Precision Isn't Everything: A Hybrid Approach to Grammatical Error Detection

Some grammatical error detection methods, including the ones currently used by the Educational Testing Service’s e-rater system (Attali and Burstein, 2006), are tuned for precision because of the perceived high cost of false positives (i.e., marking fluent English as ungrammatical). Precision, however, is not optimal for all tasks, particularly the HOO 2012 Shared Task on grammatical errors, which uses F-score for evaluation. In this paper, we extend e-rater’s preposition and determiner error detection modules with a largescale n-gram method (Bergsma et al., 2009) that complements the existing rule-based and classifier-based methods. On the HOO 2012 Shared Task, the hybrid method performed better than its component methods in terms of F-score, and it was competitive with submissions from other HOO 2012 participants.