Research Of Chinese Named Netity Recognition Based On Deep Learning

Chinese Named Entity Recognition is the one of the most basic task in Nature Language Process. It plays an important role in many NLP tasks, such as Question Answering, Information Extraction and so on. This problem has been widely studied by NLP scholars all around the world while many achievements have been made in this domain. However, with the rapid development of the Internet, un-rules and open-domain text data is growing faster and faster, where it puts forward new challenges for Chinese Named Entity Recognition technology.Based on the above introduction, this paper mainly does work as follows:(1) Most of solutions with the current technology and development trend for Named Entity Recognition have been reviewed in this paper. In view of the defects and the special difficulties of Chinese Named Entity Recognition, we discussed the probability of using deep learning to address Chinese Named Entity Recognition.(2) A stack auto-encoders neural network is proposed to build model for Chinese Named Entity Recognition. A classical solution is used to deal with the mapping problem from Chinese text to input vector for deep learning model. We also deduce the vectorized forward and back-propagation formula which is easy to implement in project. At the same time, we summarize a set of effective tricks for initialization and adjustment of parameters to optimize the process of training and entity labeling.(3) Large amount of contrast experiments is taken during experiments based on this approach. The results of experiments show that this deep neural network has a good effect on recognizing Chinese Named Entity. More specific, the comparison result in the People’s Daily corpus set reach the-state-of-the-art, and take more advantages than Conditional Random Fields on the recall of identifying location name and organization name, which is increased by 9.60% and 8.84% respectively. The F-1 score of this deep neural network is increased by 3.76% and 2.35%.(4) A system of Chinese Named Entity Recognition has been implemented based on deep neural network. Instead of traditional approaches which combined rules and statistical methods, we propose incremental learning and boundary entropy as a semi-supervised post-processing method by taking the advantages of deep architecture model. This approach can cover the lack of the Chinese labeled corpus and reduce the time consumption during training model and system maintenance. All of these approaches can make the model deal with Chinese mass data quickly and efficiently under less supervised.This deep neural network has a good effect on recognizing Chinese Named Entity. The application in daily practice shows that this system has good robustness and more easy to maintain.