Multimodal Negative-Attitude Recognition Toward Automatic Conflict-Scene Detection in Negotiation Dialog

This study aims to present a method for detecting conflict scenes during negotiations through negative-attitude recognition using multimodal features of face-to-face negotiation interactions. This research is conducted based on a new multimodal data corpus that includes annotation of the participants’ negative attitude during the conflicting scenes. Semantic orientation (positive/negative) of the words used by the participants in the negotiation dialog, as well as the speaking turn and prosodic features, are extracted as multimodal features. Linear SVM was used to fuse the multimodal features, and the latest fusion technique was used for estimating the negative attitudes. When the proposed method is applied, classification accuracy of negative/non-negative attitude detection reaches 62.8%; recall is 0.641, precision is 0.645, and F value is 0.638.

full version link