基于循环神经网络的电力一次设备实体关系抽取模型研究

Research on entity relation extraction model of power primary equipment based on recurrent neural network

为提升变电站一次设备启动过程方案编制的数字化水平,文中对自然语言处理中的实体识别及关系抽取方法进行了研究。在定义了生成相关规则库及方案模板库所需数学模型的基础上,讨论了循环神经网络(RNN)在该场景的实际应用方法。应用过程中,针对传统RNN网络在处理长时依赖信息时会出现梯度消失的问题,文中引入了长短期记忆单元,并基于该单元设计了一种双向的循环神经网络。通过实际工程数据集上进行的对比测试结果表明,在进行实体识别时,所提算法对于复杂实体的识别效果更优,对操作指令的识别F值能达到94.15%;而在进行关系抽取时,该算法的F值则可达95.21%。

In order to improve the digitalization level of the scheme preparation for the starting process of substation primary equipment,the entity recognition and relation extraction methods in naturallanguageprocessing are studied in this paper.On the basis of defining the mathematical model needed to generate the relevant rule base and scheme template base,the practical application method of Recurrent Neural Network(RNN) in this scenario is discussed.In the process of application,in view of the shortcoming that the gradient disappears when the traditional RNN network processes long-term dependent information,this paper introduces a long short-term memory unit,and designs a bidirectional recurrent neural network based on this unit.The comparison test results on actual engineering data sets show that the proposed algorithm has better recognition effect for complex entities,and the recognition F value of operation instructions can reach 94.15%.In relation extraction,the F value of the algorithm can reach 95.21%.