Abstract: Multi-label classification of electric power customer service work order texts was demonstrated to be crucial for enhancing service efficiency and user satisfaction. To mitigate the insufficient modeling of label relationships and class imbalance in multi-label text classification of electric power customer service work orders, a multi-label text classification method integrating feature enhancement and contrastive learning was proposed. First, the text features of customer service work orders were extracted by employing a pre-trained language model. Then, an innovative text feature enhancement method was developed by integrating global encoding with multi-head attention mechanisms and local encoding modules based on convolutional neural networks. Finally, a K-Nearest Neighbors (KNN) framework enhanced with contrastive learning was implemented for multi-label text classification. Positive samples were generated using the Regularized Dropout (R-Drop) method, while negative samples were reweighted. Supervised contrastive learning loss was incorporated during training to enhance the quality of neighbors retrieved by the KNN mechanism. Experimental results indicate that the proposed method achieves a Micro-F1 score of 92.17% on the electric power customer service work order dataset, surpassing the BERT model by 1.62 percentage points. Additionally, Micro-F1 scores of 75.2% and 88.5% were achieved on the public multi-label text classification datasets AAPD and RCV1-V2, respectively, demonstrating the practical feasibility of improving work order processing accuracy and validating the effectiveness of the proposed method in complex multi-label scenarios

Key words: multi-label text classification, power customer service tickets, contrastive learning, feature enhancement, pre-trained language model

摘要： 摘 要: 电力客服工单多标签文本分类在提升服务效率与用户满意度方面具有重要意义。针对电力客服工单多标签文本分类中标签关系建模不足与类别不平衡问题，提出了一种融合特征增强和对比学习的电力客服工单多标签文本分类方法。首先通过预训练语言模型提取客服工单文本特征，然后结合多头注意力机制的全局编码与卷积神经网络的局部编码模块，设计了一种创新的文本特征增强方法，以有效捕捉电力工单文本中的重要信息并提升特征表达能力。同时引入对比学习改进K近邻算法（K-Nearest Neighbors，KNN）的多标签文本分类框架，采用R-Drop（Regularized Dropout）方法生成正样本，并对负样本进行重新加权，结合监督对比学习损失函数训练，提高了KNN机制推理期间检索到的邻居的质量，有效地缓解了样本不平衡带来的负面影响。实验结果表明，该方法在电力客服工单数据集上的微平均F1分数结果为92.17%，较BERT模型提高了1.62个百分点，优于现有对比方法，同时也分别在多标签文本分类公共数据集AAPD和RCV1-V2上取得了75.2%和88.5%的优异结果，不仅在提升工单处理准确性和服务效率方面展现出较高的应用价值，也进一步验证了在复杂多标签文本分类任务中的有效性。

关键词: 多标签文本分类, 电力客服工单, 对比学习, 特征增强, 预训练语言模型