Although knowledge graphs store a large number of facts in the form of triplets, they are still limited by incompleteness. Hence, Knowledge Graph Completion (KGC), defined as inferring missing entities or relations based on observed facts, has long been a fundamental issue for various knowledge driven downstream applications. Prevailing KG embedding methods for KGC like TransE rely solely on mining structural information of existing facts, thus failing to handle generalization issue as they are inapplicable to unseen entities. Recently, a series of researches employ pre-trained encoders to learn textual representation for triples i.e., textual-encoding methods. While exhibiting great generalization for unseen entities, they are still inferior compared with above KG embedding based ones. In this paper, we devise a novel textual-encoding learning framework for KGC. To enrich textual prior knowledge for more informative prediction, it features three hierarchical maskings which can utilize far contexts of input text so that textual prior knowledge can be elicited. Besides, to solve predictive ambiguity caused by improper relational modeling, a relational-aware structure learning scheme is applied based on textual embeddings. Extensive experimental results on several popular datasets suggest the effectiveness of our approach even compared with recent state-of-the-arts in this task.

尽管知识图以三元组的形式存储大量事实，但它们仍然受到不完整性的限制。因此，知识图补全（KGC），定义为基于观察到的事实推断缺失的实体或关系，长期以来一直是各种知识驱动的下游应用程序的基本问题。像 TransE 这样的 KGC 流行的 KG 嵌入方法仅依赖于挖掘现有事实的结构信息，因此无法处理泛化问题，因为它们不适用于不可见的实体。最近，一系列研究采用预训练的编码器来学习三元组的文本表示，即文本编码方法。虽然对不可见实体表现出很好的泛化能力，但与上述基于 KG 嵌入的实体相比，它们仍然较差。在本文中，我们为 KGC 设计了一种新颖的文本编码学习框架。为了丰富文本先验知识以进行信息更丰富的预测，它具有三个分层掩蔽，可以利用输入文本的远上下文，以便可以引出文本先验知识。此外，为了解决由不正确的关系建模引起的预测歧义，应用了基于文本嵌入的关系感知结构学习方案。对几个流行数据集的广泛实验结果表明，即使与该任务中的最新技术相比，我们的方法也是有效的。