Improving the performance of cross-domain authorship attribution

Abstract

Most previous research on authorship attribution (AA) assumes that the training and test data are drawn from the same distribution. But in real scenarios, this assumption is too strong. Because of domain mismatches, the AA approaches that perform well on same domain scenarios will degrade performance in cross-domain settings. The goal of this research is to improve the prediction results in cross-domain AA (CDAA), where there is no training data available from the target domain. We propose three different CDAA frameworks to overcome the lack of training samples from the target domain. Our first framework is driven by the hypothesis that a simple model built from all available out-of-domain data effectively discriminates among authors for a new domain. In addition to improving the performance of CDAA, we also study the effectiveness of the three most commonly used feature types in AA. In the second framework, we explore character n-grams by separating them into ten distinct categories based on the linguistic aspect they represent. Finally, the third framework tries to represent each instance with a common feature representation that is meaningful across domains. Based on the findings of our first and second framework, we propose to use and compare two formulations of features for CDAA.

We use prediction accuracy as the performance metric. We compare the performance of proposed frameworks with state-of-the-art approaches, whenever possible. We first demonstrate that addition of training data even if it comes from out-of-topic improves the performance of cross-topic AA. Also we find that character n-grams are the most effective author discriminator for both single as well as cross-domain AA. Once we demonstrate the efficacy of character n-grams in CDAA, we then propose to categorize them to further understand their predictive value. We then demonstrate the discriminative power of each n-gram category, and propose to discard some of the worst performing categories. In the third framework, we demonstrate that structural correspondence learning can induce feature correspondences for AA, and these feature correspondences combine with our character n-gram categorization to yield superior performance on cross-domain AA.