多模态人机交互的说话人意图识别研究.docx

本科毕业设计（论文）题目: 多模态人机交互中的说话人意图识别研究学院计算机学院专业智能科学与技术班级2009211319 2013 年 6多模态人机交互中的说话人意图识别研究摘要在认知服务类机器人系统构建中，意图识别模块是至关重要的一部分。本文研究的重点是设计并实现意图分类模块，并且提取说话人言语中的关键信息。由于口语对话言语不规范现象严重，如重复、颠倒、省略等，所以要选择鲁棒的特征，为此，本文提取了三种不同的特征：关键词特征、知网特征、统计特征，并使用支撑向量机(Support Vector Machine:SVM)、KNN、Na?ve Bayes三种分类器建出扁平分类器、层级分类器。经过实验，得到如下结论：（1）KNN算法在K=3时效果最佳，Bayes分类器使用不去停用词的统计特征能得到较好性能。（2）层级分类器的效果明显优于扁平分类器。（3）使用知网特征增强了系统的覆盖能力。基于这些实验结果，本文确定了一种构建意图识别系统的方案：即使用层级分类器进行意图识别，其中第一级用Bayes分类器，第二级用SVM分类器。此外，在系统构建中，使用ICE分布式中间件，使系统跨平台、跨语言的能力得到支撑。而Python语言的使用以及模块化设计，使系统在开发中具有很大的拓展性和自由度。基于此完成了对话理解模块的任务，为多模态对话系统的建立奠定了基础。关键词：口语对话意图识别 SVM 贝叶斯知网Speaker Intention Recognition Of MultimodalHuman-Computer InteractionABSTRACTIn cognitive service robot system building, the intention recognition module is a vital part. The focus of this study is to design and implement intent classification module and to extract the key words of speakers speech. Because the abnormal phenomena of spoken dialogue is serious (such as repetition, reversed, omitted, etc), it needs to select robust features.In this article, we extract three features: keyword feature, HowNet feature, statistical feature, and build flat classifier and hierarchical classifier by using SVM(Support Vector Machine), KNN and Na?ve Bayes classifier. Through experiments, the following conclusions: (1) KNN algorithm works best when K = 3; Bayes classifier using the statistical characteristics of removing stop words do not get better performance. (2)hierarchical classifier is better than flat classification. (3) HowNet feature enhances system coverage. Based on these experimental results, this article identifies a building plan of intent identification system: uses a hierarchical classifier for intention recognition, wherein the first stage with the Bayes classifier, the second stage with a SVM classifier.In system construction, the use of distributed middleware ICE makes the cross-platform, cross-language capability got supported. By the use of Python language and modular d