SW4ALL: a CEFR Classified and Aligned Corpus for Language Learning

SW4ALL: a CEFR-Classified and Aligned Corpus for Language Learning Rodrigo Wilkens, Leonardo Zilio and C´edrick Fairon Centre de Traitement Automatique du Langage - CENTAL Universit´e catholique de Louvain - UCL {rodrigo.wilkens, leonardo.zilio, cedrick.fairon}@uclouvain.be Abstract Learning a second language is a task that requires a good amount of time and dedication. Part of the process involves the reading and writing of texts in the target language, and so, to facilitate this process, especially in terms of reading, teachers tend to search for texts that are associated to the interests and capabilities of the learners. But the search for this kind of text is also a time-consuming task. By focusing on this need for texts that are suited for different language learners, we present in this study the SW4ALL, a corpus with documents classified by language proficiency level (based on the CEFR recommendations) that allows the learner to observe ways of describing the same topic or content by using strategies from different proficiency levels. This corpus uses the alignments between the English Wikipedia and the Simple English Wikipedia for ensuring the use of similar content or topic in pairs of text, and an annotation of language levels for ensuring the difference of language proficiency level between them. Considering the size of the corpus, we used an automatic approach for the annotation, followed by an analysis to sort out annotation errors. SW4ALL contains 8,669 pairs of documents that present different levels of language proficiency. Keywords: SLA, CEFR classification, Aligned corpus 1. Introduction Learning a second language is a process that requires expo- sition to texts, especially for the acquisition of vocabulary (Rott, 1999). To retrieve texts that match learners’ language level (or proficiency) it is possible to use a corpus care- fully designed for language learners, or one can search for them in the web. Following these alternatives, systems can dig up texts aiming at finding those that are best suited to the language skills of a learner. Examples of these systems are REAP (Heilman et al., 2008), FLAIR (Chinkina et al., 2016), and READ-X (Miltsakaki and Troutt, 2007), which use the web as a corpus. The use of web allow the learn- ers to interact with a huge diversity of texts, which makes it easier to find those that correspond to their interests. But most of the web texts retrieved by search engines require a high language proficiency, even for native speakers (Vaj- jala and Meurers, 2013). On the other hand, the use of an off-line corpus ensures content quality, while hindering the search for different text topics. This dichotomy of text sources enforces different types of restrictions on the systems. An alternative for trying to get the best of both approaches is to use the web as source of texts, but restricting it to trustful domains. This is similar to SourceFinder’s method, in which on-line newspapers and magazines are downloaded and processed as text sources (Passonneau et al., 2002; Sheehan et al., 2007). However, this kind of approach doesn’t allow for …

Content truncated to 500 words