语言模型自适应的目的是减小模型与识别任务之间的语言差异。这些差异包括词典差异、风格和内容差异以及模型的概率分布差异。本文提出一种新的非迭代的中文新词提取方法和一种新的开放式词典的中文语言模型。基于这些技术,本文提出一个面向广播语音识别的语言模型自适应框架,该框架联合了以下技术: 一种新的非迭代的新词提取方法,一种新的中文开放式词典语言模型,一种基于困惑度( PPL) 的背景语料筛选方法和一个 N2gram 概率分布自适应模块。另外,本文还专门分析了在语言模型自适应过程中命名实体词的识别情况。实验表明,通过使用该框架,误识率相对下降了10 % ,实体词识别准确率提高了4 %。
Abstract
The purpose of language model (LM) adaptation is to reduce the linguistic mismatches between t raining corpus and recognition tasks. This paper proposed a new noniterative new words ext raction approach for Chinese and a novel open-vocabulary Chinese LM. To reduce lexicon mismatch , topic and style mismatch and ngram dist ribution mismatch , we also present a unified LM adaptation f ramework which combines our noniterative new words ext raction approach , a novel open-vocabulary Chinese LM , a perplexitybased corpus selection approach and an ngram dist ribution adaptation module. The recognition accuracy of name entity words is also analyzed as an effect of LM adaptation. Experiment s showed about 10 % relative character error rate reduction and 4 % (absolute) recognition accuracy increase of name entity words.
关键词
计算机应用 /
中文信息处理 /
语言模型自适应 /
新词提取 /
开放式词典
{{custom_keyword}} /
Key words
computer application /
chinese information processing /
language model adaptation /
new words ext raction /
open-vocabulary LM
{{custom_keyword}} /
{{custom_sec.title}}
{{custom_sec.title}}
{{custom_sec.content}}
参考文献
[1 ] R. Rosenfeld. Optimizing Lexical and Ngram Coverage Via J udicious Use of Linguistic Data [ A ] . Eurospeech [C] . September , 1995.
[2 ] M. Federico , N. Bertoldi. Broadcast News LM Adaptation over Time [J ] . Computer Speech and Language. October , 2004. 18 (4) : 4172435.
[3 ] R. Rosenfeld. Two decades of statistical language modeling : Where do we go from here [ A ] . In : Proceedings of IEEE ,88 (8) [C] . 2000.
[4 ] R. Iyer , M. Ostendorf . Relevance weighting for combining multi2domain data for n2gram language modeling [J ] . Computer Speech and Lang , 1999. 13 : 2672282.
[5 ] J . R. Bellergarda. An Overview of Statistical Language Model Adaptation [ A ] . In : ITRW on Adaptation Methods for Speech Recognition [ C] . 2001. 1652 174.
[6 ] Pi2Chuan Chang , Shuo2Peng Liao , Lin2shan Lee. Improved Chinese Broadcast News Transcription by Language Modeling with Temporally Consistent Training Corpora and Iterative Phrase Ext raction [ A ] . Eurospeech [C] . Aug. 2003. 4212424.
[7 ] 任纪生,王作英. 一种新的基于主题的语言模型自适应方法[J ] . 中文信息学报,2006 ,20 (4) : 82287.
[8 ] 吴根清,郑方,金凌,吴文虎. 一种在线递增式语言模型自适应方法[J ] . 中文信息学报,2002 ,16 (1) : 602 65.
[9 ] 曲卫民,张俊林,孙乐,孙玉芳. 基于记忆的自适应汉语语言模型的研究[J ] . 中文信息学报,2003 ,17 (5) : 13218.
[10 ] M. Yamamoto and K. Church. Using suffix arrays to compute term f requency and document f requency for all subst rings in a corpus [A] . In : Proceeding of the 6th Workshop on Very Large Corpora [C] . 1998.
[11 ] M. Federico and N. Bertoldi. Broadcast news LM adaptation using contemporary text s [ A ] . Eurospeech [C] . 2001.
[12 ] M. Federico. Bayesian estimation methods for Ngram language model adaptation [ A ] . ICSL P [ C ] . 1996. 2402243.
[13 ] R. Kneser , J . Peters , D. Klakow. Language model adaptation using dynamic marginals [A] . Eurospeech [C] . September 1997. 197121974.
[14 ] Sheng Gao , Bo Xu , Taiyi Huang. A New Framework for Mandarin LVCSR based on onepass decoder [A] . ISCSL P [C] . 2000. 49252.
{{custom_fnGroup.title_cn}}
脚注
{{custom_fn.content}}
基金
国家863 计划资助项目(2006AA010103)
{{custom_fund}}
PDF(1144 KB)
