Russian natural language processing for computer-assisted language learning: Capturing the benefits of deep morphological analysis in real-life applications

Abstract

In this dissertation, I investigate practical and theoretical issues surrounding the use of natural language processing technology in the context of Russian Computer-Assisted Language-Learning, with particular emphasis on morphological analysis.

In Part I, I present linguistic and practical issues surrounding the development and evaluation of two foundational technologies: a two-level morphological analyzer, and a constraint grammar to contextually disambiguate homonymy in the analyzer’s output. The analyzer was specially designed for L2 learner applications—with stress annotation and rule-based morphosyntactic disambiguation—and it is competitive with state-of-the-art Russian analyzers. The constraint grammar is designed to have high recall, allowing an L2-learner application to base decisions on all possible readings, and not just the single most likely reading. The constraint grammar resolves 44% of the ambiguity output by the morphological analyzer. A voting setup combining the constraint grammar with a trigram hidden markov model tagger demonstrates how a high-recall grammar can boost performance of probabilistic taggers, which are better suited to capturing highly idiosyncratic facts about collocational tendencies.

In Part II, I present linguistic, theoretical, practical issues surrounding the application of the morphological analyzer and constraint grammar to three real-life computer-assisted language-learning tasks: automatic stress annotation, automatic grammar exercise generation from authentic texts, and automatic evaluation of text readability. The automatic stress placement task is vital for Russian language-learning applications. The morphological analyzer and constraint grammar yield state-of-the-art results, resolving 42% of stress ambiguity in a corpus of running text.

In order to demonstrate the value of a high-recall constraint grammar, I developed Russian grammar activities for the VIEW platform, a system for providing automatic Visual Input Enhancement of Web documents. This system allows teachers and learners to automatically generate grammatical highlighting, identification activities, multiple-choice activities, and fill-in-the-blank activities, enabling them to study grammar using texts that are interesting or relevant to them. I show that the morphological analysis described above is instrumental not only for generating exercises, but also for providing adaptive feedback, a feature which typically requires encoding specific learner language features.

A final test-case for morphological analysis in Russian language-learning is automatic readability assessment, which can help learners and teachers find texts at appropriate reading levels. I show that features based on morphology are among the most informative for L2 readability assessment.