Journal of Frontiers of Computer Science and Technology ›› 2021, Vol. 15 ›› Issue (2): 206-218.DOI: 10.3778/j.issn.1673-9418.2003049

• Surveys and Frontiers • Previous Articles     Next Articles

Review of Transfer Learning for Named Entity Recognition

LI Meng, LI Yanling, LIN Min   

  1. College of Computer Science and Technology, Inner Mongolia Normal University, Hohhot 010022, China
  • Online:2021-02-01 Published:2021-02-01

命名实体识别的迁移学习研究综述

李猛李艳玲林民   

  1. 内蒙古师范大学 计算机科学技术学院,呼和浩特 010022

Abstract:

Named entity recognition (NER) is one of the core application tasks of natural language processing. Traditional and deep NER methods rely heavily on a large amount of labeled training data with the same distri-bution, and the portability of the model is very poor. However, data are small and personalized in practical appli-cations, and it is very difficult to collect enough training data. Transfer learning is used in NER, the data and model of the source domain are utilized to complete the target task model construction, increase the amount of labeled data in the target domain and reduce the demand for the amount of labeled data of the target model. It has a very good effect in dealing with the task of low-resource NER. Firstly, the method and difficulty of NER and the method of transfer learning are summarized. Then, transfer learning methods applied to NER including data-based transfer learning, model-based transfer learning and adversarial transfer learning in recent years are comprehensively reviewed, and adversarial transfer learning methods are mainly described. Finally, this paper further expounds the current problems and looks forward to the future research directions.

Key words: named entity recognition (NER), transfer learning, adversarial transfer learning, deep learning

摘要:

命名实体识别(NER)是自然语言处理的核心应用任务之一。传统和深度命名实体识别方法严重依赖于大量具有相同分布的标注训练数据,模型可移植性差。然而在实际应用中数据往往都是小数据、个性化数据, 收集足够的训练数据是非常困难的。在命名实体识别中引入迁移学习,利用源域数据和模型完成目标域任务模型构建,提高目标领域的标注数据量和降低目标域模型对标注数据数量的需求,在处理资源匮乏命名实体识别任务上,具有非常好的效果。首先对命名实体识别方法和难点以及迁移学习方法进行概述;然后对近些年应用于命名实体识别的迁移学习方法,包括基于数据迁移学习、基于模型迁移学习和对抗迁移学习,进行全面综述,重点阐述了对抗迁移学习方法;最后进一步思考当前存在的问题并对未来的研究方向进行了展望。

关键词: 命名实体识别(NER), 迁移学习, 对抗迁移学习, 深度学习