Automatic Inference of Interaction Occasion in Multiparty Meetings: Spontaneous or Reactive

摘要

In this paper, we propose an automatic multimodal approach for inferring interaction occasions (spontaneous or reactive) in multiparty meetings. A variety of features such as head gesture, attention from others, attention towards others, speech tone, speaking time, and lexical cue are integrated. A support vector machines classifier is used to classify interaction occasions based on these features. Our experimental results verified that the proposed approach was really effective, which successfully inferred the occasions of human interactions with a recognition rate by number of 0.870, an accuracy by time of 0.773, and a class average accuracy of 0.812. We also found that the reactive interactions are easier to recognize than the spontaneous ones, and the lexical cue is very important in detecting spontaneous interactions.