摘要
近年来,生物恐怖威胁已成为国家安全的重大挑战,准确快速地识别生物恐怖威胁信息并对其进行分类成为亟待解决的关键问题。然而,传统的文本分类技术在应对生物恐怖威胁时面临数据稀缺和威胁因子复杂的问题。为此,本文提出了一种基于提示学习的零样本文本分类方法,设计了基于掩码策略的MaskBERT模型,并集成了提示插入模块和提示匹配模块。该方法利用预训练语言模型的知识,无须依赖外部知识库,成功实现了文本与类别的有效匹配,提高了分类的准确性和语义丰富性。在生物恐怖威胁信息指纹数据集上进行的对比实验和消融实验表明,本文提出的模型在准确率、召回率和F1值上分别达93.4%、92.3%和92.1%。相较于传统文本分类模型BERT、FPT-BERT、DepRNN、CPFT、CNN-BERT、SN-FT和HGAT,本模型对不同生物恐怖威胁信息的文本分类准确率更高,表明其具有良好的分类性能,能够准确而全面地识别生物恐怖威胁信息。
In recent years,biological terrorism has emerged as a significant challenge to national security.Accurate,rapid identification and classification of biological terrorism information has become a critical issue that urgently needs to be addressed.However,traditional text classification techniques face challenges such as data scarcity and the complexity of threat factors when dealing with biological terrorism threat.To address these issues,a zero-shot text classification method based on prompt learning was proposed,a Mask BERT model based on a masking strategy was designed,and integrated prompt insertion and prompt matching modules were introduced.This approach leveraged the knowledge of pre-trained language models without relying on external knowledge bases.It successfully achieved effective matching between text and categories,enhancing classification accuracy and semantic richness.Comparative and ablation experiments on the bioterrorism threat information fingerprint dataset showed that the model proposed in this paper achieved precision,recall,and F1 scores of 93.4%,92.3%and 92.1%,respectively.Compared to classic text classification models such as BERT,FPT-BERT,DepRNN,CPFT,CNN-BERT,SN-FT and HGAT,the model proposed in this paper demonstrated higher classification accuracy for various bioterrorism threat information texts,indicating that this model had excellent classification performance and could accurately and comprehensively identify bioterrorism threat information.
作者
吴龙涛
黄李洲
黄凰
施加松
WU Longtao;HUANG Lizhou;HUANG Huang;SHI Jiasong(Chemical Defense Institute,Academy of Military Sciences,Beijing 102205,China)
出处
《防化研究》
2024年第3期63-71,共9页
CBRN DEFENSE
基金
国家重点研发计划项目(2021YFC2600501)。
关键词
零样本学习
文本分类
提示学习
BERT
生物威胁
zero-shot learning
text classification
prompt learning
BERT
biological threat
