Native Language Identification with PPM

Abstract

This paper reports on our work in the NLI shared task 2013 on Native Language Identification. The task is to automatically detect the native language of the TOEFL essays authors in a set of given test documents in English. The task was solved by a system that used the PPM compression algorithm based on an n-gram statistical model. We submitted four runs; word-based PPMC algorithm with normalization and without, character-based PPMC algorithm with normalization and without. The worst result was obtained on training and testing data during the evaluation procedure using the character-based PPM method and normalization: accuracy = 31.9%; the best one was macroaverage F-measure = 0.708 with the word-based PPMC algorithm without normalization.