Authorship Verification: A Review of Recent Advances

Abstract

Authorship verification attempts to decide whether the author of a given set of texts is also the author of a disputed text. In comparison to closed-set and open-set attribution, the most popular tasks in relevant literature, the verification setting has some important advantages. First, it is more general since any attribution problem can be decomposed into a series of verification cases. Then, certain factors that affect the performance of closed-set and open-set attribution, like the candidate set size and the distribution of training texts over the candidate authors have limited impact on authorship verification. It is, therefore, more feasible to estimate the error rate of authorship attribution technology, needed in the framework of forensic applications, when focusing on the verification setting.Recently, there has been increasing interest for authorship verification, mainly due to the PAN shared tasks organized in 2013, 2014, and 2015. Multiple methods were developed and tested in new benchmark corpora covering several languages and genres. This paper presents a review of recent advances in this field focusing on the evaluation results of PAN shared tasks. Moreover, it discusses successes, failures, and open issues.

Improving author verification based on topic modeling

“…The former attempts to discover and exploit differences in the set of documents of known authorship (Jankowska et al, 2014;Khonji & Iraqi, 2014;Seidman, 2013). The majority of PAN submissions follow this paradigm (Stamatatos, 2016). On the other hand, profile-based approaches handle all available known texts cumulatively (Kocher & Savoy, 2017;Pacheco et al, 2015;Potha & Stamatatos, 2014).…”

“…A supervised learning algorithm can be used to learn a general verification model (Fréry et al, ; Hürlimann, Weck, van den Berg, Šuster, & Nissim, ; Pacheco et al, ). These eager verification approaches (Stamatatos, ) can follow either the intrinsic paradigm (each verification instance is represented using only information from d u and D known ) (Hürlimann et al, ) or the extrinsic paradigm (each instance is represented using information from d u , D known , and D external ) (Pacheco et al, ). The effectiveness of such methods heavily depend on the representativeness of the training set of instances (Stamatatos et al, ).…”

Authorship verification

“…The extraction of stylistic information from documents is usually based on very frequent patterns found in texts. In authorship attribution research there is a wide variety of measures that attempt to capture nuances of the personal style of the authors [129,131]. The basic categories of text representation features used for stylistic purposes are described below:…”

“…A supervised learning algorithm can be used to learn a general verification model [28,43,87,88,93]. These eager verification approaches [131] can follow either the intrinsic paradigm (each verification instance is represented using only information from d unknown and D known ) [8,43,86,117] or the extrinsic paradigm (each instance is represented using information from d unknown , D known , and D external ) [88,93]. The effectiveness of such methods heavily depend on the representativeness of the training set of instances [137].…”

“…In contrast to eager methods, lazy verification approaches [131] do not attempt to extract a general verification model in the training phase. Rather, they treat each test verification case separately.…”