A Universal Method for Author Identification Using Statistical Properties of Text

Author identification is a major topic in Natural Language Processing whose applications go far beyond recognizing the original author of a text to detecting fraud. Each author has a unique writing style which is revealed by analyzing statistical features of his/her text. Traditionally, statistical features such as word frequencies and n-gram character frequencies were used. Such features depended on language properties and not just on the identity of the author. Alternatively, text features consisting of blocks of 8 consecutive bytes extracted from documents, were used to identify authors. A set of distinct blocks for each author is determined by analyzing text known to be written by that author. Given text of unknown authorship, its distinct blocks are extracted and compared to the set of unique distinct blocks for each known author. The author of the text is then identified as the one who has the highest overlap between his/her distinct blocks and those of the text. This method was tested on Arabic and English texts with 100% accuracy. The language independence of the method is tested in this paper. The method is applied to Spanish, French and German texts. Using this method, the authors of the texts were identified with an accuracy of 100% proving that this method is indeed language independent.