Maximizing Classification Accuracy in Native Language Identification

This paper reports our contribution to the 2013 NLI Shared Task. The purpose of the task was to train a machine-learning system to identify the native-language affiliations of 1,100 texts written in English by nonnative speakers as part of a high-stakes test of general academic English proficiency. We trained our system on the new TOEFL11 corpus, which includes 11,000 essays written by nonnative speakers from 11 native-language backgrounds. Our final system used an SVM classifier with over 400,000 unique features consisting of lexical and POS n-grams occurring in at least two texts in the training set. Our system identified the correct nativelanguage affiliations of 83.6% of the texts in the test set. This was the highest classification accuracy achieved in the 2013 NLI Shared Task.