An Off-the-shelf Approach to Authorship Attribution

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 895–904, Dublin, Ireland, August 23-29 2014. An Off-the-shelf Approach to Authorship Attribution Jamal Abdul Nasir Dep. of Computer Science LUMS Lahore Pakistan jamaln@lums.edu.pk Nico G¨ornitz Dep. of Computer Science TU Berlin Germany goernitz@tu-berlin.de Ulf Brefeld Dep. of Computer Science TU Darmstadt Germany brefeld@cs.tu-darmstadt.de Abstract Authorship detection is a challenging task due to many design choices the user has to decide on. The performance highly depends on the right set of features, the amount of data, in-sample vs. out-of-sample settings, and profile- vs. instance-based approaches. So far, the variety of combinations renders off-the-shelf methods for authorship detection inappropriate. We propose a novel and generally deployable method that does not share these limitations. We treat authorship attribution as an anomaly detection problem where author regions are learned in feature space. The choice of the right feature space for a given task is identified automatically by representing the optimal solution as a linear mixture of multiple kernel functions (MKL). Our approach allows to include labelled as well as unlabelled examples to remedy the in-sample and out-of-sample problems. Empirically, we observe our proposed novel technique either to be better or on par with baseline competitors. However, our method relieves the user from critical design choices (e.g., feature set) and can therefore be used as an off-the-shelf method for authorship attribution. 1 Introduction Automatically attributing a piece of text to its author is one of the oldest problems studied in linguis- tics (Mendenhall, 1887). Despite being an old problem, authorship attribution is still highly topical and todays applications range from plagiarism detection (Maurer et al., 2006), identifying the origin of anonymous harassments in emails, blogs, and chat rooms (Tan et al., 2013) to copyright and estate is- sues as well as resolving historical questions of disputed authorship (Mosteller and Wallace, 1964; Fung, 2003). Intrinsically, the goal of authorship detection is to identify the characteristic traits of an author. The idea is that, these traits distinguish an author from others in terms of writing style, use of words, etc. Thus, prior work often focuses on designing and extracting features from text to capture these traits. There is a great deal of features proposed for authorship detection, including word or character n-grams (Burrows, 1987; Houvardas and Stamatatos, 2006), part-of-speech (Stamatatos et al., 2001), probabilistic context- free grammars (Raghavan et al., 2010), or linguistic features (Koppel et al., 2006). However, indicative features for one author do not necessarily help to characterise another. A major problem in authorship detection is therefore to find the right set of features for a given task at hand (Forman, 2003). Algorithmically, a variety of different models have been studied in the context of authorship detection, ranging from probabilistic approaches (Seroussi et al., 2011) and similarity-based methods (Koppel et al., 2011) to vector space models (Fung, 2003; Li et al., 2006). The approaches either treat documents as independent (instance-based) or concatenate documents by the same author (profile-based). Intuitively, the latter …

Content truncated to 500 words