Topic or Style? Exploring the Most Useful Features for Authorship Attribution

Proceedings of the 27th International Conference on Computational Linguistics, pages 343–353 Santa Fe, New Mexico, USA, August 20-26, 2018. 343 Topic or Style? Exploring the Most Useful Features for Authorship Attribution Yunita Sari, Mark Stevenson and Andreas Vlachos Department of Computer Science University of Sheffield, UK {y.sari, mark.stevenson, a.vlachos}@sheffield.ac.uk Abstract Approaches to authorship attribution, the task of identifying the author of a document, are based on analysis of individuals’ writing style and/or preferred topics. Although the problem has been widely explored, no previous studies have analysed the relationship between dataset character- istics and effectiveness of different types of features. This study carries out an analysis of four widely used datasets to explore how different types of features affect authorship attribution ac- curacy under varying conditions. The results of the analysis are applied to authorship attribution models based on both discrete and continuous representations. We apply the conclusions from our analysis to an extension of an existing approach to authorship attribution and outperform the prior state-of-the-art on two out of the four datasets used. 1 Introduction Authorship attribution plays an important role in many applications, including plagiarism detection and forensic investigation. Approaches to this problem attempt to identify a document’s author through analysis of individual’s writing style and/or topics they tend to write about. The problem has been extensively studied and a wide range of features has been explored (Stamatatos, 2013; Schwartz et al., 2013; Seroussi et al., 2013; H¨urlimann et al., 2015). However, there has been a lack of analysis of the behavior of features across multiple datasets or using a range of classifiers. Consequently, it is difficult to determine which types of features will be most useful for a particular authorship attribution dataset. Authorship attribution is a unique task which is closely related to both the representation of indi- viduals’ writing style and text categorization. In some cases, where there is a clear topical distinction between the documents written by different authors, content-related features such as those used in text categorization may be effective. However, style-based features are more likely to be effective for datasets containing a more homogeneous set of topics. Previous work on feature exploration for authorship at- tribution, focused on the overall effectiveness of features without considering the characteristics of the datasets to which they were applied, e.g. (Grieve, 2007; Guthrie, 2008; Stamatatos, 2009; Brennan et al., 2012; Sapkota et al., 2015). A wide range of features have been applied to the authorship attribution problem and many previous studies concluded that using character n-grams is often effective, e.g. (Peng et al., 2003; Koppel et al., 2011; Schwartz et al., 2013; Sapkota et al., 2015; Sari et al., 2017; Shrestha et al., 2017). Thus, character n-grams have become the go-to features for this task to capture both an author’s topical preferences and writing style. This study explores how the characteristics of a dataset affect the usefulness of different types of features for the authorship attribution task. Experiments are carried out using four datasets that have previously been …

Content truncated to 500 words