Simultaneous prediction of dialog acts and address types in three-party conversations