计算机科学与探索 ›› 2024, Vol. 18 ›› Issue (8): 1998-2013.DOI: 10.3778/j.issn.1673-9418.2311033

• 前沿·综述 • 上一篇    下一篇

基于逻辑推理的机器阅读理解综述

李晴,李艳玲,董杰,葛凤培,林民   

  1. 1. 内蒙古师范大学 计算机科学技术学院,呼和浩特 010022
    2. 无穷维哈密顿系统及其算法应用教育部重点实验室(内蒙古师范大学),呼和浩特 010022
    3. 北京邮电大学 图书馆,北京 100876
  • 出版日期:2024-08-01 发布日期:2024-07-29

Survey of Machine Reading Comprehension Based on Logical Reasoning

LI Qing, LI Yanling, DONG Jie, GE Fengpei, LIN Min   

  1. 1. College of Computer Science and Technology, Inner Mongolia Normal University, Hohhot 010022, China
    2. Key Laboratory of Infinite-Dimensional Hamiltonian System and Its Algorithm Application (Inner Mongolia Normal University), Ministry of Education, Hohhot 010022, China
    3. Library, Beijing University of Posts and Telecommunications, Beijing 100876, China
  • Online:2024-08-01 Published:2024-07-29

摘要: 机器阅读理解是自然语言处理领域中的核心任务之一,该任务目标是使机器能够理解自然语言文本,并正确回答关于文本内容的问题。随着自然语言处理相关方法和模型的发展,研究者们开始关注机器阅读理解中更具挑战性的推理型问题,这些问题通常要求模型不仅理解文本中的浅层信息,还要能够在逻辑上进行思考和推理,以回答更加复杂的问题。对基于逻辑推理的机器阅读理解相关的最新成果进行全面的归纳。介绍基于逻辑推理的机器阅读理解任务。介绍该任务的相关方法,并根据侧重点的不同将这些方法分成四类:基于符号神经网络的方法、基于图神经网络的方法、基于预训练的方法和基于大模型的微调策略。重点描述四类方法的代表性工作。在LogiQA和ReClor两个逻辑推理主流数据集上探讨方法的优缺点,并总结基于逻辑推理的机器阅读理解任务的未来研究方向。

关键词: 机器阅读理解, 逻辑推理, 智能问答

Abstract: Machine reading comprehension is one of the core tasks in the field of natural language processing, the goal of which is to enable machines to understand natural language text and correctly answer questions about the content of the text. With the development of methods and models related to natural language processing, researchers have begun to focus on the more challenging inferential problems in machine reading comprehension, which often require models to not only understand superficial information in text, but also to be able to think and reason logically to answer more complex questions. This paper gives a comprehensive summary of the latest achievements of machine reading comprehension based on logical reasoning. Firstly, the task of machine reading comprehension based on logical reasoning is introduced. Secondly, this paper introduces relevant methods for this task and classifies them into four categories based on different focuses: methods based on symbolic neural networks, methods based on graph neural networks, methods based on pre-training, and fine-tuning strategies based on large models, with a focus on describing the representative work of the four methods. Finally, the advantages and disadvantages of the methods are discussed on LogiQA and ReClor datasets, and the future research directions of machine reading comprehension task based on logical reasoning are summarized.

Key words: machine reading comprehension, logical reasoning, intelligent questions and answers