面向武器装备领域的复杂三元组抽取方法  

Complex Triplet Extraction Method for Weapon Equipment Field

作　　者：游新冬 刘陌村 葛昊杰 肖刚 吕学强[1] YOU Xindong;LIU Mocun;GE Haojie;XIAO Gang;L Xueqiang(Beijing Key Laboratory of Internet Culture and Digital Dissemination Research,Beijing Information Science and Technology University,Beijing 100101,China;General Key Laboratory of Complex System Simulation,Systems Engineering Research Institute,Academy of Military Sciences,Beijing 100101,China)

机构地区：[1]北京信息科技大学网络文化与数字传播北京市重点实验室,北京100101 [2]军事科学院系统工程研究院复杂系统仿真总体重点实验室,北京100101

出　　处：《小型微型计算机系统》2024年第3期521-528,共8页Journal of Chinese Computer Systems

基　　金：国防科技重点实验室基金项目(6412006200404)资助;国家自然科学基金项目(62171043)资助.

摘　　要：为解决武器装备领域中单实体重叠和实体对重叠的复杂三元组的抽取问题,提出了挂载武器装备领域知识结合多轮对抗攻击的复杂三元组抽取方法(RDA),该方法通过武器装备领域微调后的Bert获取更具领域语义的文本向量;利用在嵌入层发起多轮对抗的方式,实现模型层面的数据增强,减少模型对标注样本规模的依赖;采用单层指针网络获取头实体对头实体的类别进行判定,利用维基百科知识库对武器装备领域的实体类别解释信息的向量,对武器装备类别信息以字为最小粒度进行融合,缓解分层标注的天然缺陷;最后在横纵两个维度基于不同粒度的序列标注实现复杂三元组的抽取.在武器装备领域的数据集上精准率达到88.54%,召回率达到75.88%,F1值达到81.72%,取得了SOTA效果.实验表明提出的RDA方法对武器装备领域的信息利用更加充分,有效地缓解武器装备领域遇到的单实体重叠问题(SEO)和实体对重叠(EPO)问题.To solve the problem of extracting complex triples with single entity overlap and entity-to-overlap in the weaponry domain,Relation extraction in Domain of weaponary combined with multi-round Adversarial(RDA)is proposed,which obtains more domain semantic text vectors through fine-tuned Bert in the weaponry domain;uses the initiation of multi-round adversarial at the embedding layer to achieve model-level data enhancement The method uses a single-layer pointer network to obtain the head entity to head entity category determination and a vector of entity category interpretation information of the weaponry domain using the Wikipedia knowledge base to fuse the weaponry category information at the smallest granularity of words to alleviate the natural shortcomings of hierarchical annotation;finally,a complex triad is achieved in both horizontal and vertical dimensions based on sequence annotation of different granularities The extraction of complex triples in both horizontal and vertical dimensions is achieved based on sequence annotation at different granularities.The SOTA effect is achieved with an accuracy rate of 88.54%,a recall rate of 75.88%,and an F1 value of 81.72%on the weapon and equipment domain dataset.The experiments show that the proposed RDA method makes fuller use of information in the weaponry domain and effectively alleviates the single entity overlap problem(SEO)and entity pair overlap(EPO)problem encountered in the weaponry domain.

关 键 词：三元组抽取 武器装备领域 复杂命名实体识别 单层指针网络 多轮对抗攻击 RDA 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]