Effect of Machine Translation on Authorship Attribution

Abstract

In this investigation, the effect of automatic machine translation on stylometry is investigated. For that purpose, an Arabic corpus called Hundred of Arabic Travelers (HAT), containing 100 authors, is used. The idea is to translate all the texts of this corpus, which are written in Arabic, to the French language by using Microsoft office translate. An authorship attribution system is applied on both datasets in order to attribute the author identity for each text before and after translation. Thus, a comparative evaluation based on the author attribution score is made between the two datasets. Several types of features are tested, namely: rare words, words, word bi-gram, word tri-gram, character bi-gram, character tri-gram, and character tetra-gram. Those features are used with a centroid nearest neighbor distance for classification. The experimental results have shown that the effect of machine translation reduces the stylometric identification performances, but preserves some characteristics of the author, which makes the identification still possible even after translation. The accuracy of authorship attribution, with 100 authors, on the translated documents is about 80% of correct attribution, while the best accuracy obtained on the original documents is 97%.