학습자 코퍼스를 이용한 영어 전치사 오류 교정 모델 개발

Abstract

With growing demands for computerized tools in ESL (English as a Second Language) and EFL (English as a Foreign Language) classrooms, applying latest advancement in natural language processing to developing models for diagnosing and correcting errors in learner language poses an interesting research question which touches on issues of diverse nature: engineering- oriented, theoretical and also practical. In this study, we present a method of statistically modeling preposition usage errors by training a classifier ex- clusively on an error-annotated corpus of L2 essays. The data set, Chungdahm English Learner Corpus, is a large-scale corpus containing over 130 million words and over 860,000 individual essays, written by middle school students whose native language is Korean. We train a maximum entropy classifier on the preposition instances in the corpus based on a small number of simplistic contextual features and report a good level of performance at over 90% precision and 29% recall in identifying and error and suggesting a grammatical alternative. In comparison with the more widely practiced method of building language correction models based on well-formed texts produced by native users of the language, the approach presented in this study invites some interesting theoretical and empirical considerations, namely the nature of the resultant model as one of a particular sub-language, the English of Korean middle students in this case, and also its extendability to other variations of the English language.