计算机工程与应用 ›› 2020, Vol. 56 ›› Issue (23): 12-22.DOI: 10.3778/j.issn.1002-8331.2006-0040

• 热点与综述 • 上一篇    下一篇

自然语言处理预训练模型的研究综述

余同瑞,金冉,韩晓臻,李家辉,郁婷   

  1. 1.浙江万里学院 大数据与软件工程学院,浙江 宁波 315100
    2.浙江大学 计算机科学与技术学院,杭州 310027
  • 出版日期:2020-12-01 发布日期:2020-11-30

Review of Pre-training Models for Natural Language Processing

YU Tongrui, JIN Ran, HAN Xiaozhen, LI Jiahui, YU Ting   

  1. 1.College of Big Data and Software Engineering, Zhejiang Wanli University, Ningbo, Zhejiang 315100, China
    2.College of Computer Science and Technology, Zhejiang University, Hangzhou 310027, China
  • Online:2020-12-01 Published:2020-11-30

摘要:

近年来,深度学习技术被广泛应用于各个领域,基于深度学习的预处理模型将自然语言处理带入一个新时代。预训练模型的目标是如何使预训练好的模型处于良好的初始状态,在下游任务中达到更好的性能表现。对预训练技术及其发展历史进行介绍,并按照模型特点划分为基于概率统计的传统模型和基于深度学习的新式模型进行综述;简要分析传统预训练模型的特点及局限性,重点介绍基于深度学习的预训练模型,并针对它们在下游任务的表现进行对比评估;梳理出具有启发意义的新式预训练模型,简述这些模型的改进机制以及在下游任务中取得的性能提升;总结目前预训练的模型所面临的问题,并对后续发展趋势进行展望。

关键词: 深度学习, 自然语言处理, 预处理, 词向量, 语言模型

Abstract:

In recent years, deep learning technology has been advancing, pre-training technology for deep learning brings natural language processing into a new era. Pre-training model aims to how to make pre-trained model stay in good initial state and achieve better performances in subsequent downstream tasks. This paper firstly introduces pre-training technology and its development history. And then, this paper further classifies it into the following two types, namely probability-statistics-based traditional model and deep-learning-based new model, according to different features of pre-training models to conduct corresponding detailed introductions. Firstly, it briefly analyzes the characteristics and limitations of today’s pre-training models and highlights the existing deep-learning-based pre-training models. And based on their performances in downstream tasks, it gives necessary comparisons and assessments accordingly. Finally, it combs out a series of whole-new pre-training models with instructive significances,briefly describes corresponding feasible improvement mechanisms and the performance enhancements achieved in downstream tasks, summarizes the problems existing therein, as well as prospectes its development trend in near future.

Key words: deep learning, natural language processing, pre-training, word embedding, language model