Abstract
Gaze and language are major pillars in multimodal communication. Gaze is a non-verbal mechanism that conveys crucial social signals in face-to-face conversation. However, compared to language, gaze has been less studied as a communication modality. The purpose of the present study is 2-fold: (i) to investigate gaze direction (i.e., aversion and face gaze) and its relation to speech in a face-to-face interaction; and (ii) to propose a computational model for multimodal communication, which predicts gaze direction using high-level speech features. Twenty-eight pairs of participants participated in data collection. The experimental setting was a mock job interview. The eye movements were recorded for both participants. The speech data were annotated by ISO 24617-2 Standard for Dialogue Act Annotation, as well as manual tags based on previous social gaze studies. A comparative analysis was conducted by Convolutional Neural Network (CNN) models that employed specific architectures, namely, VGGNet and ResNet. The results showed that the frequency and the duration of gaze differ significantly depending on the role of participant. Moreover, the ResNet models achieve higher than 70% accuracy in predicting gaze direction.