Computer Engineering and Applications ›› 2010, Vol. 46 ›› Issue (36): 139-143.DOI: 10.3778/j.issn.1002-8331.2010.36.038

• 数据库、信号与信息处理 • Previous Articles     Next Articles

Predicting Chinese prosodic phrase with height of syntax tree

YANG Hong-wu1,WANG Xiao-li1,CHEN Long2,PEI Dong1,GUO Wei-tong1,CAI Lian-hong3   

  1. 1.College of Physics and Electronic Engineering,Northwest Normal University,Lanzhou 730070,China
    2.Graduate School at Shenzhen,Tsinghua University,Shenzhen,Guangdong 518005,China
    3.Department of Computer Science and Technology,Tsinghua University,Beijing 100084,China
  • Received:2010-06-11 Revised:2010-10-08 Online:2010-12-21 Published:2010-12-21
  • Contact: YANG Hong-wu

基于语法树高度的汉语韵律短语预测

杨鸿武1,王晓丽1,陈 龙2,裴 东1,郭威彤1,蔡莲红3   

  1. 1.西北师范大学 物理与电子工程学院,兰州 730070
    2.清华大学 深圳研究生院,广东 深圳 518005
    3.清华大学 计算机系,北京 100084

  • 通讯作者: 杨鸿武

Abstract: Predicting precise prosodic structure is one of the most important aspects for improving the naturalness of synthesized speech in text to speech synthesis.10000 sentences with part-of-speech tags are used for manually labeling prosodic word and prosodic phrase under a linguistic expert’s guidance.A new feature based on the height of syntax tree is proposed in the paper for describing the degree of closeness between the adjacent lexicon words according to the hierarchal syntax structure.Analysis on the relationships between height of syntax tree and prosodic phrase boundary shows that the height of syntax tree,part-of-speech and length of lexicon word are the three most important features for prosodic phrase boundary prediction.Therefore these three features are employed to train a TBL based model with 400 training sentences.Experiments demonstrate that the approach achieves 75.2% of precision,77.1% of recall and 76.1% of F-Score.

Key words: prosodic structure prediction, syntax structure, prosodic phrase, height of syntax tree, Transformation-based Error-driven Learning(TBL)

摘要: 在文语转换系统中,从文本中预测出准确的韵律结构对于提高合成语音的自然度具有重要的作用。利用10 000句标注了词性标记的文本语料,在语言学专家的指导下,人工标注了语料的韵律词和韵律短语。选择了标注结果一致性最高的500句语句,标注了语法层级结构,并利用语法树高度描述语法词之间连接的紧密程度。通过分析韵律短语边界与语法结构的关系,发现韵律短语边界受语法树高度、语法词词性和语法词词长的影响,因此选择了这三个特征,利用TBL算法和400句训练语句训练了预测模型。测试集上的预测结果表明,提出的方法在小规模训练语料下,韵律短语预测的精确率达到了75.2%,召回率达到了77.1%,F-Score达到了76.1%。

关键词: 韵律结构预测, 语法结构, 韵律短语, 语法树高度, 错误驱动的规则学习算法(TBL)

CLC Number: