Experiments with Convolutional Neural Networks for Multi-Label Authorship Attribution

Experiments with Convolutional Neural Networks for Multi-Label Authorship Attribution Dainis Boumber∗, Yifan Zhang∗, Arjun Mukherjee Department of Computer Science, University of Houston Philip Guthrie Hoffman Hall, 3551 Cullen Blvd., Room 501, Houston, TX 77204-3010, USA {dboumber, yzhang114}@uh.edu, arjun@cs.uh.edu Abstract We explore the use of Convolutional Neural Networks (CNNs) for multi-label Authorship Attribution (AA) problems and propose a CNN specifically designed for such tasks. By averaging the author probability distributions at sentence level for the longer documents and treating smaller documents as sentences, our multi-label design adapts to single-label datasets and various document sizes, retaining the capabilities of a traditional CNN. As a part of this work, we also create and make available to the public a multi-label Authorship Attribution dataset (MLPA-400), consisting of 400 scientific publications by 20 authors from the field of Machine Learning. Proposed Multi-label CNN is evaluated against a large number of algorithms on MLPA-400 and PAN-2012, a traditional single-label AA benchmark dataset. Experimental results demonstrate that our method outperforms several state-of-the-art models on the proposed task. Keywords: multi-label authorship attribution, convolutional neural networks, datasets 1. Introduction Authorship Attribution uses textual features to distinguish between texts written by different authors (Stamatatos, 2009). A typical AA problem is a single-label text- categorization task: given a set of candidate authors for whom text samples of undisputed authorship are available, a text is assigned to one candidate author (Sebastiani and Ricerche, 2002; Stamatatos, 2009). Less common yet equally important is the case that involves identifying mul- tiple co-authors of a document. This is the problem of multi-label AA and as we will see it is significantly harder than classical AA. We briefly review classical AA and then introduce our work in multi-label AA. Classification algorithms utilizing lexical, semantic, syn- tactic, stylistic, and character n-gram features have been explored by Graham et al. (2005), Gamon (2004), Sap- kota et al. (2015), and Shrestha et al. (2017). Qian et al. (2014) proposed a tri-training method to solve AA under limited training data per author. It extended standard co- training using three views: lexical, character and syntactic and was shown to have better generalization performance. This method assumes that a large set of unlabeled docu- ments authored by the same given closed set of authors is available. Sapkota et al. (2016) leveraged Domain Adap- tation in an AA scenario where articles on different top- ics may be written by the same author and labeled training data is limited. The method introduced was a modifica- tion of Structural Correspondence Learning (Blitzer et al., 2006) and requires a large set of unlabeled documents per- taining to the target domain and written by the same au- thors. Seroussi et al. (2012) used latent topic features to improve attribution. Although useful, it requires a large text collection per author. AA via text distortion has been used on traditional PAN corpora (Stamatatos, 2017). AA with a large number of authors and limited training data has ∗The first two authors contributed equally to the work but are named alphabetically. …

Content truncated to 500 words