摘要
针对中文命名实体识别中长短期记忆网络识别模型缺乏并行性的问题,提出一种融合双向准循环神经网络(BQRNN)与条件随机场(CRF)的中文命名实体识别模型。通过BQRNN网络在序列维度和特征维度上并行获取序列化文本的内部特征,由CRF层选取最终的标签序列,在模型中添加Attention机制,增强BQRNN网络输出的特征信息。实验结果表明,该模型与BLSTM-CRF模型相比F1值提高了1.81%,缩短了约40%的运行时间。
Aiming at the problem that Chinese named entity recognition models based on long-short term memory network lack parallelism,a Chinese named entity recognition model based on bidirectional quasi-recurrent neural networks(BQRNN)and conditional random fields(CRF)was proposed.The internal features of serialized text were obtained in parallel on the sequence dimension and feature dimension through the BQRNN,the final tag sequence was selected from the CRF layer,and the Attention mechanism was added to the model to enhance the feature information of BQRNN output.Experimental results show that compared with the BLSTM-CRF model,the F1 value is increased by 1.81%and the running time is shortened by about 40%.
作者
王栋
李业刚
张晓
蒲相忠
WANG Dong;LI Ye-gang;ZHANG Xiao;PU Xiang-zhong(College of Computer Science and Technology,Shandong University of Technology,Zibo 255049,China)
出处
《计算机工程与设计》
北大核心
2020年第7期2038-2043,共6页
Computer Engineering and Design
基金
国家自然科学基金面上基金项目(61671064)。
关键词
命名实体
准循环神经网络
条件随机场
注意力机制
并行性
name entity
quasi-recurrent neural networks
conditional random fields
Attention mechanism
parallelism
