Model Combination for Correcting Preposition Selection Errors

Many grammatical error correction approaches use classifiers with specially-engineered features to predict corrections. A simpler alternative is to use n-gram language model scores. Rozovskaya and Roth (2011) reported that classifiers outperformed a language modeling approach. Here, we report a more nuanced result: a classifier approach yielded results with higher precision while a language modeling approach provided better recall. Most importantly, we found that a combined approach using a logistic regression ensemble outperformed both a classifier and a language modeling approach.