融合ChineseBERT的电力安全隐患分类研究  

Research on classification of power safety dangers based on ChineseBERT

作　　者：孙耀平 赵洁 李润伟 马强[2] 周琰[2] SUN Yaoping;ZHAO Jie;LI Runwei;MA Qiang;ZHOU Yan(Inner Mongolia Branch,Beijing Jingneng Clean Energy Co.,Ltd.,Hohhot 010070,China;National Meteorological Information Center,Beijing 100081,China)

机构地区：[1]北京京能清洁能源电力股份有限公司内蒙古分公司,内蒙古呼和浩特010070 [2]国家气象信息中心,北京100081

出　　处：《电子设计工程》2023年第11期16-20,共5页Electronic Design Engineering

基　　金：国家重点研发计划(2016YFA0602101);国家气象信息中心信息网络安全与“信创”技术研发创新团队项目(NMIC-202011-05)。

摘　　要：针对循环模块特征提取能力不强、训练效率低等问题,提出了融合ChineseBERT的电力安全隐患分类模型(ChineseBERT-BiSRU++-AT)。ChineseBERT模型通过结合当前词的上下文动态训练文本向量表示,解决了一词多义问题;BiSRU++模块捕捉长距离序列特征和句子内部依赖信息,软注意力机制计算特征对结果的影响程度,由分类层输出结果。在真实电力公司电网安全隐患数据集的实验结果证明,与其他模型相比,ChineseBERT-BiSRU++-AT模型分类性能最高,F1值达到了97.44%,BiSRU++模块训练效率优于传统循环模块。To address the problems of weak feature extraction ability and low training efficiency of recurrent module,a classification model of power safety dangers based on ChineseBERT is proposed.ChineseBERT model solves polysemy by dynamically training the text vector representation combined with the context of the current word;BiSRU++module captures long⁃distance sequence features and sentence internal dependency information,calculates the influence of features on the results by soft attention mechanism,and outputs the results by classification layer.The experimental results show that,compared to other models,the data set of power grid security risks of real power companies show that the classification performance of ChineseBERT-BiSRU++-AT model is the highest,the F1 value reaches 97.44%,and the training efficiency of BiSRU++module is better than that of traditional cycle module.

关 键 词：文本分类 ChineseBERT 简单循环单元 软注意力 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]