Linguistic Cues to Deception and Perceived Deception in Interview Dialogues

Proceedings of NAACL-HLT 2018, pages 1941–1950 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics Linguistic Cues to Deception and Perceived Deception in Interview Dialogues Sarah Ita Levitan, Angel Maredia, & Julia Hirschberg Department of Computer Science Columbia University New York, NY, USA {sarahita@cs, asm2221, julia@cs}.columbia.edu Abstract We explore deception detection in interview dialogues. We analyze a set of linguistic fea- tures in both truthful and deceptive responses to interview questions. We also study the per- ception of deception, identifying characteris- tics of statements that are perceived as truth- ful or deceptive by interviewers. Our analysis show significant differences between truthful and deceptive question responses, as well as variations in deception patterns across gender and native language. This analysis motivated our selection of features for machine learning experiments aimed at classifying globally de- ceptive speech. Our best classification perfor- mance is 72.74 F1-Score (about 27% better than human performance), which is achieved using a combination of linguistic features and individual traits. 1 Introduction Deception detection is a critical problem studied by psychologists, criminologists, and computer scientists. In recent years the NLP and speech communities have increased their interest in de- ception detection. Language cues are inexpen- sive and easy to collect, and research examining text-based and speech-based cues to deception has been quite promising. Prior work has examined deceptive language in several domains, including fake reviews, mock crime scenes, and opinions about topics such as abortion or the death penalty. In this work we explore the domain of interview dialogues, which are similar to many real-world deception conditions. Previous work has presented the results of clas- sification experiments using linguistic features, attempting to identify which features contribute most to classification accuracy. However, stud- ies often do not include an empirical analysis of features. We might know that a particular feature set (e.g. LIWC categories) is useful for deception classification, but we lack insight about the nature of the deceptive and truthful language that makes the feature set useful, and whether the differences in language use are statistically significant. In this work we conduct an empirical analysis of feature sets and report on the different characteristics of truthful and deceptive language. In addition, pre- vious work has focused on the characteristics of deceptive language, and not on the characteristics of perceived deceptive language. We are also in- terested in human perception of deception; that is, what are the characteristics of language that lis- teners perceive as truthful or deceptive? We ex- amine a unique dataset that includes information about both the deceiver and the interviewer, along with interviewer judgments of deception. Along with an analysis of deceptive and truthful speech, we analyze the believed and disbelieved speech, according to reported interviewer judgments. Fi- nally, previous work has focused on general infer- ences about deception; here we include analysis of gender and native language, to study their ef- fect on deceptive behavior, and also their effect on perception of deception. This work contributes to the critical problem of …

Content truncated to 500 words