| Summary: | 碩士 === 國立成功大學 === 資訊工程學系碩博士班 === 91 === Recently, most dialogue systems have been designed for a single domain. Also, few researches focused on the construction of the knowledgebase for the dialogue systems. The construction of a knowledgebase with inference and the application to multidomain become important topics and are worth researching. In order to achieve this goal, we construct a medical ontology as the knowledgebase for the system semi-automatically. The medical ontology is used to establish three related services, including registration information service, clinic information service, and FAQ (Frequency Asked Question) service. Then, this approach uses the dialogue management module to integrate the services according to the user’s intention.
In this thesis, we use the bilingual knowledge, WordNet and HowNet to establish a universal ontology first. The island-driven algorithm is then adopted to extract the medical ontology as a knowledgebase for the system. We apply the partial pattern tree (PPT) for intention detection. Finally, the detected intention and its corresponding semantic frames are filled and used to control the whole dialogue process. In addition, the generation of the responses to the input is based on the predefined templates.
In order to evaluate the system performance, we asked 50 college students to test the system. The correct rate for intention detection is 86.2% and the success rate for system operation is 77%. The average length of the dialogues is about 9.2 turns. The naturalness for the response is 78.5%. The correct rate for FAQ module with inference achieves 82% and is improved by 15% in comparison to the keyword-based approach.
|
|---|
