Leveraging Discourse Information Effectively for Authorship Attribution

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 584–593, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Leveraging Discourse Information Effectively for Authorship Attribution∗ Elisa Ferracane1, Su Wang1,2 and Raymond J. Mooney3 1Department of Linguistics, The University of Texas at Austin 2Department of Statistics and Data Science, The University of Texas at Austin 3Department of Computer Science, The University of Texas at Austin elisa@ferracane.com, shrekwang@utexas.edu, mooney@cs.utexas.edu Abstract We explore techniques to maximize the ef- fectiveness of discourse information in the task of authorship attribution. We present a novel method to embed discourse fea- tures in a Convolutional Neural Network text classifier, which achieves a state-of- the-art result by a significant margin. We empirically investigate several featuriza- tion methods to understand the conditions under which discourse features contribute non-trivial performance gains, and analyze discourse embeddings.1 1 Introduction Authorship attribution (AA) is the task of identi- fying the author of a text, given a set of author- labeled training texts. This task typically makes use of stylometric cues at the surface lexical and syntactic level (Stamatatos et al., 2015), although Feng and Hirst (2014) and Feng (2015) go be- yond the sentence level, showing that discourse information can help. However, they achieve lim- ited performance gains and lack an in-depth anal- ysis of discourse featurization techniques. More recently, convolutional neural networks (CNNs) have demonstrated considerable success on AA relying only on character-level n-grams (Ruder et al., 2016; Shrestha et al., 2017). The strength of these models is evidenced by findings that tra- ditional stylometric features such as word n-grams and POS-tags do not improve, and can sometimes even hurt performance (Ruder et al., 2016; Sari et al., 2017). However, none of these CNN models make use of discourse. ∗The first two authors contributed equally to this work. 1https://github.com/elisaF/authorship- attribution-discourse Our work builds upon these prior studies by exploring an effective method to (i) featurize the discourse information, and (ii) integrate discourse features into the best text classifier (i.e., CNN- based models), in the expectation of achieving state-of-the-art results in AA. Feng and Hirst (2014) (henceforth F&H14) made the first comprehensive attempt at using discourse information for AA. They employ an entity-grid model, an approach introduced by Barzilay and Lapata (2008) for the task of ordering sentences. This model tracks how the grammati- cal relations of salient entities (e.g., subj, obj, etc.) change between pairs of sentences in a doc- ument, thus capturing a form of discourse coher- ence. The grid is summarized into a vector of tran- sition probabilities. However, because the model only records the transition between two consec- utive sentences at a time, the coherence is lo- cal. Feng (2015) (henceforth F15) further extends the entity-grid model by replacing grammatical relations with discourse relations from Rhetori- cal Structure Theory (Mann and Thompson, 1988, RST). Their study uses a linear-kernel SVM to perform pairwise author classifications, where a non-discourse model captures lexical and syntac- tic features. They find that adding the entity- grid with …

Content truncated to 500 words