Study on Probabilistic Control Mechanism for Establishing Harmonic Cognitive State between Human and Systems

本文基于统计技术，对建立人与信息系统的和谐认知状态进行了研究，取得了以下成果：1.提出了一种基于麦克风阵列的移动用户语音事件检测与跟踪的统计方法。通过对真实的办公环境中采集的数据进行测试，语音区间估计的正确率达到85%。2.提出了一种包含固定短语的问答任务中用户话语的建模方法。我们将所提出的方法应用于语音驱动的问答系统，并证实识别口语问题的准确率和任务完成率显着提高。我们还提出了一种方法来收集语音驱动的信息检索系统的自发口语查询。利用该方法收集了NTCIR-3和NTCIR-4信息检索任务下的语音查询。3.提出了一种交互式导游系统用户知识水平的统计估计方法。我们收集了12个用户的对话数据，通过绿野仙踪的方法。我们评估的性能的估计方法的问题，区分用户分为4类。4.我们收集了20个用户的自然口语对话，用于控制家庭电子系统的任务。我们转录并标记了语音数据。这些结果支持了基于动态贝叶斯网络的统计方法对于估计用户在各种交互任务中的状态是有效的。未来的工作包括改进估计方法和评估与收集的数据详细，研究如何使用估计的状态系统决策和回答，并构建各种对话系统。

We conducted research aiming to establish harmonic cognitive state between human and information systems based on statistical techniques, and achieved the following results.1.We proposed a statistical method of detecting and tracking moving users' speech events using a microphone array,. The method was evaluated with data acquired in real office environment 85% correct rate for speech interval estimation was achieved.2.We proposed a method of modeling users' utterances in the question answering task which include fixed phrases. We applied the proposed method to a speech-driven question answering system and confirmed that the accuracy of recognizing spoken questions and the task achievement rate were improved significantly. We also proposed a method to collect spontaneously spoken queries for speech-driven information retrieval systems. Using the method we collected spoken queries under the task of NTCIR-3 and NTCIR-4 information retrieval.3.We proposed a statistical method of estimating users' knowledge level for interactive visitor guide systems. We collected dialog data for 12 users through the Wizard of Oz method. We evaluated the performance of the estimation method for the problem of discriminating users into 4 classes. The correct rate about 87% was achieved.4.We collected natural spoken dialog of 20 users for the task of controlling home electronic systems. We transcribed and labeled the speech data. These result supports that the statistical methods based on dynamic Bayesian networks are effective for estimating users' status in the various interaction tasks. Future works include improving the estimation methods and evaluating with collected data in detail, investigating the way of using the estimated state for systems decision making and answering, and constructing various dialog systems.