Detecting and Correcting Learner Korean Particle Omission Errors

International Joint Conference on Natural Language Processing, pages 1419–1427, Nagoya, Japan, 14-18 October 2013. Detecting and Correcting Learner Korean Particle Omission Errors Ross Israel Indiana University raisrael@indiana.edu Markus Dickinson Indiana University md7@indiana.edu Sun-Hee Lee Wellesley College slee6@wellesley.edu Abstract We detect errors in Korean post-positional particle usage, focusing on optimizing omission detection, as omissions are the single-biggest factor in particle errors for learners of Korean. We also develop a sys- tem for predicting the correct choice of a particle. For omission detection, we model the task largely on English grammatical er- ror detection, but employ Korean-specific features and filters; likewise, output analy- sis and the omission correction system il- lustrate how unique properties of Korean, such as the distinct types of particles used, need to be accounted for in adapting the system, thereby moving the field one step closer to robust multi-lingual methods. 1 Introduction Grammatical error detection is useful to produce an improved final document for writing assistance, provide feedback to language learners, provide features for automatic essay scoring, and post- edit machine translation output (see references in Chodorow et al., 2012, sec. 2). Within this grow- ing field, most of the work has focused on En- glish, but there has been a small community of researchers working on other languages. We con- tinue this trend by advancing the state-of-the-art in detecting errors in Korean particle usage. Expanding to other languages and language families obviously presents new challenges, such as being able to handle word segmentation and greater morphological complexity (e.g., Basque (de Ilarraza et al., 2008), Korean (Lee et al., 2012), Hungarian (Dickinson and Ledbetter, 2012), Japanese (Mizumoto et al., 2011)); greater vari- eties of word order (Czech (Hana et al., 2010), German (Boyd, 2012)); case ending errors (Czech, German, Hungarian); differing definitions of func- tion words (Korean, Japanese, Basque); and so forth. Investing in methods which apply across languages will make techniques more robust and applicable for even more languages. An additional challenge for many of these lan- guages is the lack of resources. Much previous work on detecting errors in Korean, for example, focused less on techniques and more on acquiring training data (Dickinson et al., 2011) and evalu- ation data (Lee et al., 2012). We thus desire tech- niques that work using smaller and/or unannotated data sets that may be less reliable than some of the corpora for better-resourced languages. We focus on detecting errors in the presence or absence of Korean postpositional particles. Korean is a Less Commonly Taught Language (LCTL) needing proficient speakers and more pedagogical research (see Dickinson et al., 2008, sec. 2), making computational tools for Korean language learning important. Particles are used to mark properties akin to prepositions and also to case markers, as discussed in section 2. This makes our task applicable to similar languages like Japanese and more generally to agglutinative lan- guages like Basque, Hungarian, and Turkish, as discussed in section 3 on related work. Particles are our focus because of the high prevalence of particle errors in …

Content truncated to 500 words