Reduce & Attribute: Two-Step Authorship Attribution for Large-Scale Problems

Proceedings of the 23rd Conference on Computational Natural Language Learning, pages 951–960 Hong Kong, China, November 3-4, 2019. c⃝2019 Association for Computational Linguistics 951 Reduce & Attribute: Two-Step Authorship Attribution for Large-Scale Problems Michael Tschuggnall Universit¨at Innsbruck michael.tschuggnall@uibk.ac.at Benjamin Murauer Universit¨at Innsbruck b.murauer@posteo.de G¨unther Specht Universit¨at Innsbruck guenther.specht@uibk.ac.at Abstract Authorship attribution is an active research area which has been actively studied for many decades. Nevertheless, the majority of ap- proaches consider problem sizes of a few can- didate authors only, making them difficult to apply to recent scenarios incorporating thou- sands of authors emerging due to the manifold means to digitally share text. In this study, we focus on such large scale problems and pro- pose to effectively reduce the number of can- didate authors before applying common attri- bution techniques. By utilizing document em- beddings, we show on a novel, comprehen- sive dataset collection that the set of candi- date authors can be reduced with high accu- racy. Moreover, we show that common author- ship attribution methods substantially benefit from a preliminary reduction if thousands of authors are involved. 1 Introduction Correctly determining the author of an anonymous text has been researched for several decades. Un- doubtedly the most groundbreaking work in this area has been conducted in 1964 by Mosteller and Wallace, who attempted to automatically as- sign the authors of the Federalist Papers by utiliz- ing a simple, but yet efficient statistical approach operating on function words (Mosteller and Wal- lace, 1964). By showing that writers can indeed be distinguished by their writing style, many ap- proaches have been published in the following years, proposing enhancements by incorporating a variety of so-called stylometric features, methods and learning techniques (Stamatatos, 2009). By categorizing the problem of authorship attribution as a special form of text categorization (Sebas- tiani, 2002), also the respective methods in terms of different machine learning algorithms are ef- fectively in use. With the advent of deep learn- ing, also several approaches have been proposed recently which utilize comprehensive neural net- works. Nevertheless, a recent comparison of all submitted approaches to the cross-domain author- ship attribution task at PAN1 indicates that deep learning is currently not able to surpass traditional methods (Kestemont et al., 2018). Regardless of the features and methods used, the efficacy of an approach can only be measured by using appropriate datasets. Thereby, the ma- jority of existing approaches focus only on a small number of candidate authors (up to 20, most of the times ten or less, e.g., (Stamatatos, 2009; Juola, 2012)). Only a few studies have examined the performance of authorship attribution approaches on larger amounts of possible authors, e.g., 114 (Madigan et al., 2005), 145 (Luyckx and Daele- mans, 2008), 808 (Hitschler et al., 2017) or 1,000 (Shrestha et al., 2017) candidates. To the best of our knowledge, only three studies dive into the multiple thousands of authors: 10,000 (Koppel et al., 2006, 2011) and even 100,000 (Narayanan et al., 2012). While both studies agree that au- thorship attribution at …

Content truncated to 500 words