精彩书摘:
棒性和系统的自适应能力。同书面语相比,口语的声学特性有一定的特殊性,这类语音的基频、时长、幅度等特征都随表达内容、感情色彩等不同,变化的范围比朗读语音大得多,同时还有非语声信号和噪声,充分研究这些特性,建立精细的声学模型非常重要。而且,讲话人往往是在较强的背景噪声或多讲话人环境下发音的,如果是电话自动语音翻译系统,还存在通讯干扰等其他因素的影响,因此,提高语音识别在不同说话人、不同声学环境及通道条件下的鲁棒性,在口语翻译系统中尤其重要。另外,在语言学层面,口语句子中含有大量的修正、重复、口头语、省略等非规范语言现象,研究这些特征,对语言模型进行完善,包括建模、算法和训练等各个方面,将有助于提高语音识别的正确率。<br> (2)翻译方法有待于进一步研究。尽管统计翻译方法具有较高的鲁棒性,但是,对非规范语言现象和噪声的处理能力仍然十分有限,而且这种方法与训练语料的规模和质量密切相关。统计方法与规则方法的结合一直是人们所追求的,但是具体如何融合,多翻译引擎以什么样的集成方式可以获得最好的系统性能,统计模型赖以训练的语料规模达到多大才算充分,非语言信息(手势、表情、说话人角色等)如何融人翻译模型等,诸多问题都远远没有得到解决。<br> ……
内容简介:
计算语言学研究滥觞于上世纪五六十年代的机器翻译研究。中文的相关研究也几乎同步开始,1960年起在柏克莱加州大学研究室,王士元、邹嘉彦、C.Y.Dougherty等人已开始研究中英、中俄机器翻译。他们的中文计算语言学研究,可说是与世界最尖端科技同步的。
目录:
导读<br>Preface<br>Acknowledgements<br>Introduction<br>1.1 What This Book Is About<br>1.1.1 Why Do Spoken Language Translation?<br>1.1.2 What Are the Basic Problems?<br>1.1.3 What Is It Realistic to Attempt Today?<br>1.1.4 What Have We Achieved?<br>1.2 Overall System Architecture<br>1.3 An Illustrative Example<br>1.4 In Defence of Hand-Coded Grammars<br>1.5 Hybrid Transfer<br>1.5.1 The Need for Grammatical Knowledge<br>1.5.2 The Need for Preferences<br>1.6 Speech Processing<br>1.7 Corpora<br><br>Part 1 Language Processing and Corpora<br>Translation Using the Core Language Engine<br>2.1 Introduction: Multi-Engine Translation<br>2.2 Word-to-Word Translation<br>2.3 Quasi Logical Form<br>2.3.1 Introduction<br>2.3.2 Structure of QLF<br>2.3.3 QLF as a Transfer Formalism: Examples <br>2.3.4 Head-Head Relations in QLF<br>2.4 Unification Grammar and QLFs<br>2.4.1 The CLE Unification Grammar Formalism <br>2.4.2 Unification Grammar Example: French Noun Phrases<br>2.4.3 Example 2a: Clauses in Swedish<br>2.4.4 Example 2b: Relative Clauses in Swedish <br>2.5 Orthographic Analysis and the Lexicon<br>2.6 Transfer Rules<br>2.6.1 Pre- and Posttransfer<br>2.7 The QLF-Based Processing Path<br>2.7.1 Linguistic Analysis<br>2.7.2 Transfer and Transfer Preferences<br>2.7.3 Generation<br>2.8 Summary<br><br>Grammar Specialisation<br>3.1 Introduction<br>3.2 Explanation-Based Learning for Grammar<br>Specialisation<br>3.2.1 A Definition of Explanation-Based Learning<br>3.2.2 Explanation-Based Learning on Unification Grammars<br>3.2.3 Category Specialisation<br>3.2.4 Elaborate Cutting-Up Criteria<br>3.3 An LR Parsing Method for Specialised Grammars <br>3.3.1 Basic LR Parsing<br>3.3.2 Prefix Merging<br>3.3.3 Abstraction<br>3.4 Empirical Results<br>3.4.1 Experimental Setup<br>3.4.2 Discussion of Results<br>3.5 Conclusions<br><br>Choosing among Interpretations<br>4.1 Properties and Discriminants<br>4.2 Constituent Pruning<br>4.2.1 Discriminants for Pruning<br>4.2.2 Deciding Which Edges to Prune<br>4.2.3 Probability Estimates for Discriminants<br>4.2.4 Relation to Other Pruning Methods<br>4.3 Choosing among QLF Analyses<br>4.3.1 Analysis Choice: An Example<br>4.3.2 Further Advantages of a Discriminant Scheme .<br>4.3.3 Numerical Metrics<br>4.4 Choosing among Transferred QLFs<br>4.5 Choosing Paths in the Chart<br><br>The TreeBanker<br>5.1 Motivation<br>5.2 Representational Issues<br>5.3 Overview of the TreeBanker<br>5.4 The Supervised Training Process<br>5.4.1 Properties and Discriminants in Training<br>5.4.2 Additional Functionality<br>5.5 Training for Transfer Choice<br>5.6 Evaluation and Conclusions<br><br>Acquisition of Lexical Entries<br>6. 1 The Lexical Acquisition Tool, LexMake<br>6.2 Acquiring Word-to-Word Transfer Rules<br>6.3 Evaluation and Conclusions<br><br>Spelling and Morphology<br>7.1 Introduction<br>7.2 The Description Language<br>7.2.1 Morphophonology<br>7.2.2 Word Formation and Interfacing to Syntax<br>7.3 Compilation<br>7.3.1 Compiling Spelling Patterns<br>7.3.2 Representing Lexical Roots<br>7.3.3 Applying Obligatory Rules<br>7.3.4 Interword Rules<br>7.3.5 Timings<br>7.4 Some Examples<br>7.4.1 Multiple-Letter Spelling Changes<br>7.4.2 Using Features to Control Rule Application<br>7.4.3 Interword Spelling Changes<br>7.5 Debugging the Rules<br>……<br>Part 2 Linguistic Coverage<br>Part 3 Speech Processing
好评度