摘要
基于神经语言模型生成汉语词语的实值向量表示,称为词语的分布式表示,相应地以这种分布式表示构造的词特征称为分布式词特征.将这种分布式词特征替换基本块识别任务中所常用的条件随机场模型中的词特征,在清华大学TCT语料上进行了汉语基本块识别任务实验,结果表明:在仅使用词窗口[-2,2]的词特征的模型中,和使用词窗口[-2,2]+词性特征的模型中,采用分布式词特征比传统的词特征的模型的标记精度分别高38.01%,1.86%,说明词语的分布式表示对汉语基本块识别任务是有作用的.
A real - valued vector representation of Chinese words based on neural language model is called dis- tributed representation of words, and the corresponding word feature is also called distributed word feature. The experiments used distributed word feature replacing traditional word feature for the identification of Chi- nese Base Chunks were carried out based on the conditional random field model on the TCT corpus of Tsing- hua University. The results show that the marking precision using distributed word features improves 38.01% than the traditional word feature model only using sliding- window word features of size [ -2,21 and 1.86 % than using sliding - window word features of size [ - 2,21 + part-of-speech feature, respectively. This indicates that the distributed representation of Chinese words is available for the identification task of Chinese base chunk.
出处
《中北大学学报(自然科学版)》
CAS
北大核心
2013年第5期582-585,共4页
Journal of North University of China(Natural Science Edition)
基金
国家自然科学基金资助项目(60873128)
关键词
神经语言模型
分布式词特征
基本块分析
边界识别
neural language model
distributed word representation
Chinese base chunk
boundary identifi- cation