Using POMDP on Conversational Response Generation with Perceptions in Role Relationships and Emotion States

• Main Authors: Sheng-Feng Li, 李勝豐
• Other Authors: Jui-Feng Yeh
• Format: Others
• Language: zh-TW
• Online Access: http://ndltd.ncl.edu.tw/handle/01275027827123973183

碩士 === 國立嘉義大學 === 資訊工程學系研究所 === 104 === In human-machine interface, the module of spoken input/output in human-machine interface will be a popular area of research in the future. On the other hand, the effectiveness of response generation is one of the important things. Even the user told to system his/her purpose clearly. The communication is still failure if it cannot correctly express efficiently when response. However, if there are different emotions and roles in speaking style when generate response, then it would made the sentences no longer such rigid. Users would think they were speak with human naturally. Therefore, it is a distinctive issue if generate response sentence with emotions and roles. In this paper, in order to generate response with emotions and roles, we use different methods to generate sentence. Then connect conceptual graph with each semantic slot(s) that should be filled in each speech act. And utilize the sentence patterns created in corpus to modeling. Using Partially Observable Markov Decision Process(POMDP) emotion and role ranker to rank candidate sentences, and using current filled states of semantic conceptual graph as current state. Ranking the best response sentence in this state. In experiment results, human assessments on sentence proper, fluency, location of punctuation, diversity, turns, length, emotions and roles all better than the baseline. On the other hand, objective evaluation on readability is almost better than baseline.