Correction Detection and Error Type Selection as an ESL Educational Aid

2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 357–361, Montr´eal, Canada, June 3-8, 2012. c⃝2012 Association for Computational Linguistics Correction Detection and Error Type Selection as an ESL Educational Aid Ben Swanson Brown University chonger@cs.brown.edu Elif Yamangil Harvard University elif@eecs.harvard.edu Abstract We present a classifier that discriminates be- tween types of corrections made by teachers of English in student essays. We define a set of linguistically motivated feature templates for a log-linear classification model, train this classifier on sentence pairs extracted from the Cambridge Learner Corpus, and achieve 89% accuracy improving upon a 33% base- line. Furthermore, we incorporate our classi- fier into a novel application that takes as input a set of corrected essays that have been sen- tence aligned with their originals and outputs the individual corrections classified by error type. We report the F-Score of our implemen- tation on this task. 1 Introduction In a typical foreign language education classroom setting, teachers are presented with student essays that are often fraught with errors. These errors can be grammatical, semantic, stylistic, simple spelling errors, etc. One task of the teacher is to isolate these errors and provide feedback to the student with cor- rections. In this body of work, we address the pos- sibility of augmenting this process with NLP tools and techniques, in the spirit of Computer Assisted Language Learning (CALL). We propose a step-wise approach in which a teacher first corrects an essay and then a computer program aligns their output with the original text and separates and classifies independent edits. With the program’s analysis the teacher would be provided accurate information that could be used in effective lesson planning tailored to the students’ strengths and weaknesses. This suggests a novel NLP task with two compo- nents: The first isolates individual corrections made by the teacher, and the second classifies these cor- rections into error types that the teacher would find useful. A suitable corpus for developing this pro- gram is the Cambridge Learner Corpus (CLC) (Yan- nakoudakis et al., 2011). The CLC contains approxi- mately 1200 essays with error corrections annotated in XML within sentences. Furthermore, these cor- rections are tagged with linguistically motivated er- ror type codes. To the best of our knowledge our proposed task is unexplored in previous work. However, there is a significant amount of related work in automated grammatical error correction (Fitzgerald et al., 2009; Gamon, 2011; West et al., 2011). The Helping Our Own (HOO) shared task (Dale and Kilgarriff, 2010) also explores this issue, with Rozovskaya et al. (2011) as the best performing system to date. While often addressing the problem of error type selection directly, previous work has dealt with the more obviously useful task of end to end error detec- tion and correction. As such, their classification sys- tems are crippled by poor recall of errors as well as the lack of information from the corrected sentence and yield very low accuracies for error detection and type …

Content truncated to 500 words