摘要
传统的命名实体识别方法是将大量手工制定的特征输入到统计学习模型中以实现对词语的标记,能够取得较好的效果,但其手工特征制定的方式增加了模型建立的难度。为了减轻传统方法中手工特征制定的工作量,首先对神经网络语言模型进行无监督训练以得到词语特征的分布式表示,然后将分布式的特征输入到深度信念网络中以发现词语的深层特征,最后进行命名实体识别。该方法在前人研究的基础上利用深度信念网络对神经网络语言模型进行了扩展,提出了一种可用于命名实体识别的深层架构。实验表明,在仅使用词特征和词性特征的条件下,该方法用于命名实体识别的性能略优于基于条件随机场模型的方法,具有一定的使用价值。
Traditional named entity recognition methods, which tag words by inputting a good deal of handmade features into statistics learning models, have achieved good results, but the manual mode of defining features makes it more diffi- cult to build the model. To decrease the workload of the manual mode, this paper firstly got the distributed representa- tion of word features by training the neural network language model without supervision, then discovered the deep fea- tures of words by inputting the distributed features into the deep belief net, finally conducted named entity recognition. The method uses the deep belief net to extend the neural network language model on the basis of research of predeces- sors, and presents a deep architecture which is available for named entity recognition. Experiments show that the me- thod applied to named entity recognition can perform better than traditional conditional random field model if both only using term feature and POS feature,and has a certain use value.
出处
《计算机科学》
CSCD
北大核心
2016年第4期224-230,共7页
Computer Science