Person Identification from Text and Speech Genre Samples

Proceedings of the 12th Conference of the European Chapter of the ACL, pages 336–344, Athens, Greece, 30 March – 3 April 2009. c⃝2009 Association for Computational Linguistics Person Identification from Text and Speech Genre Samples Jade Goldstein-Stewart U.S. Department of Defense jadeg@acm.org Ransom Winder The MITRE Corporation Hanover, MD, USA rwinder@mitre.org Roberta Evans Sabin Loyola University Baltimore, MD, USA res@loyola.edu Abstract In this paper, we describe experiments con- ducted on identifying a person using a novel unique correlated corpus of text and audio samples of the person’s communication in six genres. The text samples include essays, emails, blogs, and chat. Audio samples were collected from individual interviews and group discussions and then transcribed to text. For each genre, samples were collected for six top- ics. We show that we can identify the com- municant with an accuracy of 71% for six fold cross validation using an average of 22,000 words per individual across the six genres. For person identification in a particular genre (train on five genres, test on one), an average accuracy of 82% is achieved. For identifica- tion from topics (train on five topics, test on one), an average accuracy of 94% is achieved. We also report results on identifying a per- son’s communication in a genre using text ge- nres only as well as audio genres only. 1 Introduction Can one identify a person from samples of his/her communication? What common patterns of communication can be used to identify people? Are such patterns consistent across va- rying genres? People tend to be interested in subjects and topics that they discuss with friends, family, col- leagues and acquaintances. They can communi- cate with these people textually via email, text messages and chat rooms. They can also com- municate via verbal conversations. Other forms of communication could include blogs or even formal writings such as essays or scientific ar- ticles. People communicating in these different “genres” may have different stylistic patterns and we are interested in whether or not we could identify people from their communications in different genres. The attempt to identify authorship of written text has a long history that predates electronic computing. The idea that features such as aver- age word length and average sentence length could allow an author to be identified dates to Mendenhall (1887). Mosteller and Wallace (1964) used function words in a groundbreaking study that identified authors of The Federalist Papers. Since then many attempts at authorship attribution have used function words and other features, such as word class frequencies and measures derived from syntactic analysis, often combined using multivariable statistical tech- niques. Recently, McCarthy (2006) was able to diffe- rentiate three authors’ works, and Hill and Prov- ost (2003), using a feature of co-citations, showed that they could successfully identify scientific articles by the same person, achieving 85% accuracy when the person has authored over 100 papers. Levitan and Argamon (2006) and McCombe (2002) further investigated authorship identification of The Federalist Papers (three authors). The genre of the text may affect …

Content truncated to 500 words