Decision Support for Rapid Assessment of Truth and Deception Using Automated Assessment Technologies and Kiosk-based Embodied Conversational Agents.
Lecturer, University of Arizona
A pressing need exists for effective decision support systems to facilitate the rapid and accurate screening of large volumes of people. Millions of travelers transit through international borders and secure areas on an annual basis. Humans are exceptionally poor at detecting lying and deception and perform, on average, no better than chance. This research study focuses on the development, design and implementation of a Kiosk for Rapid Assessment of Deception (K-RAD) that integrates questioning with response processing and deception detection. An exploratory pilot study (N=68) and a primary study (N=225) were executed. The K-RAD was designed to have a three-dimensional figure, an "Embodied Conversational Agent" (ECA), deliver the questions through speech. This delivery mechanism was chosen because human subjects have been shown in the past to react emotionally to ECAs during conversational interactions, and emotional arousal is one of the cues to deception. Responses were analyzed for deception cues, focusing on kinesic, linguistic, and vocalic characteristics that can be captured for automated processing and which would be unique to this setting. The results show unique subject behaviors. Subjects exhibited minimal movement and had very little tendency to change posture. Some subjects (6%) referred to the ECA as an authority figure, using "sir" when responding. Subjects positioned themselves at varying distances from the ECA, with significant gender differences. Post-experiment surveys indicated a gender difference in overall stress, with female subjects reporting significantly higher levels, independent of the experimental condition. Postural-based logistic regression created significant classification models for the pilot (59.1% classification accuracy) and primary (57.2% & 62.8% classification accuracies) studies. Movement analysis had varying and conflicting results. A robust deception index with a 68.1% classification accuracy was achievable in the pilot study based on high-frequency movement and arm placement. Primary study deception indices were not significant. The results include a comprehensive set of observations and lessons learned regarding kiosk design, deception technologies, detection effectiveness, and future considerations to take into account when creating a next-generation K-RAD system. Many challenges remain, but the concept is functional, promising, and could revolutionize security screening and deception detection in a variety of settings.