Design And Implementation Of End To End Spoken Dialogue Systems

With the development of information technology,and the rapid development of artificial intelligence,people can enjoy the convenient service of science and technology has brought to life,and the demand for convenient services are increasingly high,and we hope to achieve human-computer interaction in a very natural way.The goal of Spoken Dialogue System(SDS)is to enable people to interact with the computer of natural language,and there is a lot of application scenarios.The system architecture of traditional spoken dialogue system is a pipeline model.The system can be divided into three sub tasks:Natural Language Understanding(NLU),Dialogue Management(DM)and Natural Language Generation(NLG).But there are two following shortcomings:one is the three sub tasks are over-dependented,which means that the error produced by the front module may propagate to the post-processing modules;two is the definition of the three sub tasks are generally associated with the specific domain,and the model is often determined according to specific tasks,it is hard to transfer to another domains.One solution that corresponds to the pipeline model is to use the end-to-end framework to output a corresponding response directly to the user-entered text dialog.Most of these end-to-end frameworks use a encoder model that encodes incoming user conversations and then generates machine responses and returns them to the user through the decoder.The existing end-to-end human-machine dialogue system is more focused on chatting robots in the open field.Due to the lack of effective modeling of user's intention and lack of domain knowledge,it is very difficult to solve the important problems in task-oriented human-machine dialogue system.To tackle the above problems,in this thesis we proposed a hierarchical discrete latent intention encoder-decoder model to introduce a discrete latent variables which can be used model the intent intention of the user,and introduced the domain knowledge to make the generated feedback related to specific tasks.Further more,in order to enhance the dialogue success rate,we introduce the reinforcement learning,and regard the selection of latent user intention as a sequential decision problem,so as to improve the success rate.In the specific restaurant ordering task,we designed and implemented a spoken dialogue system,the offline results and human evaluation results show that the proposed model out-performs recently proposed models with BLEU and achieves comparable success rate.