基于深度学习的开放领域对话系统研究综述  

Survey on Deep Learning Based Open Domain Dialogue System

作　　者：陈晨 朱晴晴[1] 严睿 柳军飞 CHEN Chen;ZHU Qing-Qing;YAN Rui;LIU Jun-Fei(School of Software and Microelectronics,Peking University,Beijing100871;Institute of Computer Science and Technology,Peking University,Beijing100871;National Engineering Research Center for Software Engineering (Peking University),Beijing100871;Office of the Cyberspace Affairs Commission Information Security and Informatization Officeof Peking University,Beijing100871)

机构地区：[1]北京大学软件与微电子学院,北京100871 [2]北京大学计算机科学技术研究所,北京100871 [3]北京大学软件工程国家工程研究中心,北京100871 [4]北京大学网络安全和信息化委员会办公室,北京100871

出　　处：《计算机学报》2019年第7期1439-1466,共28页Chinese Journal of Computers

基　　金：国家自然科学基金(61876196)资助

摘　　要：人机对话系统能够让机器通过人类语言与人进行交互,是人工智能领域的一项重要工作.因其在虚拟助手和社交聊天机器人等领域的商业价值而广受工业界和学术界的关注.近年来,互联网社交数据快速增长促进了数据驱动的开放领域对话系统的研究,尤其是将深度学习技术应用到其中取得了突破性进展.基于深度学习的开放领域对话系统使用海量社交对话数据,通过检索或者生成的方法建立对话模型学习对话模式.将深度学习融入检索式系统中研究提高对话匹配模型的效果,将深度学习融入生成式系统中构建更高质量的生成模型,成为了基于深度学习的开放领域对话系统的主要任务.本文对近几年基于深度学习的开放领域对话系统研究进展进行综述,梳理、比较和分析主要方法,整理其中的关键问题和已有解决方案,总结评测指标,展望未来研究趋势.The human-machine dialogue system enables easy interaction interface between humans and computers using natural languages,which is of growing significance in artificial intelligence. Owing to its commercial value in the fields of virtual assistants and social Chatbots,it has been widely concerned by business and academia. Dialogue systems can be classified as domain- specific and open-domain models. Recently,along with the fast prosperity of social media on the internet,research of data-driven open domain dialogue systems has been promoted. In particular, as a major breakthrough,deep learning has proven to be an extremely powerful tool in this field. The deep learning based open domain dialogue system directly constructs a dialogue model from query to reply by applying end-to-end deep learning techniques to processing massive dialogue data. Our paper starts by summarizing the background and follows by introducing the state - of-the - art methods of implementing the open domain dialogue system: retrieval-based, generation-based and the combination of both. Then,we review the methods that can address several critical problems on this domain. After that,the evaluation procedures of the open domain dialogue system are detailed. Finally,we end up with analyzing and forecasting the future development trend that can bring the dialogue system research into a new frontier.

关 键 词：对话系统 聊天机器人 深度学习 序列到序列模型 匹配模型 对话系统评测 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]