面向服务机器人的口语对话系统研究与实现

发布时间：2018-04-21 21:11

本文选题：口语对话系统 + 语义相似度　；参考：《哈尔滨工业大学》2017年硕士论文

【摘要】：随着人工智能技术的快速发展,服务机器人已经逐渐进入了日常生活的各个领域,正扮演着越来越重要的角色。同时语音识别和自然语言处理技术的日趋成熟使口语对话系统应用于服务机器人变成可能。本文主要针对服务机器人的具体应用场景,研究了面向服务机器人的口语对话系统。本文主要研究了面向服务机器人口语对话系统中预处理模块、基于常问问题集的问答模块以及基于槽特征的对话管理模块。在预处理模块中,本文首先分析了多种分词方法,包括基于词典的分词方法,基于理解的分词方法以及基于统计的分词方法,随后比较了各个分词方法的特点。并采用基于统计和词典相结合的方法对语音识别结果进行分词,并对该分词方法进行实验。之后采用停用词表去除其中的停用词。最后分析了基于传统语义资源进行关键词拓展的缺陷,即语义资源中词汇覆盖少不适用用口语对话系统。并提出了采用Word2Vec词向量工具对词串中TF-IDF值较高的词进行关键词拓展的方法。在基于常问问题集的问答模块中,本文首先建立了一种数据结构以便进行高效地候选问题集抽取,并对候选问题集的抽取比例进行研究和实验。随后分析了基于向量空间的TF-IDF相似度模型,并提出了用Word2Vec词向量工具计算目标问句和候选问句的语义相似度的方法。最后对上述两种相似度模型进行融合,通过实验确定了其权值,提高了相似度模型的匹配正确率。在基于槽特征的对话管理模块中,本文首先改进了基于向量空间TF-IDF的主题提取方法,采用一个滑动窗对对话文档进行主题提取。并引入模拟冷却来对主题热度进行监测。在本文的最后简述了整个系统的具体实现,并编写了一个用于展示的网页版应用。
[Abstract]:With the rapid development of artificial intelligence technology, service robot has gradually entered every field of daily life and is playing a more and more important role. At the same time, speech recognition and natural language processing technology are becoming more and more mature, which makes it possible for oral dialogue systems to be used in service robots. In this paper, the oral dialogue system of service-oriented robot is studied according to the specific application scenario of service robot. In this paper, the preprocessing module, the question and answer module based on the common question set and the dialogue management module based on the slot feature in the Service-Oriented Robot Oral Dialogue system are studied. In the preprocessing module, this paper first analyzes a variety of word segmentation methods, including dictionary-based word segmentation, word segmentation based on understanding and statistical segmentation, and then compares the characteristics of each word segmentation method. The segmentation method based on the combination of statistics and dictionaries is applied to the segmentation of speech recognition results, and the experiment of the segmentation method is carried out. Then use the stop-off vocabulary to remove the stop-word. Finally, the defects of keyword expansion based on traditional semantic resources are analyzed, that is, the less lexical coverage in semantic resources is not suitable for oral dialogue systems. A method of keyword extension for words with high TF-IDF value in the string is proposed by using Word2Vec word vector tool. In the question and answer module based on the set of frequently asked questions, this paper first establishes a data structure for efficient extraction of candidate question sets, and studies and experiments on the extraction ratio of candidate question sets. Then, the TF-IDF similarity model based on vector space is analyzed, and a method to calculate the semantic similarity between target question and candidate question using Word2Vec word vector tool is proposed. Finally, the above two similarity models are fused, and the weights are determined by experiments, which improves the matching accuracy of the similarity model. In the module of conversation management based on slot feature, this paper first improves the method of topic extraction based on vector space TF-IDF, and uses a sliding window to extract the topic of dialogue document. Simulation cooling is introduced to monitor the heat of the subject. At the end of this paper, the implementation of the whole system is introduced, and a web page application is written.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.1;TP242

【参考文献】