Feature analysis for native language identification

Abstract

In this study we investigate the role of different features for the task of native language identification. For this purpose, we compile a learner corpus based on a subset of the EF Cambridge Open Language Database – EFCAMDAT [10] developed at the University of Cambridge in collaboration with EF Education. The features we are taking into consideration include character n-grams, positional token frequencies, part of speech n-grams, function words, shell nouns and a set of annotated errors. Last but not least, we examine whether the essays of English learners that share the same mother tongue can be distinguished based on their country of origin.