【作者】 王健；

【导师】 何克晶； 杨敏；

【作者基本信息】 华南理工大学 ， 计算机技术（专业学位）， 2020， 硕士

【摘要】 人机对话系统是自然语言处理领域中备受学术界和工业界关注的技术,也是走向真正人工智能道路上富有挑战性的重要任务。根据应用场景的不同,人机对话系统主要分为闲聊式的开放域对话系统和面向服务的任务型对话系统。随着深度神经网络的兴起,端到端生成式对话模型因其有效性和易部署性成为近年来人机对话系统的研究热点。然而,现有的开放域对话系统仍存在对常识知识缺乏融合的问题,不能很好地结合背景知识生成高质量的回复;任务型对话系统往往需要与用户进行多轮交互才能完成特定任务,随着对话上下文变长,模型对领域特定知识库中的知识存在推理能力不足,进而难以准确响应用户的目标。因此,本文旨在研究端到端生成式对话系统中的知识感知与推理,针对现有对话生成模型的缺陷,分别从知识感知和知识推理两大方面对生成式对话展开研究,进而提出有效的模型与算法。针对开放域对话系统中的常识知识融合问题,本文提出一种基于知识感知的开放域对话生成模型,它从基于知识库的问答任务中迁移其问题表示和知识匹配的能力,用以辅助对话生成模型对输入话语中的常识知识理解,并生成符合事实的知识实体。此外,本文还提出一种回复引导注意力机制和多步解码策略,它有助于对话生成模型更好地捕获用于生成回复语句所相关的特征。针对任务型对话系统中的知识推理问题,本文提出一种对称动态记忆网络模型,它包括对话记忆管理器和知识库记忆管理器两大核心部分。对话记忆管理器通过一种动态更新机制追踪长期的多轮对话状态信息并有效捕获当前轮次的对话信息,而知识库记忆管理器则通过一种动态记忆指针有效获取领域特定知识库的知识,两者的分离与交互使得对话解码器更好地完成知识推理并生成高质量的对话回复。本文在多个公开的对话数据集上进行了对比实验,实验结果表明,本文提出的方法在多个自动评估指标和人为评估上取得显著优于基准模型的效果。本文还通过模型消融分析、错误分析和生成结果展示进一步探索了本文方法的优越性与不足之处。更多还原

【Abstract】 The human-machine dialogue system is a technology that has attracted much attention from academia to industry in the field of natural language processing.It is also a challenging task towards general artificial intelligence.According to application scenarios,the human-machine dialogue system can be mainly divided into chit-chat based open-domain dialogue system and service-based task-oriented dialogue system.With the rapid development of deep neural networks,end-to-end generative dialogue models have become a research hotspot in recent years due to their effectiveness and ease of deployment.However,existing open-domain dialogue systems still suffer from the incorporation of commonsense knowledge and cannot generate high-quality responses by combining with background knowledge.Task-oriented dialogue systems often require multiple rounds of interaction with users to complete specific tasks.As the dialogue context becomes longer,the reasoning ability of dialogue models towards the knowledge in the domain-specific knowledge base becomes poorer,which makes it difficult to respond to the user’s goals accurately.Therefore,this thesis aims to study knowledge perception and reasoning in end-to-end generative dialogue systems.For the shortcomings of existing dialog generation models,we explore generative dialogues from the two aspects,knowledge perception and knowledge reasoning.Furthermore,we propose effective models and algorithms.For the commonsense knowledge incorporation in open-domain dialogue systems,this thesis proposes an open-domain dialogue generation model based on knowledge perception,which transfers the ability of question representation and knowledge matching from the Knowledge Base Question Answering(KBQA)to assist dialogue generation.It helps the dialogue generation model understand the commonsense knowledge in the input utterance and generate knowledge entities related to the facts.In addition,this thesis proposes a response guiding attention mechanism and a multi-step decoding strategy,which help the dialogue model better capture relevant features for response generation.For the knowledge reasoning in task-oriented dialogue systems,this thesis proposes a Dual Dynamic Memory Network(DDMN)model,which includes two core modules: dialogue memory manager and knowledge base memory manager.The dialogue memory manager tracks long-term multiple turns of dialogue state information,and effectively captures the information in the current turn through a dynamic updating mechanism,while the knowledge base memory manager effectively acquires domain-specific knowledge through a dynamic memory pointer.The separation and interaction between the two modules make the dialogue decoder perform knowledge reasoning better and generate high-quality responses.In this thesis,comparative experiments are conducted on multiple open dialogue datasets.The experimental results demonstrate that the methods proposed in this thesis achieve significantly better results than the baseline models on both automatic evaluation with multiple metrics and human evaluation.This thesis also explores the superiority and shortcomings of our methods through model ablation analysis,error analysis,and display of generation results.更多还原