基于深度学习的电力安全作业实体识别方法

Entity Recognition Method for Power Safety Operation Based on Deep Learning

针对电力现场作业、设备使用安全等大量文本采用非结构化的形式制定、存储,信息自动挖掘难以实现的问题,提出一种融合注意力机制与ALBERT-BiGRU-CRF的方法。首先将ALBERT作为文本的字向量语义编码层,其次多头注意力机制融合到BiGRU网络模型中作为字符实体信息标签预测层,最终CRF作为全局标签优化层。该方法能够准确捕获专业领域字符间的依赖权重、语境、上下文语义关联等全方位特征。电力安全作业文本实体识别实验结果表明,融合注意力机制与深度学习识别方法比目前常用的算法模型识别F1值高3.05%~11.62%,具有较高准确率,识别效果较好。

Aiming at the problem that a large number of texts such as power field operation and equipment use safety are formulated and stored in unstructured form,which makes it difficult to realize automatic information mining,a method combining attention mechanism and ALBERT BiGRU CRF is proposed.Firstly,ALBERT is used as the word vector semantic coding layer of the text,secondly,the multi head attention mechanism is integrated into the BiGRU network model as the character entity information label prediction layer,and finally CRF is used as the global label optimization layer.This method can accurately capture the omni-directional features such as dependency weight,context,context semantic association and so on.The experimental results of text entity recognition in power safety operation show that the recognition method combining attention mechanism and deep learning has a higher F1 score of 3.05%~11.62%than the commonly used algorithm model,has high accuracy and good recognition effect.