Abstract: Finding an attribute to explain the relationships between a given pair of entities is valuable in many applications. However, many direct solutions fail, owing to its low precision caused by heavy dependence on text and low recall by evidence scarcity. Thus, we propose a generalization-and-inference framework and implement it to build a system: entity-relationship finder (ERF). Our main idea is conceptualizing entity pairs into proper concept pairs, as intermediate random variables to form the explanation. Although entity conceptualization has been studied, it has new challenges of collective optimization for multiple relationship instances, joint optimization for both entities, and aggregation of diluted observations into the head concepts defining the relationship. We propose conceptualization solutions and validate them as well as the framework with extensive experiments.

Key words: relation explanation, knowledge base, entity-relationship finder (ERF), probabilistic generative model

摘要： 在许多应用程序中，找到一个属性来解释给定实体对之间关系是很有价值的。然而，许多直接解决方案的失败是由于严重依赖文本而导致精度低，并且由于证据稀缺而导致召回率低。因此，提出了一个泛化和推理框架，并将其构建成一个系统：实体-关系查找（ERF）。主要思想是将实体对概念化为适当的概念对，作为中间随机变量来形成解释。尽管已经对实体概念化进行了研究，但其面临着多个关系实例的集体优化、两个实体的联合优化以及将稀释的观察结果聚集到定义关系的头部概念中的新挑战。我们提出了概念化的解决方案，并通过大量实验验证了这些方案及其框架。

关键词: 关系解释，知识库，实体-关系查找（ERF），概率生成模型