| 引用本文： |
-
湛志宏,覃开贤,彭凌华,湛铖.基于MacBERT和联合注意力增强网络的物业服务投诉分类方法[J].广西科学,2024,31(1):110-118. [点击复制]
- ZHAN Zhihong,QIN Kaixian,PENG Linghua,ZHAN Cheng.Classification Method of Property Service Complaints Based on MacBERT and Joint Attention Enhancement Networks[J].Guangxi Sciences,2024,31(1):110-118. [点击复制]
|
|
| 摘要: |
| 基于人工的物业投诉文件分类处理方法已经无法满足社会需求，并且已有投诉相关的自动分类方法在物业投诉分类问题上的性能较不足。因此，本研究提出一个基于MacBERT和联合注意力增强网络的物业服务投诉分类方法JAE-BERT4Com。JAE-BERT4Com使用基于近义词替换与合成少数过采样技术结合的样本增强策略解决类不平衡的问题，以及基于MacBERT的分层注意力、Transformers的多头注意力和关键词注意力等多重注意力联合增强的网络进行文本特征学习和分类。实验结果表明，JAE-BERT4Com能够获得比现有模型更高的准确率、F1分数和召回率，比现有较先进模型的性能更优。 |
| 关键词: 物业投诉 投诉分类 文本分类 注意力增强 深度学习 |
| DOI：10.13656/j.cnki.gxkx.20240417.011 |
| 投稿时间：2023-12-07修订日期：2024-01-04 |
| 基金项目:国家自然科学基金项目(62366011)资助。 |
|
| Classification Method of Property Service Complaints Based on MacBERT and Joint Attention Enhancement Networks |
|
ZHAN Zhihong1, QIN Kaixian2, PENG Linghua1, ZHAN Cheng3
|
| (1.Information Center of Housing and Urban-rural Development of Guangxi Zhuang Autonomous Region, Nanning, Guangxi, 530028, China;2.School of Computer and Information Engineering, Nanning Normal University, Nanning, Guangxi, 530001, China;3.Xi'an Jiaotong-Liverpool University, Suzhou, Jiangsu, 215028, China) |
| Abstract: |
| The manual-based classification method of property complaint documents has been unable to meet the needs of the society,and the existing automatic classification methods related to complaints have insufficient performance in the classification of property complaints.Therefore,this study proposes a property service complaint classification method JAE-BERT4Com based on MacBERT and joint attention enhancement network.JAE-BERT4Com uses a sample enhancement strategy based on the combination of synonym replacement and synthetic minority oversampling technology to solve the problem of class imbalance.And a multi-attention joint enhancement network based on MacBERT's hierarchical attention,Transformers' multi-head attention and keyword attention is designed to perform text feature learning and classification.The experimental results show that JAE-BERT4Com can obtain higher accuracy,F1 score and recall rate than the existing models,and has better performance than the existing advanced models. |
| Key words: property complaints complaint classification text categorization attention enhancement deep learning |
