Using Mostly Native Data to Correct Errors in Learners’ Writing

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 163–171, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Using Mostly Native Data to Correct Errors in Learners' Writing: A Meta-Classifier Approach Michael Gamon Microsoft Research One Microsoft Way Redmond, WA 98052 mgamon@microsoft.com Abstract We present results from a range of experi- ments on article and preposition error correc- tion for non-native speakers of English. We first compare a language model and error- specific classifiers (all trained on large Eng- lish corpora) with respect to their performance in error detection and correction. We then combine the language model and the classifi- ers in a meta-classification approach by com- bining evidence from the classifiers and the language model as input features to the meta- classifier. The meta-classifier in turn is trained on error-annotated learner data, optimizing the error detection and correction performance on this domain. The meta-classification approach results in substantial gains over the classifier- only and language-model-only scenario. Since the meta-classifier requires error-annotated data for training, we investigate how much training data is needed to improve results over the baseline of not using a meta-classifier. All evaluations are conducted on a large error- annotated corpus of learner English. 1 Introduction Research on the automatic correction of grammati- cal errors has undergone a renaissance in the past decade. This is, at least in part, based on the recog- nition that non-native speakers of English now outnumber native speakers by 2:1 in some esti- mates, so any tool in this domain could be of tre- mendous value. While earlier work in both native and non-native error correction was focused on the construction of grammars and analysis systems to detect and correct specific errors (see Heift and Schulze, 2005 for a detailed overview), more re- cent approaches have been based on data-driven methods. The majority of the data-driven methods use a classification technique to determine whether a word is used appropriately in its context, continu- ing the tradition established for contextual spelling correction by Golding (1995) and Golding and Roth (1996). The words investigated are typically articles and prepositions. They have two distinct advantages as the subject matter for investigation: They are a closed class and they comprise a sub- stantial proportion of learners’ errors. The investi- gation of preposition corrections can even be narrowed further: amongst the more than 150 Eng- lish prepositions, the usage of the ten most fre- quent prepositions accounts for 82% of preposition errors in the 20 million word Cambridge Universi- ty Press Learners’ Corpus. Learning correct article use is most difficult for native speakers of an L1 that does not overtly mark definiteness and indefi- niteness as English does. Prepositions, on the oth- er hand, pose difficulties for language learners from all L1 backgrounds (Dalgish, 1995; Bitchener et al., 2005). Contextual classification methods represent the context of a preposition or article as a feature vec- tor gleaned from a window of a few words around the preposition/article. Different systems …

Content truncated to 500 words