基于POMDP的口语对话系统研究与实现

发布时间：2019-04-08 19:27

【摘要】：随着信息技术的发展,人们几乎可以在任意时间、任意地点实时地使用互联网提供的便捷服务,人与计算设备的交互越来越频繁,因此对交互的便利性、自然度也提出了越来越高的要求。口语对话系统(Spoken Dialogue System)的目标是让人们通过语言这种最自然、便利的方式与计算机进行交流,有着广泛的应用需求。对话管理是口语对话系统的核心,得到了广泛深入的研究,其中,部分可观测马尔可夫决策过程(POMDP)目前被认为是最重要的一个模型。本文针对该模型存在的两个问题,开展了如下工作:针对现有的基于POMDP的对话管理模型中槽个数及每个槽取值都有限的问题,提出了一种对象动态绑定方法。该方法虽然能处理的糟数量依然为有限,但可以允许某个槽的取值为无限。针对口语对话系统中语音识别的不确定性,提出了一种基于语音识别的N-Best结果估计POMDP模型观测概率的方法。结合上述两个改进方法提出了一个改进的POMDP模型,设计了其求解算法。并以面向教学任务的口语对话为例,设计实现了一个对话系统,在该系统上的评测表明,本文提出的改进模型能够处理对象数为无限的教学任务,并且基于N-Best的观测概率估计方法提高了系统的性能。
[Abstract]:With the development of information technology, people can use the convenient services provided by the Internet almost at any time and anywhere in real time. The interaction between people and computing devices is becoming more and more frequent, so it is convenient to interact with each other. Naturalness also puts forward higher and higher requirements. The goal of spoken dialogue system (Spoken Dialogue System) is to make people communicate with computers in the most natural and convenient way, which has a wide range of application needs. Dialogue management is the core of spoken dialogue system and has been widely studied. Among them, part of observable Markov decision-making process (POMDP) is considered to be the most important model at present. Aiming at the two problems existing in the model, this paper has carried out the following work: aiming at the problem that the number of slots and the value of each slot in the existing POMDP-based dialogue management model are limited, a method of object dynamic binding is proposed. Although the amount of mess that can be dealt with by this method is still limited, it can allow the value of a slot to be infinite. Aiming at the uncertainty of speech recognition in spoken dialogue system, a method based on speech recognition based on N-Best results to estimate the observation probability of POMDP model is proposed. Combined with the above two improved methods, an improved POMDP model is proposed and its solution algorithm is designed. Taking the oral dialogue for teaching task as an example, a dialogue system is designed and implemented. The evaluation on this system shows that the improved model can deal with the teaching task with unlimited number of objects, and the evaluation on the system shows that the improved model can deal with the teaching task with unlimited number of objects. The observation probability estimation method based on N-Best improves the performance of the system.
【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

【相似文献】