结合双层多头自注意力和BiLSTM-CRF的军事武器实体识别  

Military Weapon Entity Recognition Combined with Double-layer Multi-head Self-attention and BiLSTM-CRF

作　　者：俞海亮 彭冬亮[2] 谷雨[2] YU Hailiang;PENG Dongliang;GU Yu(HDU-ITMO Joint Institute,Hangzhou Dianzi University,Hangzhou 310018,China;School of Automation,Hangzhou Dianzi University,Hangzhou 310018,China)

机构地区：[1]杭州电子科技大学圣光机联合学院,浙江杭州310018 [2]杭州电子科技大学自动化学院,浙江杭州310018

出　　处：《无线电工程》2022年第5期775-782,共8页Radio Engineering

基　　金：浙江省自然科学基金(LY21F030010)。

摘　　要：军事武器实体识别是军事领域本体构建的一项重要任务,基于深度学习方法实现自动军事武器实体识别能够提升军事情报信息检索的效率。为提升军事武器实体识别的精确率,面向网络公开非结构化军事新闻数据,提出了一种结合双层多头自注意力机制和BiLSTM-CRF模型的武器实体识别方法。在BiLSTM-CRF模型的基础上,采用双层自注意力机制,分别在嵌入层提取重要输入特征以及BiLSTM层提取关键字符信息,并结合军事武器实体构词特点,建立正则匹配模板对识别结果进行校正。构建了包含1196条数据的军事武器数据集,测试结果表明,提出方法的精确率、召回率和F1值分别为0.9293,0.9301和0.9297,相比于经典深度学习模型的最优结果,在精确率、召回率以及F1值上分别提升了1.15%,0.97%和0.97%。Military weapon entity recognition is an important task of ontology construction in military field.The realization of automatic military weapon entity recognition based on deep learning method can improve the efficiency of military intelligence information retrieval.To improve the accuracy of military weapon entity recognition,a weapon entity recognition method combined with double-layer multi-head self-attention mechanism and BiLSTM-CRF model is proposed,which can be used for the unstructured military weapon news data published on the Internet.Based on the BiLSTM-CRF model,this method uses double-layer self-attention mechanism to extract important input features in the embedded layer and key character information in the BiLSTM layer respectively.Combined with the word-building characteristics of military weapon entities,a regular matching template is established to correct the recognition results.A military weapon data set containing 1196 pieces of data is constructed.The test results show that the accuracy rate,recall rate and F1 value of the proposed method are 0.9293,0.9301 and 0.9297 respectively,which,as compared with the best results using classical deep learning model,are improved by 1.15%,0.97%and 0.97%respectively.

关 键 词：武器实体识别 双层多头自注意力 BiLSTM-CRF 校正 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]