitrary websites so that casual end users can modify their texts on demand. After surveying 52 users, we conclude that peo- ple is leaned toward using this method in lieu of other comparable alternatives. We validate our prototype in a controlled study with 10 users, showing that language re- sources can be easily generated. 1 Introduction Today most websites are looking forward to mak- ing their contents available in more than one lan- guage, mainly to reach a global audience, to gain a competitive advantage, or just because of legal requirements. To this end, adapting user interface (UI) texts through translation—or “localization”— is a central task, since its result affects system us- ability and acceptability. Actually, translation is just one of the activities of localization yet the most important overall (Keniston, 1997). Recently there have been significant improve- ments in machine translation (MT) technology, to the extent that, in particular contexts such as medical prescriptions or knowledge-base articles, machine-translated content is qualitatively compa- rable to that of human-translated (Dillinger and Laurie, 2009). However, for MT systems to excel at UI localization not only it is needed an impor- tant amount of training data, but also the data must be especially tailored to the particularities of UI messages. Indeed, translating the text in an inter- c⃝ 2014 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. face is a challenging task, even for trained human translators (Munt´es-Mulero et al., 2012). Parallel data offer a rich source of additional knowledge about language, and a sound basis for both translation and contrastive studies (McEnery and Xiao, 2007). Although there are some valu- able tools to build multilingual parallel corpora, they are still limited when it comes to the exploita- tion of UI-based resources. Thus, we propose a novel approach: delegating the corpus generation to the end users of software applications, as a re- sult of a regular interaction with such applications. To support our approach, we developed a proof- of-concept web-based prototype, motivated by the fact that nowadays people use web browsers more than any other class of software. Moreover, soft- ware translation poses two interesting challenges: 1) user interface (UI) strings appear anywhere in the developer’s language of choice whereas con- tent is typically generated and consumed in the user’s language; 2) UI bilingual sentences can be enriched with metadata to handle disambiguation. 2 Related Work In the past, several methods have been developed to build parallel corpora by automatic means, e.g., by mining Wikipedia (Smith et al., 2010), web pages with a …

Content truncated to 500 words