Titel: Using Dynamic Bayesian Networks in Modelling Conversations Instituut: University of Twente (HMI) Plaats: Enschede Land: Nederland Begindatum: 01-05-2011 Voltooid: Nee Begeleider: Rieks op den Akker Onderwerpen: Multimodal Interactions, Intelligent Agents, Speech and Language Technology, Information Engineering, Computational Intelligence, Human factors

Beschrijving:

Research in automatic addressee classification (Natasa Jovanovic' work on addressee identification) as well as in segmentation and classification of dialogue act segments in conversations (see Rutger Rienks' work and Daan Verbree's master thesis) has shown that dynamic Bayesian models for sequence classification outperform static classifiers for these tasks (see work of Alfred Dielemann and Steve Renals, Lendvai and Geertzen, and Op den Akker). Up to now these tasks: topic segmentation of conversations, dialogue act segmentation, and addressee identification (who the speaker is talking to) have largely been studied and modeled in isolation. In this assignment we develop a dynamic Bayesian network that models the dynamics of a conversation using and predicting speaker, dialogue act type and addressee, using information about interests of speakers into specific topics. In this model each of the participants are modeled by a Dynamic Bayesian Decision network, where observation and action nodes connect individual networks. Participants and interaction parameters are computed using the data of real meetings collected in European projects AMI and Amida.