Research On End-to-End Task-Oriented Dialogue System Based On Knowledge Reasoning

The human-machine dialogue system is a new way of human-machine interaction,and is also a core technology in the field of artificial intelligence.Among them,the task-oriented dialogue system is designed to help users to accomplish tasks in a specific field.It has attracted much attention on academia and industry due to its wide range of application scenarios.Traditional pipeline task-oriented dialogue systems are divided into several modules and trained separately,which rely on a large number of manual annotations,hence it is difficult to adapt to the new task field.With the development of deep learning,end-to-end task-oriented dialogue systems are superior to pipeline systems in scalability and ease of deployment,which have considerable attention in the field of task-oriented dialogue system.However,most of the existing dialogue models still face some challenges in knowledge reasoning,such as:(1)Rare words and unknown words are easy to appear in real application scenarios,but the existing methods rely heavily on entity embedding as entity representation,resulting in insufficient generalization ability.(2)It is very important for task-oriented dialogue models to conduct knowledge reasoning from domain knowledge base according to dialogue context,but the existing models have weakness in knowledge encoding and knowledge fusion,which leads to insufficient reasoning ability of these models.First,to deal with the rare words and unknown words,this thesis proposes a sequence-tosequence model with enhanced entity representation called SEER.In order to reduce the dependence on entity embedding,SEER considers entities from dialogue history and external knowledge base separately,and obtains more robust entity representation based on their respective characteristics:(1)For the entities from dialogue history,the model learns contextsensitive entity representation.(2)For the entities from external knowledge base,the model learns structure-aware entity representation.In addition,this thesis proposes a switching network that helps decoder to incoporate the two heterogeneous representations into dialogue generation.Then,to solve the problem of knowledge encoding and knowledge fusion,this thesis proposes a Flow-to-Graph sequence-to-sequence model called FG2 Seq.For knowledge encoding,this thesis proposes a Flow-to-Graph operation,which can not only enhance entity representation by introducing semantic information of dialogue context through flow mechanism,but also capture relationship structure information between entities through graph convolution network.For knowledge fusion,we adopt the strategy of generation and then filling,which is beneficial for model to incorporate knowledge into dialogue generation.In this thesis,comparative experiments are conducted on three public multi-turn taskoriented dialogue datasets.The experimental results show that the proposed methods outperform the baselines on both automatic matric and human evaluation metric.This thesis further demonstrates the superiority and shortcomings of proposed methods through model ablation analysis,error analysis,and case study.