Authorship Attribution Using Text Distortion

Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 1138–1149, Valencia, Spain, April 3-7, 2017. c⃝2017 Association for Computational Linguistics Authorship Attribution Using Text Distortion Efstathios Stamatatos University of the Aegean 83200, Karlovassi, Samos, Greece stamatatos@aegean.gr Abstract Authorship attribution is associated with important applications in forensics and hu- manities research. A crucial point in this field is to quantify the personal style of writing, ideally in a way that is not af- fected by changes in topic or genre. In this paper, we present a novel method that en- hances authorship attribution effectiveness by introducing a text distortion step be- fore extracting stylometric measures. The proposed method attempts to mask topic- specific information that is not related to the personal style of authors. Based on ex- periments on two main tasks in authorship attribution, closed-set attribution and au- thorship verification, we demonstrate that the proposed approach can enhance exist- ing methods especially under cross-topic conditions, where the training and test cor- pora do not match in topic. 1 Introduction Authorship attribution is the task of determining the author of a disputed text given a set of candi- date authors and samples of their writing (Juola, 2008; Stamatatos, 2009). This task has gained increasing popularity since it is associated with important forensic applications, e.g., identifying the authors of anonymous messages in extremist forums, verifying the author of threatening email messages, etc. (Abbasi and Chen, 2005; Lambers and Veenman, 2009; Coulthard, 2013), as well as humanities and historical research, e.g., unmask- ing the authors of novels published anonymously or under aliases, verifying the authenticity of lit- erary works by specific authors, etc. (Koppel and Seidman, 2013; Juola, 2013; Stover et al., 2016). The majority of published works in authorship attribution focus on closed-set attribution where it is assumed that the author of the text under investigation is necessarily a member of a given well-defined set of candidate authors (Stamatatos et al., 2000; Gamon, 2004; Escalante et al., 2011; Schwartz et al., 2013; Savoy, 2013; Seroussi et al., 2014). This setting fits many forensic appli- cations where usually specific individuals have ac- cess to certain resources, have knowledge of cer- tain issues, etc. (Coulthard, 2013) A more general framework is open-set attribution (Koppel et al., 2011). A special case of the latter is authorship verification where the set of candidate authors is singleton (Stamatatos et al., 2000; van Halteren, 2004; Koppel et al., 2007; Jankowska et al., 2014; Koppel and Winter, 2014). This is essentially a one-class classification problem since the negative class (i.e., all texts by all other authors) is huge and extremely heterogeneous. Recently, the verifi- cation setting became popular in research commu- nity mainly due to the corresponding PAN shared tasks (Stamatatos et al., 2014; Stamatatos et al., 2015). In authorship attribution it is not always real- istic to assume that the texts of known author- ship and the texts under investigation belong in the same genre and are …

Content truncated to 500 words