基于BERT-BiLSTM-CRF模型的电子目标图谱实体抽取方法

Electronic target pattern entity extraction method based on BERT-BiLSTM-CRF model

针对电子目标情报数据种类繁多、关联关系复杂,电子目标图谱存在实体抽取混乱、语义容易发生歧义等问题,将BiLSTM-CRF模型和BERT模型相结合,提出了一种电子目标图谱实体抽取方法.该方法将BERT模型中训练的词向量传递给BiLSTM模型中做特征;然后在CRF模型中得到全局序列排列,实现电子目标图谱的实体抽取.实验结果表明,与Word2Vec和BERT不同字嵌入相比,BERT的字嵌入平均识别率提高3.22%;与BiLSTM、CRF、BiLSTM等不同模型层次相比,BERT-BiLSTM-CRF的平均识别率比其他3种模型最好的平均值还要高2.99%,说明该方法能够进一步提高电子目标相关命名实体识别的效果.

Aiming at the problems of wide variety of electronic target intelligence data,complex correlated relationships,confusion of entity extraction in electronic target atlas and semantics prone to ambiguity,this paper combines BiLSTM-CRF model with BERT model to propose a new entity extraction method for electronic target atlas.This method passes the word vectors trained in the BERT model to the BiLSTM model as features.Then,the global sequence arrangement is obtained in the CRF model to realize the entity extraction of the electronic target map.Experimental results show that compared with Word2Vec and BERT different word embeddings,BERT’s average recognition rate of word embedding is increased by 3.22%,and that compared with BiLSTM,CRF,BiLSTM and other different model levels,the average recognition rate of BERT-BiLSTM-CRF is 2.99%higher than the best average of the other three models,indicating that the proposed method can further improve the recognition effect of the named entities related to electronic targets.