基于BERT模型的暗网犯罪情报挖掘技术研究  

Research on Dark Web Crime Intelligence Mining Technology Based on BERT Model

作　　者：周宇[1] 蔡都 ZHOU Yu;CAI Du(Jiangsu Provincial Public Security Department,Nanjing 210024,China)

机构地区：[1]江苏省公安厅,江苏南京210024

出　　处：《现代信息科技》2024年第23期165-169,174,共6页Modern Information Technology

摘　　要：实现暗网违法犯罪情报的规模化产出是打击暗网违法犯罪的一项重要前置任务。当前研究较难解决暗网数据量不足的问题,且主要针对西文暗网数据进行。为实现中文暗网文本的针对性分析,提出了一种基于多任务学习的BERT-BiLSTM违法犯罪分类和命名实体识别多任务学习模型,其在文本分类和命名实体识别任务间共享BERT-BiLSTM层,并分别采用全连接层和条件随机场(CRF)层作为文本分类和实体识别的输出层,以加强不同任务间的知识共享。在自建的中文暗网数据集上的实验结果表明,该多任务学习模型相比基线模型在两类任务上均有一定性能提升。Achieving the large-scale production of illegal and criminal intelligence on the dark web is a crucial preliminary task for combating illegal and criminal activities on the dark web.Current research struggles to address the issue of insufficient dark web data and primarily focuses on Western language dark web data.In order to achieve targeted analysis of Chinese dark web texts,this paper proposes a multi-task learning model for BERT-BiLSTM illegal and criminal classification and named entity recognition based on multi-task learning.It shares the BERT-BiLSTM layer between the text classification and named entity recognition tasks,and adopts the fully connected layer and the Conditional Random Field(CRF)layer as the output layers for text classification and entity recognition respectively,so as to strengthen knowledge sharing between different tasks.The experimental results on the self-constructed Chinese dark web dataset show that,compared with the baseline model,this multi-task learning model achieves certain performance improvements in both types of tasks.

关 键 词：暗网 犯罪治理 多任务学习 BERT 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP391[自动化与计算机技术—控制科学与工程]