融合深度主动学习的军事物流实体识别研究  

Research on military logistics entity recognition based on deep active learning

作　　者：罗少锋 詹威威 王静 陈可夫 LUO Shao-feng;ZHAN Wei-wei;WANG Jing;CHEN Ke-fu(Systems Engineering Institute,AMS,PLA,Beijing 100166,China;Troila Technology Development Co.,Ltd.,Tianjin 300133,China)

机构地区：[1]军事科学院系统工程研究院,北京100166 [2]天津卓朗科技发展有限公司,天津300133

出　　处：《信息技术》2024年第12期115-123,共9页Information Technology

基　　金：军队后勤科研项目(BS118R002)。

摘　　要：针对军事领域实体识别模型训练需要大量标注语料,且开源语料数据缺乏等问题,提出一种融合主动学习与深度学习的军事物流需求实体识别模型训练方法。将不同选样策略的主动学习方法融入BERT-BiLSTM-CRF模型训练中,通过筛选有价值的训练样本,快速提升军事物流需求实体识别模型的泛化能力,同时也减少了样本标注的数量,实现使用少量标注语料训练模型。实验表明,在保证模型性能的前提下,不确定性和多样性的选样策略分别使人工标注量减少56.25%、62.5%,有效降低模型训练所需样本量,对语料缺少的军事物流领域需求实体识别任务具有重要意义。To solve the problems that a large amount of annotated corpus is required for entity recognition model training in military domain and the lack of open source corpus data,this paper proposes a training method for entity recognition model of military logistics requirements that integrates active learning and deep learning.By incorporating active learning methods with different sample selection strategies into BERT-BiLSTM-CRF model training,the generalization ability of the military logistics demand entity recognition model is rapidly improved by screening valuable training samples,and the number of sample annotations is also reduced to achieve a small amount of annotated corpus to train the model.Experiments show that the uncertainty and diversity sample selection strategies reduce the expert annotation by 56.25%and 62.5%,respectively,under the premise of guaranteeing the model performance,effectively reducing the sample size required for model training,which is important for the demand entity recognition task in the military logistics domain where the corpus is lacking.

关 键 词：军事物流 BERT 实体识别 主动学习 深度学习 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术] E95[自动化与计算机技术—计算机科学与技术]