Informing Determiner and Preposition Error Correction with Hierarchical Word Clustering

We extend our n-gram-based data-driven prediction approach from the Helping Our Own (HOO) 2011 Shared Task (Boyd and Meurers, 2011) to identify determiner and preposition errors in non-native English essays from the Cambridge Learner Corpus FCE Dataset (Yannakoudakis et al., 2011) as part of the HOO 2012 Shared Task. Our system focuses on three error categories: missing determiner, incorrect determiner, and incorrect preposition. Approximately two-thirds of the errors annotated in HOO 2012 training and test data fall into these three categories. To improve our approach, we developed a missing determiner detector and incorporated word clustering (Brown et al., 1992) into the n-gram prediction approach.