预训练模型下航天情报实体识别方法

为了快速处理航天情报,基于数据驱动的深度学习技术,提出融合多源异构知识标注中文航天情报数据集的方法流程,以及基于预训练(pre-training)模型的航天情报实体识别(AIER)方法;通过对航天情报进行命名实体识别,达到对航天情报进行信息抽取的目的.通过融合BERT(bidirectional encoder representation from transformers)预训练模型和条件随机场(CRF)模型构建AIER模型(BERT-CRF模型),将其与隐马尔可夫模型(HMM)、条件随机场(CRF)模型、双向长短期记忆网络加条件随机场(BiLSTM-CRF)模型进行实体识别对比实验.结果表明:基于预训练模型的AIER模型能够取得93.68%的准确率、97.56%的召回率和95.58%的F₁值;相比于其他方法,基于预训练模型方法的性能得到提高.

Aerospace Intelligence Entity Recognition Method Based on Pre-Training Model

In order to quickly process aerospace intelligence, based on a data-driven deeplearning technology, a method of fusing multi-source heterogeneous knowledge to label Chinese aerospace intelligence data sets is proposed, and the aerospace intelligence entity recognition(AIER)method based on pre-training models is formed. Through the identification of named entities for aerospace intelligence, the purpose of information extraction for aerospace intelligence is achieved. This paper aims to construct the AIER model(BERT-CRF model)by fusing the bidirectional encoder representations from transformers(BERT)pre-training model and the conditional random field(CRF)model, and combine it with the hidden Markov model(HMM)and CRF model, bidirectional long short-term memory network plus conditional random field(BiLSTM-CRF model)model for entity recognition contrast experiments. The results show that the AIER model based on the pre-training model can achieve 93.68% accuracy, 97.56% recall rate and 95.58% F₁ value; compared with other methods, the pre-training model method is much improved on performance.