TAT: An Author Profiling Tool with Application to Arabic Emails

TAT: an author profiling tool with application to Arabic emails Dominique Estival, Tanja Gaustad, Son Bao Pham, Will Radford Appen Pty Ltd Chatswood NSW 2067, Australia {destival,tgaustad,sbpham,wradford}@appen.com.au Ben Hutchinson∗ Google Sydney, Australia benhutch@google.com Abstract This paper reports on the application of the Text Attribution Tool (TAT) to profiling the authors of Arabic emails. The TAT sys- tem has been developed for the purpose of language-independent author profiling and has now been trained on two email corpora, English and Arabic. We describe the overall TAT system and the Machine Learning ex- periments resulting in classifiers for the dif- ferent author traits. Predictions for demo- graphic and psychometric author traits show improvements over the baseline for some of the author traits with both the English and the Arabic data. Arabic presents particular challenges for NLP and this paper describes more specifically the text processing compo- nents developed to handle Arabic emails. 1 Introduction The goal of the TAT project is to develop a language- independent Text Attribution Tool (TAT) which can provide information on authors for a variety of docu- ment types and a range of languages. In the first im- plementation, the TAT has been developed for pro- filing the authors of email messages in English and Arabic, with other languages to be added in the fu- ture. (Estival et al., 2007) describes the Machine Learning experiments and the results obtained for the English email data. In this paper, we will focus on the results obtained for Arabic emails and will ∗The work presented in this paper was carried out while this author was working at Appen Pty Ltd. describe the aspects of the TAT which are particular to the processing of our Arabic email data. We first introduce the two tasks of author attri- bution and author profiling in Section 2. In Sec- tion 3, we describe the data set on which our tool was trained and tested. Section 4 contains an overall description of the TAT system, followed by a more specific description of the processing steps required for Arabic emails in Section 5. After presenting the general experimental setup in Section 6, we re- port on the results achieved for several demographic and psychometric traits in Section 7 and propose our general conclusions in Section 8. 2 Author attribution and author profiling The ability to recognise the identity or certain char- acteristics of authors automatically from texts has a number of potential applications. Author identifi- cation and author profiling can provide valuable in- formation for marketing intelligence (Glance et al., 2005), while the rapidly growing field of sentiment analysis and classification (Oberlander and Now- son, 2006) is another application where profiling can make a contribution. Also, author profiling forensics may be helpful in narrowing the choice of potential authors when identifying the source of a threat (Cor- ney et al., 2002; Argamon et al., 2005; Abbasi and Chen, 2005a). Author attribution is the task of deciding for a given text which author, usually from a predefined set of …

Content truncated to 500 words