A System for Labeling and Predicting Group Interaction in Meetings

Automatic processing of human meetings requires high-level information about group interaction as well as accurate speech recognition (speech-to-text). The goal of this project was to develop a system for labeling meeting styles and use these labels to evaluate automatic prediction. Meeting Acts (MAs) are descriptors for group interaction that specify the high-level function taking place in a meeting. Dialog Acts (DAs) are labels that describe the function of individual utterances. Using the Meeting Corpus, a collection of transcribed meetings and their accompanying DA labels, we have developed a system of MA labels that include Reporting, Negotiation, Planning, and Brainstorming. Several researchers have applied this labeling system to selected meetings in the corpus. When labels were applied to the same meeting by different labelers, significant agreement was found. Prediction results showing the relationship between high-level information (MAs) and given sentence-level information (DAs) will be presented.