Modality Specific Meta Features for Authorship Attribution in Web Forum Posts

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 156–164, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Modality Specific Meta Features for Authorship Attribution in Web Forum Posts Thamar Solorio and Sangita Pillay The University of Alabama at Birmingham 1300 University Boulevard Birmingham, AL 35294, USA {solorio,rsangita}@cis.uab.edu Sindhu Raghavan The University of Texas at Austin 1 University Station C0500 Austin, TX 78712, USA sindhu@cis.uab.edu Manuel Montes y G´omez The University of Alabama at Birmingham 1300 University Boulevard Birmingham, AL 35294, USA National Institute of Astrophysics, Optics, and Electronics Luis Enrique Erro No. 1 Tonantzintla, Puebla, Mexico mmontesg@ccc.inaoep.mx Abstract This paper presents a new method for Au- thorship Attribution (AA) on online fo- rum posts. The idea behind the method is to generate meta features that cap- ture modality specific similarity relations among texts from different authors. Each modality represents a particular linguistic dimension (syntactic, lexical, stylistic). To evaluate this approach we measure predic- tion accuracy on data from an online fo- rum with up to 100 candidate authors. We also compare our results with a state of the art approach that has shown to per- form well across different genres. We have found the meta features to be especially helpful in the online forum domain, where the documents are very short, showing this to be a very promising direction for AA on a realistic web forum scenario. 1 Introduction Authorship attribution (AA) refers to the task of analyzing a document to identify the potential au- thor who wrote the text. Earlier work on this problem involved gathering statistics about the fre- quency of words with specific length, together with other stylistic characteristics extracted from written samples that were in most cases an en- tire book or volume (Mendenhall, 1887; Mosteller and Wallace, 1964). Current approaches to AA relay on casting this problem as a text classifica- tion task, where instead of aiming to do a thematic classification of documents the goal is to have the models learn the distinguishable characteristics in the written work of authors. The focus of analysis on more recent work has also shifted from book- length pieces to documents with length ranging from a couple of blocks (Hirst and Feiguina, 2007) to samples with at most 140 characters (Layton et al., 2010). AA can help settle disputes over the original creators of a given piece of text. But other prac- tical applications include using AA for building a prosecution case against an online abuser. This is an important application, especially when we con- sider the raising trends in cyber-bullying and other electronic forms of teen violence1. Achieving good accuracy in AA on spontaneous online data is, however, far more challenging than the typical scenario for AA. One of the major com- plicating factors involves the limited amount of training data. In the typical scenario, we may have an entire document (several pages long), or even an entire book, while in the case of online data from social media we will …

Content truncated to 500 words