基于混合式迁移学习的命名实体识别算法

NAMED ENTITY RECOGNITION ALGORITHM BASED ON MIXED TRANSFER LEARNING

针对命名实体识别领域中大量标注数据难于获取而带来的问题,提出基于混合式迁移学习的命名实体识别算法——MT-NER。利用样本之间的距离作为权衡样本相似性的标准,进行样本迁移以扩充目标域样本;利用模型迁移建立带有finetune的新命名实体识别网络结构,用扩充后的目标域数据集来训练网络。以医疗领域为例的实验结果分析表明,MT-NER算法在小样本数据中的实体识别效果最佳,精度达到93.31%,召回率达到89.5%,F1值达到0.9317,与BiLSTM-CRF模型相比分别提升了6.33百分点、3.65百分点和0.0891。

In the field of named entity recognition,it is difficult to obtain a large number of labeled data.To solve this problem,this paper proposes a named entity recognition algorithm based on mixed transfer learning named MT-NER.The distance between the samples was used as the criterion to balance the similarity of the samples,and the instances-based transfer learning was carried out to expand the target domain samples.A new named entity recognition network structure with finetune was established by the models-based transfer learning,and the expanded target domain data set was used to train the network.Taking the medical field as an example,experiments show that MT-NER algorithm has the best effect in entity recognition in small sample data,with an accuracy of 93.31%,a recall rate of 89.5%and a F1 value of 0.9317.Compared with the BiLSTM-CRF model,the accuracy,recall rate and F1 value of MT-NER are improved by 6.33,3.65 and 8.91 percentage points.