Generating Learner-Like Morphological Errors in Russian

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 259–267, Beijing, August 2010 Generating Learner-Like Morphological Errors in Russian Markus Dickinson Indiana University md7@indiana.edu Abstract To speed up the process of categorizing learner errors and obtaining data for lan- guages which lack error-annotated data, we describe a linguistically-informed method for generating learner-like mor- phological errors, focusing on Russian. We outline a procedure to select likely er- rors, relying on guiding stem and suffix combinations from a segmented lexicon to match particular error categories and rely- ing on grammatical information from the original context. 1 Introduction Work on detecting grammatical errors in the lan- guage of non-native speakers covers a range of errors, but it has largely focused on syntax in a small number of languages (e.g., Vandeven- ter Faltin, 2003; Tetreault and Chodorow, 2008). In more morphologically-rich languages, learn- ers naturally make many errors in morphology (Dickinson and Herring, 2008). Yet for many lan- guages, there is a major bottleneck in system de- velopment: there are not enough error-annotated learner corpora which can be mined to discover the nature of learner errors, let alone enough data to train or evaluate a system. Our perspective is that one can speed up the process of determin- ing the nature of learner errors via semi-automatic means, by generating plausible errors. We set out to generate linguistically-plausible morphological errors for Russian, a language with rich inflections. Generating learner-like errors has practical and theoretical benefits. First, there is the issue of obtaining training data; as Foster and Andersen (2009) state, “The ideal situation for a grammatical error detection system is one where a large amount of labelled positive and negative ev- idence is available.” Generated errors can bridge this gap by creating realistic negative evidence (see also Rozovskaya and Roth, 2010). As for evaluation data, generated errors have at least one advantage over real errors, in that we know pre- cisely what the correct form is supposed to be, a problem for real learner data (e.g., Boyd, 2010). By starting with a coarse error taxonomy, gen- erating errors can improve categorization. Gener- ated errors provide data for an expert—e.g., a lan- guage teacher—to search through, expanding the taxonomy with new error types or subtypes and/or deprecating error types which are unlikely. Given the lack of real learner data, this has the potential to speed up error categorization and subsequent system development. Furthermore, error genera- tion techniques can be re-used, adjusting the er- rors for different learner levels, first languages, and so forth. The error generation process can benefit by us- ing linguistic properties to mimic learner varia- tions. This can lead to more realistic errors, a ben- efit for machine learning (Foster and Andersen, 2009), and can also provide feedback for the lin- guistic representation used to generate errors by, e.g., demonstrating under which linguistic condi- tions certain error types are generated and under which they are not. We are specifically interested in generating Russian morphological errors. To do this, we need …

Content truncated to 500 words