Authorship Attribution with Latent Dirichlet Allocation

Proceedings of the Fifteenth Conference on Computational Natural Language Learning, pages 181–189, Portland, Oregon, USA, 23–24 June 2011. c⃝2011 Association for Computational Linguistics Authorship Attribution with Latent Dirichlet Allocation Yanir Seroussi Ingrid Zukerman Faculty of Information Technology, Monash University Clayton, Victoria 3800, Australia firstname.lastname@monash.edu Fabian Bohnert Abstract The problem of authorship attribution – at- tributing texts to their original authors – has been an active research area since the end of the 19th century, attracting increased interest in the last decade. Most of the work on au- thorship attribution focuses on scenarios with only a few candidate authors, but recently con- sidered cases with tens to thousands of can- didate authors were found to be much more challenging. In this paper, we propose ways of employing Latent Dirichlet Allocation in authorship attribution. We show that our ap- proach yields state-of-the-art performance for both a few and many candidate authors, in cases where these authors wrote enough texts to be modelled effectively. 1 Introduction The problem of authorship attribution – attributing texts to their original authors – has received con- siderable attention in the last decade (Juola, 2006; Stamatatos, 2009). Most of the work in this field fo- cuses on cases where texts must be attributed to one of a few candidate authors, e.g., (Mosteller and Wal- lace, 1964; Gamon, 2004). Recently, researchers have turned their attention to scenarios with tens to thousands of candidate authors (Koppel et al., 2011). In this paper, we study authorship attribution with few to many candidate authors, and introduce a new method that achieves state-of-the-art performance in the latter case. Our approach to authorship attribution consists of building models of authors and their texts using La- tent Dirichlet Allocation (LDA) (Blei et al., 2003). We compare these models to models built from texts with unknown authors to find the most likely authors of these texts (Section 3.2). Our evaluation shows that our approach yields a higher accuracy than the method recently introduced by Koppel et al. (2011) in several cases where prolific authors are consid- ered, while requiring less runtime (Section 4). This paper is structured as follows. Related work is surveyed in Section 2. Our LDA-based approach to authorship attribution is described in Section 3, together with the baselines we considered in our evaluation. Section 4 presents and discusses the re- sults of our evaluation, and Section 5 discusses our conclusions and plans for future work. 2 Related Work The field of authorship attribution predates modern computing. For example, in the late 19th century, Mendenhall (1887) suggested that word length can be used to distinguish works by different authors. In recent years, increased interest in authorship attribu- tion was fuelled by advances in machine learning, information retrieval, and natural language process- ing (Juola, 2006; Stamatatos, 2009). Commonly used features in authorship attribu- tion range from “shallow” features, such as token and character n-gram frequencies, to features that require deeper analysis, such as part-of-speech and rewrite rule frequencies (Stamatatos, 2009). As in other text classification …

Content truncated to 500 words