通过数据挖掘任务掌握用户痛点,以及构建电力 AI 客服是国家电网公司提升服务质量的两大途径.实现上述途径面临如何针对电力文本实现准确高效分类的问题.现有文本分类技术通常使用深度学习模型进行特征表示,之后使用Softmax层作为分类器实现分类.但有时SVM作为分类器在对高维复杂的文本张量进行分类时效果可能更佳,而直接使用 SVM 进行分类无法对进行特征表示的深度学习模型进行参数调优.基于上述背景和技术现状,文章提出了一种基于BGAS(BERT-BiGRU-Attention-SVM/Softmax)模型的文本分类方法,先使用Softmax层对特征表示部分进行参数调优,再将分类器替换为SVM,以达到最佳的文本分类效果.为检验BGAS模型的性能,分别设计了分类实验和鲁棒性实验.实验中F1值分别达到了0.844 1和0.733 5,与最佳基线模型相比F1值分别提升了0.024 5和0.019 4.
Research on the Improvement of Power Grid Service Quality Based on the BGAS Model
Grasping user pain points through data mining tasks and building power AI customer service are the two major ways for State Grid Corporation of China to improve service quality.The implementation of the above approaches is faced with the problem of how to achieve accurate and efficient classification of power texts.Current text classification techniques typically utilize deep learning models for feature representation,followed by the implementation of a Softmax layer as a classifier.However,at times,support vector machine(SVM)as a classifier may yield better results when classifying high-dimensional and complex text tensors.However,directly employing SVM for classification cannot optimize the parameters of the deep learning model used for feature representation.Based on the aforementioned background and technological landscape,this paper proposes a text classification approach based on the BGAS(BERT-BiGRU-Attention-SVM/Softmax)model.This method initially employs the Softmax layer to optimize the parameters of the feature representation section,followed by replacing the classifier with SVM to achieve optimal text classification results.To assess the performance of the BGAS model,classification experiments and robustness experiments are respectively designed.In the experiments,F1 scores reached 0.8441 and 0.7335,representing improvements of 0.0245 and 0.0194 compared to the best baseline model.
natural language processingBERTBiGRUattention mechanismSVM
万方数据
维普
