Research On Knowledge-driven Open-domain Multi-turn Dialogue Model

With the continuous progress of the Internet,the conversation system has become more and more popular.It is not only used in smart phones and smart home devices with voice assistants,but also can be widely used in these areas such as customer service and online shopping.It provides great convenience for people’s daily life to help people solve problems and obtain a better user experience.However,there are still many problems and challenges in the current dialogue system.Current dialogue models are easy to generate safe responses,lack information and personalization,and has high repeatability of words,which affects the fluency and naturalness of the dialogue and easily leads to poor user experience.Based on these phenomena,this thesis aims to propose an open domain dialogue model that can generate high-quality answers,improve the original dialogue model based on neural network by introducing external knowledge,so that the dialogue system can make better use of external knowledge to produce more natural,accurate and diverse responses.The research results of this thesis will help improve the quality and efficiency of the dialog system,so as to better meet the needs of users.The main research contents include open domain dialogue model based on language and world knowledge,multi-round dialogue model integrating retelling content and real-time knowledge,design and implementation of a dialogue system,and demonstrate its advantages through man-machine dialogue examples.It includes the following three research contents.1.This thesis proposes a knowledge-enhanced open-domain multi-turn dialogue model that effectively extracts vocabulary corresponding to broader concepts of actual words in the dialogue history by utilizing language knowledge(semantic primitives)and world knowledge(domain words)and replacing them.The model not only learns the coarse-grained information of the dialogue history after replacement,but also grasps the fine-grained information of the original dialogue history as a whole,thus eliminating ambiguity and enriching the representation effect of the dialogue text.At the same time,language knowledge and world knowledge are effectively integrated.This model can integrate multiple information(i.e.,dialogue history enhanced by language knowledge,expanded triple-form world knowledge,knowledge management,and knowledge copying)to make the generated replies knowledgeable and diverse.Experimental results and visualization effects on two benchmark multi-turn dialogue corpora show that the proposed model achieves good results on both automatic and human evaluation metrics.2.This thesis proposes a multi-turn dialogue model that integrates paraphrasing content and real-time knowledge.The model retrieves contextually relevant information through a search engine and applies regularization to filter out irrelevant content and avoid noise during model training.Additionally,the RODS model is adopted to generate paraphrase of the dialogue context and obtain the core content of the context,generating a summary of the main content to ensure that the model can better grasp the dialogue topic during training.By combining paraphrasing and realtime knowledge,the model can improve the coherence and information accuracy of the dialogue and further enhance the semantic expression ability of the dialogue.3.This thesis designs and implements a dialogue system based on the proposed multi-turn dialogue model.The analysis of requirements and functions precedes the introduction to the architecture and different functional components of the system.Each functional module’s implementation steps are elaborated in detail,followed by a demonstration of the system’s various functions.Through human-machine dialogue examples,it is shown that the system can generate accurate and meaningful dialogue responses during conversation,further demonstrating the efficiency and practicality of the system.The system can help users quickly obtain the required information,while also possessing certain intelligence and self-learning capabilities,continuously optimizing and improving the system’s performance to better meet user needs.