Ng, Vincent2019-10-072019-10-072019-052019-05May 2019https://hdl.handle.net/10735.1/6965Dialogue systems have attracted a lot of attention since some conversational products like Google Assistant and Amazon Echo smart speaker have achieved big successes recently. In this work, we try to build a pipeline-based task-oriented dialogue system, which is the core technology behind these famous products. Our system consists of three modules: a GLAD dialogue state tracker, a policy learning module and a response generation module. They are sequentially connected. The contributions of this work are two-fold. Firstly, we propose an effective approach to improve the controllability of language generation. The experimental results show that this strategy significantly increases the key information accuracy in the generated dialogue responses. Secondly, we introduce a practical method to build a taskoriented dialogue system. Compared to models that are completely based on neural networks, the modularity of our system helps convert a hard problem into several smaller ones that are more specific and easier to solve.application/pdfen©2019 Yize PangNatural language processing (Computer science)Pipelining (Electronics)Automatic speech recognitionThesis2019-10-07