基于矩阵分解和注意力多任务学习的客服投诉工单分类

投诉工单自动分类是通信运营商客服数字化、智能化发展的要求。客服投诉工单的类别有多层,每一层有多个标签,层级之间有所关联,属于典型的层次多标签文本分类问题,现有解决方法大多数基于分类器同时处理所有的分类标签,或者对每一层级分别使用多个分类器进行处理,忽略了层次结构之间的依赖。提出了一种基于矩阵分解和注意力的多任务学习的方法(MF-AMLA),处理层次多标签文本分类任务。在通信运营商客服场景真实投诉工单分类数据下,与该场景常用的机器学习算法和深度学习算法的Top1F1值相比分别最大提高了21.1%和5.7%。已在某移动运营商客服系统上线,模型输出的正确率97%以上,对客服坐席单位时间的处理效率提升22.1%。

Customer service complaint work order classification based on matrix factorization and attention multi-task learning

The automatic classification of complaint work orders is the requirement of the digital and intelligent de-velopment of customer service of communication operators.The categories of customer service complaint work or-ders have multiple levels,each level has multiple labels,and the levels are related,which belongs to a typical hierar-chical multi-label text classification(HMTC)problem.Most of the existing solutions are based on classifiers to process all classification labels at the same time,or use multiple classifiers for each level,ignoring the dependence between hierarchies.A matrix factorization and attention-based multi-task learning approach(MF-AMLA)to deal with hierarchical multi-label text classification tasks was proposed.Under the classification data of real complaint work orders in the customer service scenario of communication operators,the maximum Top1 F1 value of MF-AMLA is increased by 21.1%and 5.7% respectively compared with the commonly used machine learning algorithm and deep learning algorithm in this scenario.It has been launched in the customer service system of one mobile operator,the accuracy of model output is more than 97%,and the processing efficiency of customer service agent unit time has been improved by 22.1%.