Experiments in Open Domain Deception Detection

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1120–1125, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Experiments in Open Domain Deception Detection Ver´onica P´erez-Rosas and Rada Mihalcea Computer Science and Engineering University of Michigan vrncapr@umich.edu, mihalcea@umich.edu Abstract The widespread use of deception in online sources has motivated the need for meth- ods to automatically profile and identify deceivers. This work explores deception, gender and age detection in short texts us- ing a machine learning approach. First, we collect a new open domain deception dataset also containing demographic data such as gender and age. Second, we ex- tract feature sets including n-grams, shal- low and deep syntactic features, seman- tic features, and syntactic complexity and readability metrics. Third, we build clas- sifiers that aim to predict deception, gen- der, and age. Our findings show that while deception detection can be performed in short texts even in the absence of a pre- determined domain, gender and age pre- diction in deceptive texts is a challenging task. We further explore the linguistic dif- ferences in deceptive content that relate to deceivers gender and age and find evi- dence that both age and gender play an im- portant role in people’s word choices when fabricating lies. 1 Introduction Given the potential ethical and security risks as- sociated with deceitful interactions, it is important to build computational tools able not only to detect deceivers but also to provide insights into the na- ture of deceptive behaviors. In particular, informa- tion related to the demographics of the deceivers could be potentially useful, as recent studies have shown that online users lie frequently about their appearance, gender, age or even education level. There are multiple scenarios where it would be desirable to identify deceivers’ demographics; for instance, identifying the age and gender of SMS senders or Twitter users might help improve parental controls, spam filtering, and user’s secu- rity and privacy. In this paper, we present a study on deception detection in an open domain, and also present an analysis of deceptive behavior in association with gender and age. Unlike previous studies, where domain-specific conversational transcripts and re- views have been used, this research targets the identification of deceit in short texts where domain and context are not available. We aim to build deception, age, and gender classifiers using short texts, and also explore the prediction of gender and age in deceptive content. Moreover, we present an analysis of the topics discussed by deceivers given their age and gender based on the assumption that, when lying in an open domain setting, deceivers will show natural bias towards specific topics re- lated to gender and age. 2 Related work To date, several studies have explored the iden- tification of deceptive content in a variety of do- mains, including online dating, forums, social net- works, and consumer reviews. (Toma and Han- cock, 2010) conducted linguistic analyses in on- line dating profiles and identified correlations be- tween deceptive profiles and self references, …

Content truncated to 500 words