摘要
军事情报中的命名实体识别(NER)是一项至关重要的任务,尤其在面临嵌套实体和少样本学习问题时更具挑战性。为应对上述问题,提出一种基于强化学习的军事情报嵌套命名实体识别方法。首先,采用全局指针的形式将实体的开始和结束位置视为一个整体进行识别,有助于解决嵌套实体的识别问题;然后,通过强化学习优化模型的误差学习策略,使模型能够更好地在有限的训练样本下进行学习;最后,引入元学习的思想,帮助模型从少量的任务中提取通用的知识和模式,进一步增强模型的泛化能力。在军事情报数据集上的试验结果表明,该方法在嵌套实体的抽取效果上相较于现有的基准模型有大幅提升,证明了其有效性和可行性。
Named entity recognition(NER)is a crucial task in military intelligence,especially chal-lenging when confronting nested entities and few-shot learning problems.For the above problems,a military intelligence nested named entity recognition based on reinforcement learning method is pro-posed.Firstly,used in a global pointer form,the start and end positions of an entity are regarded as a whole for recognition,and it is helpful to address the nested entity recognition problem.Then,through the optimization of the model's error learning strategy via reinforcement learning,the model can better learn with limited training samples.Finally,the concept of meta-learning is incorporated to assist the model in extracting general knowledge and patterns from a small number of tasks.Thus,the model's generalization capability is enhanced.Experimental results on a military intelligence dataset show that the effect of the method on extracting nested entities is greatly improved,compared with the existing baseline models.Its effectiveness and feasibility are proved.
作者
方正
蒋明鹏
周高峰
张广庆
FANG Zheng;JIANG Mingpeng;ZHOU Gaofeng;ZHANG Guangqing(CETC Cloud(Beijing)Science&Technology Co.,Ltd,Bejing 100041,China)
出处
《指挥信息系统与技术》
2024年第4期39-45,共7页
Command Information System and Technology
关键词
命名实体识别
强化学习
嵌套命名实体识别
named entity recognition(NER)
reinforcement learning
nested named entity recognition
