摘要
针对军事情报分析领域难以快速准确抽取军事目标活动相关属性和事件要素问题,本文提出一种基于层级式双向-长短期记忆神经网络-条件随机场(Bi-LSTM-CRF)模型的军事目标实体识别方法。结合军事目标及属性特点,采用树形结构定义层级式目标及属性实体、活动要素及属性实体,细化了实体类别粒度,实现实体识别同时自动关联实体及相关属性。之后依据层级式特点进行军事情报语料标注,将训练好的词向量和训练语料输入Bi-LSTM-CRF模型,其中模型的CRF层依据标签转移条件添加固定约束矩阵,弥补了样本覆盖面补全问题,有效提高实体识别精度。
For the field of military intelligence analysis,it is difficult to quickly and accurately extract the attributes and event elements related to military target activities.We propose a military target entity recognition method based on hierarchical bilateral long-term and short-term memory neural network conditional random field model.Combining military objectives and attributes characteristics,using a tree structure to define hierarchical goals,active elements and their attributes,this entity class granularity is refined to related attributes.After that,the military intelligence is expected to be marked according to the hierarchical characteristics,and the trained word vectors and training are expected to be injected into the model.Among them,the conditional random field layer of the model adds a fixed constraint matrix according to the label transfer condition,which makes up for the problem of complementing the sample coverage and effectively improves the entity recognition.
作者
徐树奎
曹劲然
Xu Shukui;Cao Jinran(The 28th Research Institute of China Electronics Technology group Corporation,Nanjing 210007,China)
出处
《信息化研究》
2019年第6期18-22,46,共6页
INFORMATIZATION RESEARCH
关键词
实体识别
层级式
双向-长短期记忆神经网络-条件随机场
entity recognition
hierarchical
bilateral long-term and short-term memory neural network conditional random field
