Annotation Scheme for Emotionally Relevant Behavior in Multiparty Conversation

We present the development of an annotation scheme for labeling emotions and emotion-related phenomena in multiparty meetings. Our initial experiments suggest that when attempting to characterize emotional episodes, people show preference for describing the intentions and coping behaviors of speakers rather than inferring their emotional state. Accordingly, we propose a pragmatic annotation scheme for emotionally relevant behavior, in addition to a simple assessment of valence on a separate tier. The aim of the annotation is to produce a two-part label per speaker utterance, suitable for subsequent browsing of meeting corpora, machine intervention in meetings, and computational modeling of multiparty discourse in general. We explore interlabeler agreement and correlation between the two tiers.