【作者】 龚媛；

【导师】 钟义信；

【作者基本信息】 北京邮电大学 ， 信号与信息处理， 2008， 硕士

【摘要】 目前对语音识别后处理的研究正呈现出多样化,语言学知识在研究过程中越来越受到重视,应该更加深入地应用语言学知识,应用自然语言理解方面的各种现有及正在兴起的方法来改善语音识别系统的性能。本课题以此为指导,主要针对“奥运多语言综合信息服务”项目的典型示范系统“CityGuide”,研究语音识别后语句检错纠错方法。将采用基于自然语言理解方法,即主要从语法、语义和语用三个方面出发,重点关注语用信息对识别正确率提高的贡献。本文的主要研究工作和成果有:1,在智能移动终端的语音识别引擎之后引入基于自然语言理解模块,特别需要指出的是,在原有算法(包括语法、语义算法)基础上增加了语用算法和一些辅助算法,使语音识别的正确率约从52%提高到70%。2,目前该演示系统已完成在智能手机上的实验性设计、实现与测试,并尝试引入智能移动平台的语音引擎,实现语音识别及识别后利用自然语言理解方法来进行纠错。目前系统主要支持单句语音输入,所支持语种为中文/英文两种语言。3,提出了一种基于元搜索技术的在线语料知识库采集、学习、构建和更新优化方案,特别针对语言本身存在一定的模糊性和不确定性的特点,探讨了模糊理论在文本分类中的应用,提出了一种梯形隶属度函数法将分类结果模糊化,以及引入模糊熵的概念来评估文本模糊化分类的性能,克服了原有实验系统语料库规模小、领域局限性大、来源不够丰富、缺乏时效性的缺点。更多还原

【Abstract】 At present the post-processing of speech recognition research is showing a diversity of linguistic knowledge in the course of the study, more and more attention should be paid to the knowledge of applied linguistics, in order to improve the performance of speech recognition systems; we should use various existing and emerging methods of natural language understanding.According to the national 863 project of Olympics oriented Multilingual Intelligent Information Service System, this thesis studies mainly on text correction for ASR (Automatic Speech Recognition) result in a demo system called CityGuide. All information will be based on the theory of natural language understanding, that is, mainly from the syntax, semantics and pragmatics of the three aspects, focusing on contribution of the pragmatics information to increase the correct rate. The main research work and achievements are:1, A new module of CI based NLU is added after the ASR module in IMP. Original tests have shown that this module could improve the precision of ASR result to some extent. As to CityGuide corpus testing, after pragmatics and other information is added, the precision of ASR could be improved from 52% to 70%.2, A demo system for this module is implemented in IMP, and original testing is finished. More effort is made to import an ASR program in IMP to connect the ASR and correction directly. Currently the system supports one sentence voice input a time. Chinese and English languages are both acceptable.3, Based on a proposed online search technology corpus knowledgebase acquisition, learning, building and updating optimization programme, in particular for the ambiguity and uncertainty of the language, discussed the application of the fuzzy theory in the text classification, proposed a trapezoidal membership function, and the classification results will be ambiguous, as well as the introduction of the concept of fuzzy entropy to assess the fuzzy text of the classification performance, overcome the shortcomings, thoses are the original small-scale experimental system corpus, the limitations of the field, the source is not rich enough, the lack of limitation.更多还原