Innovation in Language Learning

Recent advances in natural language processing (NLP) have made it possible to implement automatic text analysis in foreign language learning applications. The use of NLP has been used primarily in evaluating learner errors, but it has also been used to enhance authentic texts for learners (Authentic Text Intelligent Computer-Assisted Language Learning, i.e. ATICALL). For instance, the WERTi/VIEW platform makes it possible to convert text from virtually any webpage into focus-on-form activities, including textual highlighting, target identification, multiple choice, and fill-in-the-blank. In this presentation, I report on recent progress in developing WERTi/VIEW exercises for Russian.

Previous work in ATICALL has been focused on languages with relatively little inflectional morphology, e.g. English, Spanish and German. Since Russian has more extensive inflectional morphology, it poses new challenges and opportunities for ATICALL textual enhancement.

Since ATICALL activities are dependent on the density of target structures, I use frequency data to justify the creation of particular exercises. For example, grammatical forms that are uncommon in written texts, such as vocative case (e.g. Saš ‘Sasha.VOC’) or 1st person plural imperatives (e.g. pojdemte! ‘let’s go!’), are not good candidates for ATICALL activities.

Those target grammatical forms that were selected are so frequent that one of the primary tasks of activity generation was to avoid overwhelming the learner with too many exercises per document. I discuss several criteria by which less optimal tokens are excluded. First, tokens whose context in less readable (more complex structures, or unknown vocabulary) are excluded. Second, some tokens are rated as more likely to lead to uptake according to research in second language acquisition, i.e. operationalizations of VanPatten’s principles of input processing. Third, some criteria are simply arbitrary assertions, based on teacher intuitions or computational limitations. For example, it was judged that more than one exercise within the same sentence could be confusing. Also, if the computer is unable to output a well-formed, unambiguous analysis of a sentence, then any target tokens contained therein are excluded in order to avoid potential mistakes. Future plans for integration of student models is also briefly discussed.