毕业生知识库

随着互联网的发展和普及，网络上不同类型的事件得以快速和广泛传播，给人们的生活和社会经济带来不同程度的影响。将事件数据有效地组织，进而提升对事件的综合分析和综合推理能力，在多领域具有十分重要的作用。面向事件知识图谱的问答能帮助人们以自然语言的方式快速和准确地获取事件相关知识，进而提升对事件的了解和认识。面向事件知识图谱的关系推理能预测出事件之间缺失的关系，从而对事件知识进行补全。因此，面向事件知识图谱的问答和关系推理具有重要的研究意义和应用价值。

本论文首先构建了面向事件的知识图谱，该事件知识图谱包含145万个事件和2117种关系类型。针对现有知识图谱问答研究存在的不足，本论文在问句和候选关系链的表示学习过程中，充分考虑到不同关系链与问句的语义差异性，建立了对比语义相似性学习的事件问答方法；针对现有知识图谱推理研究存在的不足，本论文提出了一种事件表示学习与神经逻辑相结合的事件关系推理方法，从而补全事件知识图谱中缺失的关系。最后，建立了一个事件知识图谱分析系统。

本论文的主要工作和研究贡献归纳如下：

1. 以往的知识图谱问答研究一方面只关注实体，而忽略了事件；另一方面不能充分建模复杂问句与候选关系链的语义表示。因此，本论文重点研究以事件为中心的多跳问答，提出了一种对比语义相似性学习的问答方法。本方法识别出问句中的主题事件或主题实体，然后构建一个与问句相关的检索子图并获取所有的候选关系链。考虑到各个候选关系链和问句语义相似性差异，本方法采用对比学习框架构建了一个公共的语义向量空间，使得问句与正确关系链距离更近而与错误关系链距离更远，从而获取更精准的问句答案。在基准数据集上的实验结果验证了所提出方法的有效性。

2. 以往的知识图谱推理研究大多通过三元组学习结点和关系的分布式表示，从而计算三元组的评分，然后以此为基础进行推理。然而，这种方法在可解释性上存在不足，也没有考虑到事件的组成要素以及文本描述等重要信息。因此，本论文提出了一种事件表示学习与神经逻辑规则相结合的事件关系推理方法。本方法使用事件的名称、时间、地点、参与者以及文本描述学习事件的分布式表示，使用循环神经网络学习推理规则并通过学习到的规则推理事件之间的关系。本论文从公开的数据中构建了一个事件关系推理数据集，在该数据集上的实验结果验证了所提出方法的有效性。

3. 本论文构建了一个事件知识图谱分析系统，包括事件知识图谱、事件问答以及事件关系推理三个子系统。所构建的事件知识图谱结点包含145万个事件、1558万个实体，结点之间的关系包含三大类、共2117种关系。在此基础上，事件问答子系统使用三元组匹配的方法回答简单问句，使用对比语义相似性学习的方法回答复杂问句；事件关系推理子系统使用事件表示学习和神经逻辑规则相结合的方法推理事件之间的关系，从而实现面向事件知识图谱的分析系统构建。

With the development and popularization of the Internet, events can be rapidly and widely disseminated, causing varying degrees of impact on people's lives and the social economy. Effectively organizing event data can enhance the comprehensive analysis and reasoning ability of events, and play a very important role in multiple fields. Question answering with event knowledge graphs can help people quickly and accurately obtain event-related knowledge in natural language, thereby improving their understanding and awareness of events. Relation reasoning with event knowledge graphs can predict missing relations between events and complete event knowledge. Question answering and relation reasoning with event knowledge graphs have important research significance and application value. 

This thesis firstly constructs an event-oriented knowledge graph, which contains 1.45 million events and 2,117 types of relationships.To address the shortcomings of existing knowledge graph question answering researches, this thesis proposes a contrastive semantic similarity learning method for event question answering, which fully considers the semantic differences between different relation chains and questions during the representation learning of relation chains and questions. To address the shortcomings of existing knowledge graph reasoning researches, this thesis proposes an event relation reasoning method that combines event representation learning and neural logic, in order to complete the missing relations in the event knowledge graph. Finally, an event knowledge graph analysis system is constructed.

The major work and contributions of this thesis are summarized as follows:

1. Previous researches on knowledge graph question answering have focused on entities while ignoring events, and have been unable to adequately model the semantic representation of complex questions and candidate relation chains. Therefore, this thesis focuses on event-centric multi-hop question answering and proposes a contrastive semantic similarity learning method for question answering. This method identifies the topic event or entity in the questions, then constructs a retrieval subgraph related to the question, and obtains all candidate relation chains. Considering the differences in semantic similarity between each candidate relation chain and the question, this thesis uses a contrastive learning framework to construct a common semantic vector space, so that the question is closer to the correct relation chain and farther from the incorrect relation chains, thereby obtaining more accurate question answers. Experimental results on the benchmark dataset verify the effectiveness of the proposed method.

2. Previous researches on knowledge graph reasoning mostly learn the distributed representations of nodes and relationships through triplets, and then compute the scores of the triplets to perform reasoning. However, these methods have limitations in terms of interpretability and fail to take into account important information such as the constituents and textual descriptions of events. Therefore, this thesis proposes an event relation reasoning method which combines event representation learning and neural logic rules. The method learns the distributed representation of events using their names, time, locations, participants, and textual descriptions, then uses a recurrent neural network to learn reasoning rules and infer the relations between events through the learned rules. The thesis constructs an event relation reasoning dataset from publicly available data and the experimental results on the dataset validate the effectiveness of the proposed method.

3. This thesis has constructed an event knowledge graph analysis system, which includes three subsystems, namely, event knowledge graph, event question answering, and event relation reasoning. The constructed event knowledge graph contains 1.45 million events, 15.58 million entities, and a total of 2,117 types of relationships in three categories. Based on the event knowledge graph, the event question answering subsystem uses the method of triple matching to answer simple questions and the method of contrastive semantic similarity learning to answer complex questions. The event relation reasoning subsystem uses the combination of event representation learning and neural logical rules to infer the relations between events. Therefore, the analysis system for the event knowledge graph has been constructed.