面向事件预测的因果知识发现、验证和推理研究

As a special kind of commonsense knowledge, causality represents the causal relation between events which plays an important role in the construction of commonsense knowledge base. Due to its good ability in discribing the evolution of an event, the application of causal knowledge would produce an effect on our national security. However, there is a lack of the more systematic research on acquisition of large-scale causality, and research on how to represent and make use of acquired causal knowledge. To improve the scale and accuracy of causality acquistion, we proposed a systematic approach to mine causal relations between events, and investigated the evaluation and application of causal knowledge. This project mainly focuses on the following research issues:(1) extraction of causal relations from the web corpora;(2)extraction of event structures and analysis of the semantic constraints on event arguments;(3) evaluation of acquired causal knowledge; (4) uncertainty reasoning model and computational method based on evidential network. Through the above reasearch efforts, this project will construct a large-scale causal knowledge base which supports the application of event prediction. Furthermore, it will provide a general research framework for other types of relation acquistion between events.

因果知识作为一种描述事件关系的常识知识，是大规模知识库构建中不可或缺的一部分。由于因果知识能够很好地描述事件的演化过程，其在安全领域的应用关系到国家稳定与社会安全。面向事件预测的应用需求，研究大规模因果知识的获取，同时构建基于因果知识的推理模型是当前亟待解决的重要问题和研究热点。针对目前因果知识获取的规模小、精度低、应用难等问题，本课题将系统地研究因果知识的发现、验证和推理方法，主要内容包括：(1)面向Web语料的因果关系识别和提取方法；（2）因果事件的角色分析和语义约束学习方法；（3）因果知识的可信度评估方法；（4）基于证据网络的因果知识推理方法。通过本课题的研究，旨在构建支持事件预测的大规模因果知识库，同时为通用的事件关系获取提供统一的研究框架。

本项目在执行过程中，按照项目申请书和计划书中的研究计划，开展相应的研究内容，未出现研究目标和研究内容的调整。同时，根据本领域的发展趋势，增加事件推荐技术研究。通过本项目的研究，构建了大规模的事件因果知识库，将面向Web语料的事件关系抽取精度提升5.6%、覆盖率提高17%；将隐式关系发现速度提高2~3倍；提出新的事件表示方法，提高了事件知识的可用性；面向社交关系的事件推荐Map值提高8%；在国际重要学术会议和期刊上发表高质量论文12篇。目前所有研究目标均已完成，研究成果在多个国家重要部门获得应用。