面向智能推理的逻辑增强型分布式知识表示研究

Knowledge graph construction and application has gained massive attention in the last few years. Intelligent reasoning, which aims to infer new facts automatically from existing ones in knowledge graphs, then stands out as a key research problem. Reasoning is an important advanced thinking mode of human beings. It is also a prerequisite for machines to get real artificial intelligence. There are two major groups of reasoning techniques: embedding-based statistical reasoning and rule-based symbolic reasoning. These two groups of techniques describe the reasoning task from different perspectives, which are suitable for different application scenarios, and hence highly complementary to each other. Combining them together will definitely bring infinite opportunity for reasoning. Despite the huge potential and impact, research in this direction is still in its infancy. There is much room for improvement and a very broad prospect for future research. This project studies statistical-symbolic reasoning, which can integrate embedding-based statistical reasoning and rule-based symbolic reasoning flexibly, effectively, and seamlessly into a unified framework. Three key theoretical and technical challenges are needed to be solved, including the transfer from propositional logic to first-order logic, the transfer from hard rules to soft rules, and the transfer from one-time transmission to multi-round interactions. The research topic studied in this project is key to knowledge graphs, and even artificial intelligence. It has great theoretical value for the development of artificial intelligence, and also great practical significance for the application of knowledge graphs.

知识图谱的构建和应用是近来学术界和工业界关注的焦点。其中一个很重要的研究议题就是知识的自动推理，即按照某种策略从已有知识推断出新知识。推理是人类一种重要的高级思维形式，也是机器真正意义上拥有智能的必备前提。基于统计的分布式知识表示与基于符号的经典逻辑都是知识推理的重要技术。它们从不同侧面刻画推理任务，具有各自的应用场景，并且具有很强的互补性。如果能将两者有效结合，必将为推理带来无限机遇。然而该方向的研究仍处于起步阶段，具有巨大的提升空间和研究前景。本项目面向大规模知识图谱上的精准智能推理，围绕统计—符号推理这一核心议题，系统研究分布式知识表示和经典逻辑的耦合，重点突破由命题逻辑到一阶逻辑的跨越、由硬规则到软规则的拓展、由单次传导到多轮交互的迁移这三大技术难题。本课题面向的是知识图谱乃至人工智能领域的核心议题，对人工智能学科发展具有重大的理论价值，对知识图谱的应用普及也具有重要的现实意义。

知识图谱的构建和应用是近来学术界和工业界关注的焦点。其中一个很重要的研究议题就是知识的自动推理，即按照某种策略从已有知识推断出新知识。推理是人类一种重要的高级思维形式，也是机器真正意义上拥有智能的必备前提。基于统计的分布式知识表示与基于符号的经典逻辑都是知识推理的重要技术。它们从不同侧面刻画推理任务，具有各自的应用场景，并且具有很强的互补性。如果能将两者有效结合，必将为推理带来无限机遇。..本项目面向智能推理的逻辑增强型分布式知识表示展开深入研究，主要研究内容包括：（1）基于神经序列模型和次序嵌入的一阶逻辑全称量词建模；（2）基于变尺度松弛的带置信度规则建模；（3）基于类主动学习的逻辑推理和表示学习内在交互性建模。基于上述模块，实现了（4）面向超大规模知识图谱的精准智能推理系统。最后，利用该智能推理系统，研发了（5）语言表示与知识表示的深度融合模型KT-NET，借助语言和知识的力量进一步提升自然语言理解的效果。..项目总共发表高水平学术论文10篇，包括CCF A类国际会议长文5篇，CCF B类国际会议长文1篇，CCF B类期刊论文1篇，其中还包括CCF A类会议ACM MM最佳学生论文1篇。迄今论文他引达400余次（Google Scholar统计）。获BioNLP 2021 国际研讨会 Radiology Report Summarization 和 MRQA 2019 国际研讨会 Evaluating Generalization in Reading Comprehension 等技术评测全球冠军。参加了 2020年全国知识图谱与语义计算大会“面向中文电子病历的医疗事件抽取”技术评测，提出的基于预训练语言模型的小样本医疗事件抽取系统获第一名和技术创新奖。..本课题面向的是知识图谱乃至人工智能领域的核心议题，对人工智能学科发展具有重大的理论价值，对知识图谱的应用普及也具有重要的现实意义。

内容获取失败，请点击重试