摘要
采用分层语法规则的加权概念转移网络,并将语法规则和类似于统计语言模型中的概率分布结合起来,通过引入平滑的概念,为一些超出词典的OOV词和超出语法规则的词分配一个较小的概率,使模型具有较强的稳健性.实验结果表明:这种分层语法表示灵活、概念清晰、实现简单,可以较大地降低语言模型的混乱度;模型在概念级的预测性能可达到99%的正确率.用该语言模型为语音识别提供预测单元,可以提高识别率.
The proposed hierarchial grammar model is indeed a weighted concept transition network which combines grammar with a method similar to the probability distribution in statistical language model. The smoothing method that assigns minor probabilities for out-of-grammar and OOV phrases is used to make the model robust. The experiments demonstrate that the hierarchial flexible and simple grammar model can decrease language model's perplexity and improve the predictive performance. The accuracy rate can be raised by using the language model in the predictions for speech recognition.
出处
《上海交通大学学报》
EI
CAS
CSCD
北大核心
2003年第10期1585-1587,共3页
Journal of Shanghai Jiaotong University
基金
上海市科学技术委员会基础研究项目(01JC14033)
美国贝尔实验室上海分部资助项目
关键词
语音识别
语言模型
概念转移网络
概念层次
Algorithms
Mathematical models
Performance
Probability distributions
