Cross-Topic Authorship Attribution: Will Out-Of-Topic Data Help?

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1228–1237, Dublin, Ireland, August 23-29 2014. Cross-Topic Authorship Attribution: Will Out-Of-Topic Data Help? Upendra Sapkota and Thamar Solorio The University of Alabama at Birmingham 1300 University Boulevard Birmingham, AL 35294, USA {upendra,solorio}@cis.uab.edu Manuel Montes-y-G´omez Instituto Nacional de Astrof´ısica Optica y Electr´onica Puebla, Mexico mmontesg@ccc.inaoep.mx Steven Bethard The University of Alabama at Birmingham 1300 University Boulevard Birmingham, AL 35294, USA bethard@cis.uab.edu Paolo Rosso NLE Lab - PRHLT Research Center Universitat Polit`ecnica de Val`encia Valencia, Spain prosso@dsic.upv.es Abstract Most previous research on authorship attribution (AA) assumes that the training and test data are drawn from same distribution. But in real scenarios, this assumption is too strong. The goal of this study is to improve the prediction results in cross-topic AA (CTAA), where the training data comes from one topic but the test data comes from another. Our proposed idea is to build a predictive model for one topic using documents from all other available topics. In addition to improving the performance of CTAA, we also make a thorough analysis of the sensitivity to changes in topic of four most commonly used feature types in AA. We empirically illustrate that our proposed framework is significantly better than the one trained on a single out-of-domain topic and is as effective, in some cases, as same-topic setting. 1 Introduction Authorship Attribution is the problem of identifying who, from a number of given candidate authors, wrote the given piece of text. The authorship attribution task can be viewed as a multi-class single-label text classification task where each author indicates a class. However, the purpose of AA is to model each author’s writing style. AA methods have a wide range of applications, including Forensic Linguistics (spam filtering (de Vel et al., 2001), verifying the authorship of threatening emails), cybercrimes (identifying authors of malicious code and defending against pedophiles), and plagiarism detection (Stamatatos, 2011). The AA methods can be useful in applied areas such as law and journalism where the identification of the true author of a piece of text (such as a ransom note) may be able to save lives or help prosecute offenders. One of the outstanding problems in AA studies is the unrealistic assumption that the samples of both known and unknown authorship are drawn from the same distribution. This assumption considerably simplifies the AA task but also limits the practical usability of the methods. In practical scenarios usually the documents under investigation are from a different domain than that of the training documents. We feel the need to advance the way AA methods are designed so that the bridge between domains will be minimized to obtain the optimum performance. Therefore, we try to improve the performance of cross-topic AA (CTAA), one of the dimensions of cross-domain AA (CDAA) where training and test data come from different topics. In this paper, we focus on one of the outstanding research questions on AA: Can we reliably predict the author of a document …

Content truncated to 500 words