基于深度学习的电信运营商网络投诉工单智能分类技术研究  

Research on intelligent classifi cation technology of network complaint work order of telecom operators based on deep learning

作　　者：黄堃[1] 胡涵清 赵东明[1] 王博[2] HUANG Kun;HU Han-qing;ZHAO Dong-ming;WANG Bo(China Mobile Group Tianjin Co.,Ltd.,Tianjin 300020,China;Tianjin University,Tianjin 300354,China)

机构地区：[1]中国移动通信集团天津有限公司,天津300020 [2]天津大学,天津300354

出　　处：《电信工程技术与标准化》2023年第10期6-12,共7页Telecom Engineering Technics and Standardization

摘　　要：本文基于自然语言处理技术和深度学习算法,挖掘运营商投诉工单中结构化和非结构化内容的语义特征规律,构建了面向运营商网络投诉派单场景的大规模多标签智能分类模型TBF。本文使用包括编码器模块和解码器模块的端到端框架构建模型。编码器模块使用嵌入层将输入数据中的原始字段转化成向量表示后,使用文本卷积神经网络和双向长短期记忆网络分别对不同数据类型字段的向量表示进行特征抽取并使用前馈神经网络进行特征融合。解码器模块是多层感知机分类器组成的分类器链结构,用来接收编码器模块的融合结果并预测输出各层级投诉类别标签,从而实现对网络投诉工单的智能分类,达到节约人力成本、提升派单质效的数智化转型目的。通过在运营商实际生产环境中的测试和应用,取得了较为满意的效果,成功助力运营商的客户满意度改善。In this paper,based on natural language processing technology and deep learning algorithm,the semantic feature rules of structured and unstructured content in operator complaint work order are mined,and a largescale multi-label intelligent classifi cation model——TBF for operator network complaint order dispatching scenario is constructed.In this paper,an end-to-end framework including an encoder module and a decoder module is used to build the model.After the encoder module uses an embedding layer to convert the original fi elds in the input data into vector representations,a text convolutional neural network——TextCNN and a variant recurrent neural network——Bi-LSTM are used to extract features from the vector representations of fi elds of diff erent data types respectively,and a feedforward neural network is used to carry out feature fusion.The decoder module is a classifi er chain structure composed of multi-layer perceptron classifi ers,which is used to receive the fusion results of the encoder module and predict the output of complaint category labels at each level,so as to realize the intelligent classifi cation of the network complaint work order,and achieve the purpose of digital intelligent transformation to save labor cost and improve the effi ciency of the single element.Through the testing and application in the actual production environment of the operator,the eff ect is satisfactory,which successfully helps the operator to improve the customer satisfaction.

关 键 词：自然语言处理 深度学习 文本卷积神经网络 双向长短期记忆网络 

分 类 号：TN929.5[电子电信—通信与信息系统]