Domain Independent Authorship Attribution without Domain Adaptation

Proceedings of Recent Advances in Natural Language Processing, pages 309–315, Hissar, Bulgaria, 12-14 September 2011. Domain Independent Authorship Attribution without Domain Adaptation Rohith K Menon Department of Computer Science Stony Brook University rkmenon@cs.stonybrook.edu Yejin Choi Department of Computer Science Stony Brook University ychoi@cs.stonybrook.edu Abstract Automatic authorship attribution, by its nature, is much more advantageous if it is domain (i.e., topic and/or genre) indepen- dent. That is, many real world problems that require authorship attribution may not have in-domain training data readily avail- able. However, most previous work based on machine learning techniques focused only on in-domain text for authorship at- tribution. In this paper, we present com- prehensive evaluation of various stylomet- ric techniques for cross-domain authorship attribution. From the experiments based on the Project Gutenberg book archive, we discover that extremely simple techniques based on stopwords are surprisingly robust against domain change, essentially ridding the need for domain adaptation when sup- plied with a large amount of data. 1 Introduction Many real world problems that require authorship attribution, such as forensics (e.g., Luyckx and Daelemans (2008)) or authorship dispute for old literature (e.g., Mosteller and Wallace (1984)) may not have in-domain training data readily available. However, most previous work to date has focused on authorship attribution only for in-domain text (e.g., Stamatatos et al. (1999), Luyckx and Daele- mans (2008), Raghavan et al. (2010)). On lim- ited occasions researchers include heterogeneous (cross-domain) dataset in their experiments, but they only report the performance on heteroge- neous dataset is much lower than that of homo- geneous dataset, rather than directly tacking the problem of cross-domain or domain independent authorship attribution (e.g., Peng et al. (2003)). The lack of research for cross-domain scenarios is perhaps only reasonable, given that it is under- stood in the community that the prediction power of machine learning techniques does not transfer well over different domains (e.g., Blitzer et al. (2008)). However, the seminal work of Blitzer et al. (2006) has shown that it is possible to mitigate the problem by examining distributional differ- ence of features across different domains, and de- rive features that are robust against domain switch. Therefore, one could expect that applying domain adaptation techniques to authorship attribution can also help with cross-domain authorship attribu- tion. Before hasting into domain adaptation for au- thorship attribution, we take a slightly differ- ent push to the problem: we first examine whether there exist domain-independent features that rarely change across different domains. this is the case, and if such features are suffi-If ciently informative, then domain adaptation might not be required at all to achieve high perfor- mance in domain-independent authorship attribu- tion. Therefore, we conduct a comprehensive em- pirical evaluation using various stylistic features that are likely to be common across different top- ics and genre. From the experiments based on the Project Gutenberg book archive, we indeed discover stylistic features that are common across different domains. Against our expectations, some of such features, stop-words in particular, are extremely informative, essentially ridding of the need …

Content truncated to 500 words