Masking topic‐related information to enhance authorship attribution

Abstract

Authorship attribution attempts to reveal the authors of documents. In recent years, research in this field has grown rapidly. However, the performance of state‐of‐the‐art methods is heavily affected when text of known authorship and texts under investigation differ in topic and/or genre. So far, it is not clear how to quantify the personal style of authors in a way that is not affected by topic shifts or genre variations. In this paper, a set of text distortion methods are used attempting to mask topic‐related information. These methods transform the input texts into a more topic‐neutral form while maintaining the structure of documents associated with the personal style of the author. Using a controlled corpus that includes a fine‐grained range of topics and genres it is demonstrated how the proposed approach can be combined with existing authorship attribution methods to enhance their performance in very challenging tasks, especially in cross‐topic attribution. We also examine cross‐genre attribution and the most challenging, yet realistic, cross‐topic‐and‐genre attribution scenarios and show how the proposed techniques should be tuned to enhance performance in these tasks. Finally, we demonstrate that there are important differences in attribution effectiveness when either conversational genres, nonconversational genres, or a mix of them are considered.