基于BERT-BILSTM-LSTM的民航安全信息关系提取方法

EXTRACTION OF KEY INFORMATION OF CIVIL AVIATION SAFETY BASED ON BERT-BILSTM-LSTM

摘要: 为了提升民航安全信息实体的识别准确度和效率，提出一种BERT-BILSTM-LSTM实体关系联合提取模型。利用BERT预训练模型将文本数据转换成词向量；通过BiLSTM（bidirectional long short-term memory），结合上下文语境信息学习词向量之间的关系，使用LSTM（Long short-term memory）网络训练不同词向量的关系得到实体概率；实现对民航文本信息实体的关系提取。实验结果表明，此模型在小规模数据上相较于现存主流算法BERT-BILSTM-CRF在准确率和召回率上都有较大的性能提升。

Abstract: In order to improve the recognition accuracy and efficiency of civil aviation safety information entities, a BERT-BILSTM-LSTM entity relationship joint extraction model is proposed. The BERT pre-training model was used to convert text data into word vectors. The relationship between word vectors was learned by BiLSTM (Bidirectional long short-term memory) combined with contextual information. LSTM network was used to train the relationship between different word vectors to get entity probability. The relationship extraction of civil aviation text information entities was realized. Experimental results show that this model has a great performance improvement in accuracy and recall compared with the existing mainstream algorithm BERT-BILSTM-CRF on small-scale data.