A tractable hybrid ddn–pomdp approach to affective dialogue modeling for probabilistic frame-based dialogue systems

摘要: We propose a novel approach to developing tractable affective dialogue model for probabilistic frame-based systems. The model, based on Partially Observable Markov Decision Process (POMDP) and Dynamic Network (DDN) techniques, is composed of two main parts: the slot-level manager global manager. It has new features: (1) being able deal with large number slots (2) take into account some aspects user's state in deriving adaptive strategies. Our implemented prototype can handle hundreds slots, where each individual slot might have values. illustrated through route navigation example crisis management domain. conducted various experiments evaluate our compare it approximate POMDP techniques handcrafted policies. experimental results showed that DDN–POMDP policy outperforms three policies when action error induced by stress as well observation increases. Further, performance one-step look-ahead after optimizing its internal reward close state-of-the-art counterparts.