基于深度强化学习的知识推理研究进展综述  

Developments of Knowledge Reasoning Based on Deep Reinforcement Learning

作　　者：宋浩楠 赵刚[1] 孙若莹[1] SONG Haonan;ZHAO Gang;SUN Ruoying(School of Information Management,Beijing Information Science&Technology University,Beijing 100192,China)

机构地区：[1]北京信息科技大学信息管理学院,北京100192

出　　处：《计算机工程与应用》2022年第1期12-25,共14页Computer Engineering and Applications

基　　金：国家重点研发计划课题(2019YFB1405003)。

摘　　要：知识推理是知识图谱补全的重要方法,已在垂直搜索、智能问答等多个应用领域发挥重要作用。随着知识推理应用研究的不断深入,知识推理的可解释性受到了广泛关注。基于深度强化学习的知识推理方法具备更好的可解释性和更强的推理能力,能够更加充分地利用知识图谱中实体、关系等信息,使得推理效果更好。简要介绍知识图谱及其研究的基本情况,阐述知识推理的基本概念和近年来的研究进展,着重从封闭域推理和开放域推理两个角度,对当下基于深度强化学习知识推理方法进行了深入分析和对比,同时对所涉及到的数据集和评价指标进行了总结,并对未来研究方向进行了展望。Knowledge reasoning is an important method for the completion of knowledge graphs and has played an important role in many application fields such as vertical search and intelligent question answering.With the continuous in-depth research on the application of knowledge reasoning,the interpretability of knowledge reasoning has received extensive attention.The knowledge reasoning method based on deep reinforcement learning has better interpretability and stronger reasoning ability,which can make full use of the information of entities and relationships in the knowledge graph to make the reasoning effect better.The relevant concepts of knowledge graph and the basic research situation are intro-duced.The basic concepts of knowledge reasoning and the research progress in recent years are illuminated.Focusing on the closed domain reasoning and open domain reasoning,in-depth analysis and comparison of the current knowledge reasoning methods based on deep reinforcement learning are conducted.This paper also summarizes the data sets and evaluation methods involved.Finally,the future research directions are prospected.

关 键 词：知识图谱 知识推理 深度强化学习 链接预测 事实预测 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]