| Summary: | 碩士 === 國立成功大學 === 資訊工程學系 === 107 === The dialogue system is very popular in this field of artificial intelligence, and the research of the dialogue system has expanded from the previous mono-dialogue to multi-party dialogue. As there are more participants involving in the multiparty dialogue system than the one-to-one dialogue system, there are more interactive issues that need to be explored. The system proposed in this thesis is aimed at the influence of different participant roles on response timing decision and addressee selection in multi-party dialogue systems.
The corpus collected in this thesis is multi-party dialogues on medical consultation for the elderly, with a total of 131 dialogues and 1622 turns. In the collected corpus, there are two users and one system in each dialogue. The multiple users have different participant roles, playing the patient and a companion in the task roles, and the system is a medical expert. There are different combinations of ages and family relationships in social roles. The age is divided into five different structures, and the family relationship is categorized into four relationships and eight roles.
This thesis mainly discusses the influence of participant roles, task roles and social roles on the response timing decision and conversation role in addressee selection. We regard the user intent of each sentence as important historical information. Therefore, we use Google’s BERT for intent detection. The user's different social roles are also encoded as the user's initial vector. Considering task roles, we use three Gated Recurrent Units (GRUs), each for one task role, to update different user history information. Referring to history information and the current user utterance, we decide whether to respond in this turn after the last GRU, and the last hidden layer of GRU is used as the response timing decision embedding. In addressee selection, we improve the model by considering the concept of a conversation role, and the response timing decision embedding is considered as the interaction behavior information when selecting the addressee. We use the reinforcement learning to select the system act and decision tree to select the appropriate output template to build a multi-party dialogue system, which could have a good performance in the task of medical consultation for the elderly.
The experimental results showed that the use of social role embedding and task role modeling in response timing decision could increase the accuracy by 10% compared to the system without the use of social role embedding and task role modeling. Using the information of the response timing decision, addressee selection was increased by 3.7% using Five-fold Cross-Validation experiments compared to the existing methods.
|
|---|
