Maximal Repeats Enhance Substring-based Authorship Attribution

Proceedings of Recent Advances in Natural Language Processing, pages 63–71, Hissar, Bulgaria, Sep 7–9 2015. Maximal Repeats Enhance Substring-based Authorship Attribution Romain Brixtel Department of Organizational Behavior, Faculty of Business and Economics University of Lausanne, Quartier Dorigny, 1015 Lausanne, Switzerland romain.brixtel@unil.ch Abstract This article tackles the Authorship Attri- bution task according to the language in- dependence issue. We propose an alterna- tive of variable length character n-grams features in supervised methods: maximal repeats in strings. When character n- grams are by essence redundant, maximal repeats are a condensed way to represent any substring of a corpus. Our experi- ments show that the redundant aspect of n-grams contributes to the efficiency of character-based techniques. Therefore, we introduce a new way to weight features in vector based classifier by introducing n-th order maximal repeats (maximal re- peats detected in a set of maximal repeats). The experimental results show higher per- formance with maximal repeats, with less data than n-grams based approach (ap- proximately divided by a factor of 10). 1 Introduction Internet makes it easy to let anyone share his opin- ion, to communicate news or to disseminate his lit- erary production. A main feature of textual traces on the web is that they are mostly anonymous. Textual data mining is used to characterise au- thors, by categories (e.g. gender, age, political opinion) or as individuals. The latter case is called the Authorship Attribution (AA) issue. It consists of predicting the author of a text given a prede- fined set of candidates, thus falling in the super- vised machine learning subdomain. This problem is often expressed as the ultimate objective, find- ing the author. Technically the task is to predict a new pair, considering given pairs linking text and author. It is also known as writeprint, in reference of fingerprint in written productions. For a sur- vey, see (Koppel et al., 2009; Stamatatos, 2009; El Bouanani and Kassou, 2014). For AA, stylometry is most often used. The as- sumption is that a writer leaves unintended clues that lead to his identification. Bouanani et al. (2014) define a set of numerical features that re- mains relatively constant for a given author and sufficiently contrasts his writing style against any author’s style. In the previous studies, numerical data such as word-length, and literal data such as words or character strings were used to capture personal style features (Koppel et al., 2011). Un- like words or lemmas that belong to a priori re- sources, character strings are in compliance with a language independent objective. Supervised ma- chine learning techniques are used to learn au- thor’s profile, from a training set where text and author pairs are known. Eventually, results are used to attribute new texts to the right author. This is a multi-variate classification problem. Support Vector Machine (SVM) is one of the favorite ap- proaches to handle such complex tasks (Sun et al., 2012). This is the chosen solution here. AA therefore consists of predicting the author of a textual message given a predefined …

Content truncated to 500 words