计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (19): 189-197.DOI: 10.3778/j.issn.1002-8331.2104-0430

• 模式识别与人工智能 • 上一篇    下一篇

融合知识表示和深度强化学习的知识推理方法

宋浩楠,赵刚,王兴芬   

  1. 北京信息科技大学 信息管理学院,北京 100192
  • 出版日期:2021-10-01 发布日期:2021-09-29

Knowledge Reasoning Method Combining Knowledge Representation with Deep Reinforcement Learning

SONG Haonan, ZHAO Gang, WANG Xingfen   

  1. School of Information Management, Beijing Information Science & Technology University, Beijing 100192, China
  • Online:2021-10-01 Published:2021-09-29

摘要:

知识推理是解决知识图谱中知识缺失问题的重要方法,针对大规模知识图谱中知识推理方法仍存在可解释性差、推理准确率和效率偏低的问题,提出了一种将知识表示和深度强化学习相结合的方法RLPTransE。利用知识表示学习方法,将知识图谱映射到含有三元组语义信息的向量空间中,并在该空间中建立强化学习环境。通过单步择优策略网络和多步推理策略网络的训练,使强化学习智能体在与环境交互过程中,高效挖掘推理规则进而完成推理。在公开数据集上的实验结果表明,相比于其他先进方法,该方法在大规模数据集推理任务中取得更好的表现。

关键词: 知识推理, 深度强化学习, 知识表示, 路径控制, 规则挖掘

Abstract:

Knowledge reasoning is an important method to solve the problem of lack of knowledge in the knowledge graph. The knowledge reasoning method in the large-scale knowledge graph still has the problems of poor interpretability, low reasoning accuracy and efficiency. This paper proposes a method RLPTransE that combines knowledge representation with deep reinforcement learning. Firstly, it uses the knowledge representation learning method to map the knowledge graph to the vector space containing the semantic information of the triples, and establishes a reinforcement learning environment in the space. Then, it trains through the single-step optimization strategy network and the multi-step reasoning strategy network to enable the reinforcement learning agent to efficiently mine the reasoning rules and complete the reasoning in the process interacting with the environment. According to experimental results on public datasets, compared with state-of-the-art methods, the proposed method achieves better performance in reasoning tasks for large-scale datasets.

Key words: knowledge reasoning, deep reinforcement learning, knowledge representation, path control, rule mining