融合领域特征向量的武器装备名深度学习识别方法

MILITARY EQUIPMENT NAME DEEP LEARNING RECOGNITION WITH DOMAIN FEATURE VECTORS

提出融合领域特征向量与词向量的识别方法,将基于武器装备名特征库与维基语料训练得到的领域特征向量引入Bi-LSTM+CRF模型,并对武器装备名进行自动识别实验。引入领域特征向量后模型的识别准确率由78.30%提升到82.10%,召回率由65.25%提升到67.30%,对未登录武器装备名识别的召回率从45.08%提升到50.16%。此外,将领域特征融入条件随机场(conditionalrandomfield,CRF)模型,实验表明,在小规模语料库与领域特征支持的情况下,CRF模型的效果要优于Bi-LSTM+CRF模型且对稀疏特征的利用效率更优。

This paper proposed the recognition method of fusion domain feature vectors and word vectors.The domain feature vectors based on weapon equipment name feature library and Wikipedia corpus training were introduced into Bi-LSTM+CRF model,and the automatic recognition experiment of weapon equipment names was carried out.After introducing domain feature vectors,the recognition accuracy of the model is improved from 78.30% to 82.10%,the recall rate is increased from 65.25% to 67.30%,and the recall rate of the unlisted weapon equipment name recognition is increased from 45.08% to 50.16%.In addition,domain features were integrated into conditional random field(CRF) model.Experiments show that CRF model outperforms Bi-LSTM+CRF model in the case of small-scale corpus and domain features support,and the efficiency of using sparse features is better.