Author Verification Using Common N-Gram Profiles of Text Documents

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 387–397, Dublin, Ireland, August 23-29 2014. Author Verification Using Common N-Gram Profiles of Text Documents Magdalena Jankowska and Evangelos Milios and Vlado Keˇselj Faculty of Computer Science, Dalhousie University 6050 University Avenue Halifax, NS B3H 4R2, Canada {jankowsk, eem, vlado}@cs.dal.ca Abstract Authorship verification is the problem of answering the question whether or not a sample text document was written by a specific person, given a few other documents known to be authored by them. We propose a proximity based method for one-class classification that applies the Common N-Gram (CNG) dissimilarity measure. The CNG dissimilarity (Keˇselj et al., 2003) is based on the differences in the frequencies of n-grams of tokens (characters, words) that are most common in the considered documents. Our method utilizes the pairs of most dissimilar documents among documents of known authorship. We evaluate various variants of the method in the setting of a single classifier or an ensemble of classifiers, on a multilingual authorship verification corpus of the PAN 2013 Author Identification evaluation framework. Our method yields competitive results when compared to the results achieved by the participants of the PAN 2013 competition on the entire set, as well as separately on two subsets — English and Spanish ones — out of the three language subsets of the corpus. 1 Introduction The task of computational detection of who wrote a given text is a widely studied linguistic and machine learning problem with applications in domains such as forensics, security, criminal and civil law, or liter- ary research. The authorship verification problem is a type of such a computational authorship analysis task, in which, given a set of documents written by one author, and a sample document, we are asked whether or not this sample document was written by this given author. This is different from the more traditional problem of deciding who among a finite number of candidate authors for which we are given sample writings, wrote a document in question, and, albeit more difficult, is often considered to better reflect the real-life problems related to authorship detection (Koppel et al., 2012). We describe our one-class proximity based classification method and evaluate it on the multilingual dataset of the Authorship Identification competition task of PAN 2013 (evaluation lab on uncovering plagiarism, authorship, and social software misuse) (Juola and Stamatatos, 2013). During the competition, to which a variant of our method has been submitted (Jankowska et al., 2013), it yielded ranking 5th (joint) out of 18 with respect to the accuracy, and 1st rank out of 10 in the secondary ranking based on the area under the ROC curve (AUC), which evaluates the ordering of instances by the confidence score. In this paper we show some further experiments on how a different way of tuning the classifier parameters, using solely the training dataset of the competition, as well as an ensemble of classifiers based on our method, without any parameter tuning, leads to competitive …

Content truncated to 500 words