节点文献

基于网格的中文语音文件检索技术的研究

Research on Mandarin Spoken Document Retrieval Based on Lattice

【作者】 高运霞

【导师】 张磊;

【作者基本信息】 哈尔滨工程大学 , 信号与信息处理, 2010, 硕士

【摘要】 语音文件检索技术,可以有效地帮助人们从海量的语音信息资源中找到与自己需求相关的信息,是解决信息爆炸问题最有效的技术手段。随着语音识别技术的不断发展,将语音识别技术与传统的文本信息检索技术相结合来进行语音文件检索已经成为一个趋势。然而,语音识别系统的效果,将严重影响语音文件检索的性能。在大多数情况下,由于模型不匹配或者语料噪声的影响等,使得语音识别的效果往往不能令人满意。针对如何将语音识别技术与信息检索技术有效结合这一问题,本文从语音文件的表示形式及信息检索模型两方面进行考虑,提出了一种新的中文语音文件检索方法。一方面,对于语音文件的表示形式,采用Syllable-lattice结构。Lattice可以提供语音识别的多候选结果,它能够一定程度上减轻语音识别的误识对信息检索系统的影响。同时,基于子词的索引策略—Syllable(音节),可以有效地解决查询请求中的OOV词的问题。另一方面,对于信息检索模型,本文研究了信息检索相关技术,在传统的查询似然信息检索模型中引入了文件长度先验概率。实验表明,基于Syllable-lattice的检索系统的检索效果大大优于传统的One-best,其中,在信息检索模型中引入文件长度先验概率信息,可使基于Syllable-lattice的语音文件检索系统的检索效果达到最优,比基线检索模型提高了约30%。实验证明了所提方法是正确的、可行的、有效的。

【Abstract】 Spoken document retrieval technology can be effective in helping people find relevant information from the flood of information resources. With the advances in speech recognition technology, integrating the information retrieval technology and speech recognition together to realize spoken document retrieval system has become a trend. However, in most cases, because of the mismatch of the model, or the impact of noise, the best results of speech recognition are often unsatisfactory to be used in the spoken document retrieval system.To solve this problem, in this paper, the effects of both retrieval source and retrieval model are considered, combine them effectively to realize a new Mandarin spoken document retrieval method. For the retrieval source, the syllable-lattice providing multiple hypothesis is adopted, which can ameliorate the effect of speech recognition error on information retrieval. At the meanwhile, the syllable-based approach can effectively solve the out-of-vocabulary problem in the query. For the retrieval model, the document length prior is combined with the traditional query likelihood retrieval model.Experimental results show that the retrieval performance of lattice-based method outperforms that of one-best method. Further more, in the retrieval model with the document length prior, lattice-based approach can achieve the best performance, it can improve about 30%. The new method is proved to be correct, feasible and effective by the experiments.

节点文献中: 

本文链接的文献网络图示:

本文的引文网络