ISSN 1000-1239 CN 11-1777/TP

• 人工智能 •

基于深度神经网络的有色金属领域实体识别

1. 1(昆明理工大学信息工程与自动化学院 昆明 650500); 2(昆明理工大学材料科学与工程学院 昆明 650093) (maocunli@163.com)
• 出版日期: 2015-11-01
• 基金资助:
基金项目：国家自然科学基金项目(61262041,61472168,61163004);“科技部创新人才推进计划”中青年科技创新领军人才配套项目(2014HE001);云南省自然科学基金重点项目(2013FA030)

A Kind of Nonferrous Metal Industry Entity Recognition Model Based on Deep Neural Network Architecture

Mao Cunli1, Yu Zhengtao1, Shen Tao2, Gao Shengxiang1, Guo Jianyi1, Xian Yantuan1

1. 1(Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming 650500);2(Faculty of Material Science and Engineering, Kunming University of Science and Technology, Kunming 650093)
• Online: 2015-11-01

Abstract: Aimed at entity recognition in the field of nonferrous metal, and oriented the complex and strongly nested structure characteristics of domain entity such as product names, organizations, and placenames, it is proposed a kind of nonferrous metal industry entity recognition model based on deep neural network (DNN) architecture. In order to effectively use the characteristics of tight combination between the characters of domain entity and to bypass the Chinese words segmentation in the professional field, the model uses neural networks to automatically learn the word embeddings vector representation of Chinese characteristics as its inputs. The denoising autoencoder (DAE) of text window makes some pre-training on each DNN hidden layer. The pre-training extracts optimal feature vector combination which will be used in nonferrous metal domain entity recognition. Moreover, we detailedly describe the pre-training process that the denoising autoencoder of text window based on neuron language model makes and the construction process of deep network on nonferrous metal entity recognition. Finally, to validate the method’s effectiveness, we make some experiments of entity recognition on several entity types, such as product names, mineral names and place names in nonferrous metal domain. The experimental results show that the proposed model has good effect on the entity recognition of the professional domain.