基于RoBERTa-Span-Attack的标签指针网络军事命名实体识别  

Military named entity recognition based on RoBERTa-Span-Attack tag pointer network

作　　者：罗兵[1] 张显峰[1] 段立[1] 陈琳 LUO Bing;ZHANG Xianfeng;DUAN Li;CHEN Lin(College of Electronic Engineering,Naval Univ.of Engineering,Wuhan 430033,China)

机构地区：[1]海军工程大学电子工程学院,武汉430033

出　　处：《海军工程大学学报》2024年第1期76-82,93,共8页Journal of Naval University of Engineering

摘　　要：军事领域文本中存在大量军事实体信息,准确识别这些信息是军事文本信息提取和构建军事知识图谱的基础性任务。首先,提出了一种基于RoBERTa预训练模型、跨度和对抗训练的标签指针网络的融合深度模型(RoBERTa-Span-Attack),用于中文军事命名实体识别;然后,采用了一种基于Span的标签指针网络,同时完成实体的起止位置和类别的识别任务;最后,在模型训练过程中加入对抗训练策略,通过添加一些扰动来生成对抗样本进行训练。在军事领域数据集上的实验结果表明:所提出的军事领域命名实体识别模型相较于BERT-CRF、BERT-Softmax和BERT-Span,在识别准确度上具有更优的效果。There are plenty of military entities in the documents of military field.Identification of such information is the basic task of extracting military text information and constructing military know-ledge graph.A model based on robustly optimized BERT pre-training approach(RoBERTa) Span and confrontation training label pointer network(RoBERTa-Span-Attack) was proposed,which was used for Chinese military named entity recognition.Because RoBERTa adopts the pre-training strategy of whole word mask,it has learned the semantic representation for the whole word,which is more sui-table for the recognition of Chinese military named entities.And then,a span-based label pointer network which can recognize the starting-end position and label of entities at the same time was adopted to improve the model performance.Finally,adversarial training strategy in which disturbances were added to generate adversarial samples for training process was employed to improve the robustness of the model.Experimental results on military domain dataset demonstrate that the proposed model has better recognition accuracy than BERT-CRF,BERT-Softmax and BERT-Span.

关 键 词：军事命名实体识别 预训练模型 跨度 标签指针网络 对抗训练 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]