Construction of a Liar Corpus and Detection of Lying Situations

摘要

Recognizing lies in human interaction is one of the most important and challenging tasks in artificial intelligence. For example, it leads to the development of a human-computer interaction system that detects a liar in a conversation. In another example, some systems in the future might occasionally need to tell a lie for not hurting user's feelings, namely a white lie for trouble-free conversations. In this paper, we construct a video corpus for developing a system that detects lies in conversations. The corpus consists of 540 question-answer pairs by 9 persons; 270 true answers and 270 lies. We manually analyze the images in the corpus for recognizing lies. In addition, we apply machine learning techniques, support vector machines with local binary pattern features and geometric features, to our corpus. We discuss the results of the analysis of the corpus and the machine learning approaches to our corpus.