Discriminating Non-Native English with 350 Words
John C. Henderson, Guido Zarrella, Craig Pfeifer, John D. Burger · 2013
This paper describes MITRE’s participation in the native language identification (NLI) task at BEA-8. Our best effort performed at an accuracy of 82.6 % in the eleven-way NLI task, placing it in a statistical tie with the best performing systems. We describe the variety of machine learning approaches that we explored, including Winnow, language modeling, logistic regression and maximum-entropy models. Our primary features were word and character n-grams. We also describe several ensemble methods that we employed for combining these base systems. 1