《personalityfactorsinhumandeceptiondetectioncomparing…》由会员分享,可在线阅读,更多相关《personalityfactorsinhumandeceptiondetectioncomparing…(4页珍藏版)》请在金锄头文库上搜索。
1、Personality Factors in Human Deception Detection: Comparing Human to Machine PerformanceFrank Enos,Stefan Benus,Robin L. Cautin,Martin Graciarena,Julia Hirschberg,Elizabeth ShribergColumbia University,Manhattanville College,SRI,ICSIfrankcs.columbia.eduAbstract Previous studies of human performance i
2、n deception detection have found that humans generally are quite poor at this task, com- paring unfavorably even to the performance of automated proce- dures. However, different scenarios and speakers may be harder or easier to judge. In this paper we compare human to machine per- formance detecting
3、 deception on a single corpus, the Columbia- SRI-Colorado Corpus of deceptive speech. On average, our hu- man judges scored worse than chance and worse than current best machine learning performance on this corpus. However, not all judges scored poorly. Based on personality tests given beforethe tas
4、k, we find that several personality factors appear to correlate with the ability of a judge to detect deception in speech. Index Terms: deception, deceptive, perception, personality.1. IntroductionInterest continues to grow in the research community in the detec- tion of deceptive speech. Such work
5、also has important implica- tions for law enforcement and national security. However, despite a fair number of studies (c.f. 7), relatively little is known about how deception is revealed in the speech signal. How well humans or machines may ultimately perform at the task of detecting de- ceptive sp
6、eech remains an open question. DePaulo 7 catalogs a large number of psychological stud- ies of deception, from a long tradition focused primarily on visual cues. More recently, work has been under way to apply speech technologies and machine learning techniques to a new, cleanly recorded corpus of d
7、eceptive speech, the Columbia-SRI-Colorado (CSC) Corpus 9, 3, 8. Previous research on this corpus has pro-duced two machine learning systems that achieve classification ac- curacies of 66.4% 9 and 64.0% 8 (see Section 5). In this paper, we describe a perception study in which judges attempted to cla
8、ssify as deceptive or truthful the interviews that compose the CSC Corpus. The present work examines human per- formance at classifying the CSC Corpus with respect to two levels of truth/lie judgments. These results contextualize both previous machine learning experiments and future work on this cor
9、pus. In addition we present several strong results suggesting that partic-ular personality factors may contribute significantly to a judgessuccess at classification.2. Previous ResearchA recent meta-analysis 1 examines the results of 108 studies that attempted to determine if individual differences
10、exist in the abilityto detect deception. Ability (where chance is 50%) ranged fromthat of parole officers (40.41%, one study) to that of secret service agents, teachers, and criminals (one study each) who scored in the 6470% range. The bulk of studies (156) used students as judges; they scored on av
11、erage 54.22%.3. The CSC CorpusThe CSC corpus was designed to elicit within-speaker deceptiveand nondeceptive speech 9. Speakers received a financial incen- tive to deceive successfully, and the instructions were designed to link successful deception to the self-presentational perspective 7. That is,
12、 speakers were told that the ability to succeed at de- ception indicated other desirable personal qualities. The corpus comprises interviews of thirty-two native speak- ers of Standard American English who were recruited from the community and the Columbia University student population in ex- change
13、 for payment. Interviewees were told that the study soughtindividuals who fit a profile based on the twenty-five top en- trepreneurs of America.Interviewees answered questions andperformed tasks in six areas. The difficulty of tasks was manip-ulated so that interviewees scored too high to fit the pr
14、ofile in twoareas, too low in two, and correctly in two. Four target profiles ex- isted so that interviewees lies were balanced among the six areas. In the second phase of the study, interviewees were told thattheir scores did not fit the target profile, but that the study alsosought interviewees wh
15、o did not fit the profile but who could con- vince an interviewer that they did. They were told that those who succeeded at deceiving the interviewer would qualify for a draw- ing to receive an additional $100. Interviewees then attempted to convince the interviewer that their scores in each of the
16、six cate-gories matched the target profile. Two kinds of lies are implicit in this context. The global lie is the interviewees overall intention to deceive with respect to each score. The local lie represents statements in support of the reported score; these statements will be either true or false.1The distinction between these types of lie is subtle but important, since interviewees do not always lie at the local level to convey a global lie. For exa