Study for Spontaneous Spoken Dialogue Understanding System

自发口语对话理解系统的研究

基本信息

批准号：
09480059
负责人：
ICHIKAWA Akira
金额：
$ 7.87万
依托单位：
CHIBA UNIVERSITY
依托单位国家：
日本
项目类别：
Grant-in-Aid for Scientific Research (B).
财政年份：
1997
资助国家：
日本
起止时间：
1997 至 2000
项目状态：
已结题

项目摘要

In a spontaneous spoken dialogue understanding system, real-time response and robustness to the environment are required.To realize these requirements, we propose a multi-agent system as the system architecture. Each agent has its own function, e.g.phoneme recognition, input utterance structure reasoning from prosody, input utterance forecasting, word reasoning, parsing, etc. The output of this system is the result of co-operation of individual agents that adjust their own behavior to the environment or the input data independency. We propose the co-operation processes.A reinforcement learning method is proposed for a phoneme recognition agent as a sample agent, and adopted a continuous dynamic programming technique to deal with continuous phoneme recognition. To clarify the fundamental characteristics of the proposed method, we define some simple quasi conditions for the experiments, and confirm favorable results.The prosodic structure of the input utterance is represented as a tree form and constructed using FO, duration and pause information of each phrase of the utterance. The tree structure shows how strong successive phrases are concerned with each other.The forecasting the next input utterance uses the handling of two status transfer tables ; controlling for the dialogue in progress and for the dialogue in future.The system can be expected to achieve high adaptability to the environment(e.g., variation of speakers and tasks)and robustness. Some application systems(e.g.WWW voice browser)ware developed.

在自发口语对话理解系统中，需要实时响应和对环境的鲁棒性。为了实现这些要求，我们提出了多智能体系统作为系统架构。每个智能体都有自己的功能，例如音素识别、从韵律中推理输入话语结构、输入话语预测、单词推理、解析等。该系统的输出是各个智能体合作的结果，它们调整自己的行为环境或输入数据的独立性。我们提出了合作流程。提出了一种以音素识别智能体为样本智能体的强化学习方法，并采用连续动态规划技术来处理连续音素识别。为了阐明该方法的基本特征，我们为实验定义了一些简单的准条件，并确认了良好的结果。输入话语的韵律结构被表示为树形式，并使用每个短语的 FO、持续时间和停顿信息构建的话语。树结构显示了连续短语相互关联的程度。预测下一个输入话语使用两个状态转移表的处理；控制正在进行的对话和未来的对话。该系统有望实现对环境（例如，说话者和任务的变化）的高度适应性和鲁棒性。开发了一些应用系统（如WWW语音浏览器）软件。