开源文本中军事目标动向事件抽取方法研究  

Research on extracting military-target motion events from open-source texts

作　　者：张宇恒 郑胜[1] 陈晓玥 ZHANG Yuheng;ZHENG Sheng;CHEN Xiaoyue(School of Electrical and Information Engineering,Wuhan Institute of Technology,Wuhan 430205,China;School of Information Engineering,Hubei University of Economics,Wuhan 430205,China)

机构地区：[1]武汉工程大学电气信息学院,湖北武汉430205 [2]湖北经济学院信息工程学院,湖北武汉430205

出　　处：《武汉工程大学学报》2024年第3期299-303,共5页Journal of Wuhan Institute of Technology

基　　金：湖北省教育厅基金青年项目(Q20222203)。

摘　　要：从海量非结构化的开源军事目标动向文本中抽取指定军事目标的事件信息以及运动轨迹,是识别和预测军事目标的行动意图、挖掘战场动态信息的基础工作。针对目前事件抽取研究中忽略地点论元之间空间关系信息从而导致无法抽取移动目标的运动轨迹问题,提出划分细粒度空间关系标签的方法来识别空间关系,通过序列标注方法进行事件抽取,使用预训练语言模型进行底层语义编码、双向长短时记忆网络进行深层次特征提取、条件随机场进行标签分类的联合事件抽取模型以完成动向事件抽取。在动向事件抽取结果的基础上,使用运动轨迹抽取算法来加强空间关系信息。通过在自建的真实军事目标动向新闻数据集上进行实验,获取了84.0%的F1分数值。The extraction of event information and motion trajectory of designated military targets from massive unstructured open-source texts on military target motions constitutes a fundamental undertaking in identifying and forcasting the operational intentions of such targets and in mining battlefield dynamic information.Aiming at the problem that the current event extraction research ignores the spatial relationship information between place arguments,which leads to the inability to extract the trajectory of moving targets,we proposed a method of dividing fine-grained spatial relationship labels to identify spatial relationships,and extracting events through sequence labeling.A joint event extraction model that uses the pre-trained language model for underlying semantic encoding,bidirectional long short-term memory network for deep feature extraction,and conditional random fields for label classification to achieve motion event extraction.On the basis of the motion event extraction results,the motion trajectory extraction algorithm was used to enhance the spatial relationship information.Through experiments on the self-built real military target motion news dataset,the F1 score value of 84.0%was obtained.

关 键 词：开源军事情报 事件抽取 空间关系识别 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]