基于自注意力的迁移表示学习的国家关系研判  

State Relations Research and Judgment Based on Self-Attention Mechanism of Transfer Representation Learning

作　　者：刘卫平 潘仁前 陈伟荣[1] 张豹 LIU Weiping;PAN Renqian;CHEN Weirong;ZHANG Bao(The 28th Research Institute of China Electronics Technology Group Corporation,Nanjing 210007,China)

机构地区：[1]中国电子科技集团公司第二十八研究所,南京210007

出　　处：《指挥信息系统与技术》2021年第3期83-90,共8页Command Information System and Technology

基　　金：装备发展部“十三五”预研课题资助项目。

摘　　要：针对军事领域国家关系研判的特殊性以及标注文本不足等问题,结合最新文本编码解码算法,提出了一种基于双向语义表征模型(BERT)和双向长短时记忆网络(Bi-LSTM)的国家关系研判算法。该算法以字、字位置、语义块及词性作为输入特征,通过军事领域迁移表示学习编码模块和军事领域特征解码模块,实现国家关系研判。首先,通过BERT算法编码字、字位置及语义块特征,得到军事文本通用领域语义编码向量;然后,通过Word2Vec算法编码词性特征,构建空间映射层以消除空间不一致性,实现文本统一编码输出;最后,通过军事领域特征解码模块实现了文本特征解码,并叠加分类层以完成国家关系研判。试验结果表明,该算法在准确率、召回率和F1值上表现较佳。Aimed at the particularity of the research and judgment of state relations in the military field and the shortage of annotated texts,combined with the latest text encoding and decoding algorithm,the state relations research and judgment algorithm based on the bi-directional semantic representation model(BERT)and the bi-directional long short-term memory(Bi-LSTM)is proposed.This algorithm takes the character,the word position,the semantic block and the part-of-speech as input features,realizes the research and judgment of state relations through the coding module of transfer representation learning and the decoding module of military domain features.Firstly,the general domain semantic coding vector of military text is obtained through the encoding words,the word position and the semantic block features of BERT algorithm.Then,based on the part-of-speech features of Word2 Vector algorithm,the spatial mapping lager is constructed to eliminate spatial inconsistency,and the unified text encoding output is achieved.Finally,the text features decoding is realized through the military domain feature decoding module,and the classification layer is superimposed to complete the research and judgment of state relations.The experimental results show that this algorithm performs better in the accuracy,the recall rate and the F1 value.

关 键 词：国家关系 研判 迁移表示学习 双向长短时记忆网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]