Detecting Deceptive Language in Crime Interrogation

摘要

Deception detection is a vital research problem studied by fields as diverse as psychology, forensic science, sociology. With cooperation with National Investigation Bureau, we have 496 transcript files, each of which contains a conversation of an interrogator and a subject of a real-world polygraph test during interrogation. Researchers have explored the possibility of natural language process techniques in gaming, news articles, interviews, and criminal narratives. In this paper, we explore the effect of the frontier natural language process technique to detect deceptiveness in these conversations. We regard this task as a binary classification problem. We utilize four different methods, inclusive of part-of-speech extraction, one-hot-encoding, means of embedding vectors, and BERT pre-trained model, to capture hidden information of transcript files into vectors. After that, we take these vectors as training samples of a hierarchy neural network, which is constructed by a fully-connected layer and/or an LSTM layer. After training, our system can take a transcript file as its input and classify whether the subject is deceptive or not. An Fl score 0.733 is achieved from our system.