基于层次MDP的对话管理系统研究与实现
发布时间：2019-06-13 20:45
【摘要】：对话管理(DM:Dialogue Management)在人机对话系统(DS:Dialogue System)中扮演着重要角色。基于马氏决策过程(MDP:Markov Decision Process)的对话管理建模取得了不少进展,但也存在一些问题。其中之一是维度灾难(Curse of Dimensionality),它导致该模型不能应用在复杂的、交换信息相对庞大的对话系统中。本文在前人的研究基础上提出了一种层次MDP(Tier-MDP)对话管理模型。该模型将对话管理任务分为两层。底层处理一个或多个相互独立的对话子任务,每个子任务接收对话理解的输入,输出对话行为,传递给上一层。上层将底层输出的动作做一个函数转变作为状态,基于此进行决策,获得最后的对话行为。与MDP模型相比,Tier-MDP模型通过分层方式可以有效降低任务的状态空间的规模,同时模型复杂度比分层强化学习(HRL:Hierarchical Reinforcement Learning)低。论文进一步设计实现了 Tier-MDP对话管理模型的求解算法。论文实现了一个基于Tier-MDP对话管理模型的人机对话系统,系统可以执行基于人机对话的在线会议室预定任务,系统具有较好的性能和交互能力,表明了基于Tier-MDP的对话管理系统具有较好的性能。
[Abstract]:DIALOG MANAGEMENT plays an important role in the man-machine dialog system (DS: Dialogue System). There are some problems in the process of dialogue management based on the Markov Decision Process (MDP), but there are some problems. One of these is the Curse of Dimensionality, which leads to the fact that the model cannot be applied in a complex, exchange-information-relatively large dialog system. In this paper, a hierarchical MDP (Tier-MDP) dialog management model is presented on the basis of previous studies. The model divides the session management task into two layers. The bottom layer processes one or more independent dialog sub-tasks, each sub-task receiving an input of a dialogue understanding, outputting a dialog behavior, and transmitting to the previous layer. The upper layer converts the operation of the bottom layer output into a function transition as a state, and makes a decision based on the function, and obtains the final conversation behavior. Compared with the MDP model, the Tier-MDP model can effectively reduce the scale of the state space of the task by the layering mode, and meanwhile, the model complexity level enhancement learning (HRL: Hierarchy Remedial Learning) is low. The paper further designs the algorithm for solving the Tier-MDP dialog management model. The paper realizes a man-machine conversation system based on the Tier-MDP dialog management model. The system can carry out the online meeting room reservation task based on man-machine conversation, and the system has better performance and interactive ability, which shows that the conversation management system based on the Tier-MDP has good performance.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP311.52
本文编号：2498808
[Abstract]:DIALOG MANAGEMENT plays an important role in the man-machine dialog system (DS: Dialogue System). There are some problems in the process of dialogue management based on the Markov Decision Process (MDP), but there are some problems. One of these is the Curse of Dimensionality, which leads to the fact that the model cannot be applied in a complex, exchange-information-relatively large dialog system. In this paper, a hierarchical MDP (Tier-MDP) dialog management model is presented on the basis of previous studies. The model divides the session management task into two layers. The bottom layer processes one or more independent dialog sub-tasks, each sub-task receiving an input of a dialogue understanding, outputting a dialog behavior, and transmitting to the previous layer. The upper layer converts the operation of the bottom layer output into a function transition as a state, and makes a decision based on the function, and obtains the final conversation behavior. Compared with the MDP model, the Tier-MDP model can effectively reduce the scale of the state space of the task by the layering mode, and meanwhile, the model complexity level enhancement learning (HRL: Hierarchy Remedial Learning) is low. The paper further designs the algorithm for solving the Tier-MDP dialog management model. The paper realizes a man-machine conversation system based on the Tier-MDP dialog management model. The system can carry out the online meeting room reservation task based on man-machine conversation, and the system has better performance and interactive ability, which shows that the conversation management system based on the Tier-MDP has good performance.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP311.52
【参考文献】
相关期刊论文 前4条
1 张文彬;;微软小冰进化第三代 拥有图像识别能力[J];计算机与网络;2015年15期
2 胡宝洁;赵忠文;曾峦;张永继;;图灵机和图灵测试[J];电脑知识与技术;2006年23期
3 沈晶;顾国昌;刘海波;;分层强化学习中的Option自动生成算法[J];计算机工程与应用;2005年34期
4 王菁华,钟义信,王枞,刘建毅;口语对话管理综述[J];计算机应用研究;2005年10期
相关硕士学位论文 前1条
1 李立云;基于Option自动生成的分层强化学习方法研究[D];长沙理工大学;2008年
，本文编号：2498808
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2498808.html
