基于词嵌入约束和图卷积神经网络的法律案情分类模型构建研究  

Research on the construction of a legal text classification model basedon word embedding constraint and graph convolutional neural network

作　　者：孟春运 谈镇 栾力 ABEO Timothy Apasiba MENG Chunyun;TAN Zhen;LUAN Li;ABEO Timothy Apasiba(School of Economics and Management,Jiangsu University of Science and Technology,Zhenjiang 212100,China;School of Public Affairs,University of Science and Technology of China,Hefei 230026,China;Faculty of Applied Science and Technology,Tamale Technical University,Tamale 00233,Ghana)

机构地区：[1]江苏科技大学经济管理学院,镇江212100 [2]中国科学技术大学公共事务学院,合肥230026 [3]泰米尔技术大学应用科学与技术学院,塔马利00233

出　　处：《江苏科技大学学报(自然科学版)》2025年第2期84-91,共8页Journal of Jiangsu University of Science and Technology:Natural Science Edition

基　　金：国家社会科学基金重点项目(16AJL008)。

摘　　要：图卷积神经网络善于理解文本整体结构和语义信息,可捕获文本数据中的全局关联性,但训练数据的质量会影响其分类性能.因此,提出了基于词嵌入约束和图卷积神经网络的法律案情分类模型,该模型以法律案情数据为基础,构建全局词语共现信息的异构图,输入图卷积网络实现司法案件的归纳和分类.同时,利用预训练好的词嵌入模块提取局部连续词之间的语义相关性信息.通过余弦相似度层对高维词语特征向量进行约束,使全局词语共现和局部语义相关性特征在分类过程中协同作用.在法律文书数据集上,模型在给予数据全部标签和部分标签两种情况下均获得最优F1值,超参数敏感度测定验证了模型结构的合理性,文本分类特征的可视化进一步证实了模型的有效性.Graph convolutional neural networks excel in comprehending text structure and semantics,capturing global correlations.However,their strong reliance on training data poses challenges when data quality is suboptimal.This paper introduces a legal text classification model based on word embedding constraints and graph convolutional neural network.Built on legal case data,the model constructs a heterogeneous graph for global word co-occurrence,utilizing a GCN for case induction and classification.A pre-trained word embedding module extracts semantic relatedness between local words.Constrained by a cosine similarity layer,this information synergistically enhances global co-occurrence and local semantic relatedness during classification.The model achieves optimal F1 values on a legal dataset under both fully and partially labeled scenarios.Sensitivity analyses confirm the model′s rational structure,and visualizations validate its effectiveness in text classification.

关 键 词：深度学习 数据挖掘 文本分类 司法效率 图卷积神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]