[1]蔡惠民*,印忠文,岳世彬.基于异构图神经网络的检务知识咨询业务分类[J].计算机技术与发展,2022,32(10):201-208.[doi:10. 3969 / j. issn. 1673-629X. 2022. 10. 033]
　CAI Hui-min*,YIN Zhong-wen,YUE Shi-bin.Heterogeneous Graph Neural Network Based Business Classification for Procuratorial-knowledge Query[J].,2022,32(10):201-208.[doi:10. 3969 / j. issn. 1673-629X. 2022. 10. 033]
点击复制

基于异构图神经网络的检务知识咨询业务分类()
分享到：

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:
32
期数:
2022年10期
页码:
201-208
栏目:
新型计算应用系统
出版日期:
2022-10-10

文章信息/Info

Title:
Heterogeneous Graph Neural Network Based Business Classification for Procuratorial-knowledge Query
文章编号:
1673-629X(2022)10-0201-08
作者:
蔡惠民1; 2* ; 印忠文1; 2 ; 岳世彬1; 2
1. 中电科大数据研究院有限公司,贵州 贵阳 550022;
2. 提升政府治理能力大数据应用技术国家工程实验室,贵州 贵阳 550022
Author(s):
CAI Hui-min1; 2* ; YIN Zhong-wen1; 2 ; YUE Shi-bin1; 2
1. CETC Big Data Research Institute Co. ,Ltd. ,Guiyang 550022,China;
2. Big Data Application on Improving Government Governance Capabilities National Engineering Laboratory,Guiyang 550022,China
关键词:
智慧检务; 异构图神经网络; 意图识别; 业务分类; 检答网
Keywords:
wisdom procuratorial affairs; heterogeneous graph neural network; intention recognition; business classification; Jianda-Net
分类号:
TP391
DOI:
10. 3969 / j. issn. 1673-629X. 2022. 10. 033
摘要:
随着检察机关办理的案件日益增多,用户对检务领域知识的咨询需求逐年增大,传统依靠检察机关领域专家人工解答回复的方式难以应对大规模的咨询服务。 为了提高用户咨询服务效率,提升计算机正确理解用户提问意图的能力,提出了一种面向检务知识咨询的异构图神经网络业务类型分类模型。 该模型以基于句法依存分析的图表示、基于邻域窗口的图表示作为输入,分别以 RGCN 和 GAT 图神经网络作为特征编码器,并通过特征融合实现用户提问内容业务类型预测。 同时引入辅助分类器优化特征编码器的学习并提升模型性能,并采用 Focal Loss 损失函数有效解决了样本数据不均衡问题。 此外,该模型与传统深度学习文本分类模型、目前主流的 BERT 模型在宏平均准确率、模型大小、推理时间等多个维度进行性能对比。 对比实验显示,该模型在测试集上的宏平均准确率均优于文本分类基准模型,模型大小和推理时间远小于 BERT 模型。
Abstract:
With the increasing of cases handled by the procuratorate, the user’s demand for consulting knowledge in the field of procuratorial affairs? ? is raised year by year. However, it is difficult to deal with the large - scale consulting services by relying on the experts in procuratorate to answer and reply manually. In order to improve the efficiency of user consultation and enhance the ability of computer to correctly understand the user’s query, a heterogeneous graph neural network model is proposed for business type classification. This model uses graph representations based on both syntactic dependency analysis and neighbor window as inputs. RGCNand GAT graph neural networks are used as feature encoders respectively,and feature fusion is applied to predict the business type ofuser’s query. Meanwhile,auxiliary classifiers are introduced to optimize the learning of feature encoders and improve the performance of the model. Moreover,Focal Loss is applied to effectively address the issue of unbalanced data. Furthermore, the performance of this model is compared with the traditional deep learning text classification models and the current mainstream BERT model in many dimensions, such as macro average accuracy, model size,reasoning time. Comparative experiments show that the macro average accuracyof the proposed model on the test set is better than other benchmark text classification models. The model size and reasoning time of the proposed model are much smaller than BERT model.
更新日期/Last Update: 2022-10-10