基于实体不可知表示学习的知识图谱嵌入

Knowledge Graph Embedding Based on Entity-Agnostic Representation Learning

摘要: 针对知识图谱嵌入(KGE)引发的参数存储效率低下的问题，本研究提出了一种创新的实体不可知表示学习方法。传统的知识图谱嵌入技术通过为知识图谱中的各个元素(涵盖实体和关系)分配特定的嵌入(即向量化表达)，将其映射至连续的向量空间。然而，这种方法导致了嵌入参数随着知识图谱规模的扩大而呈现线性增长的趋势。在此基础上，我们提出了名为实体不可知表示学习(EARL)的模型。该模型仅学习一小部分实体的嵌入，这些实体被称为保留实体。为了获取完整实体集的嵌入，我们巧妙地结合关系、K近邻保留实体以及多跳邻居中的信息，以编码这些保留实体的独特特征。通过学习通用且实体不可知的编码器，我们将这些特征高效地转换为实体的嵌入。相较于传统的知识图谱嵌入技术，这种创新方法使得我们提出的EARL模型在保持高效的同时，具有更少的参数量，从而展现出更高的静态性。实验结果充分验证了EARL在链路预测任务上的卓越性能，并且在参数效率方面展现出显著优势，进一步凸显了其在减少参数使用方面的有效性。

Abstract: In order to solve the problem of inefficient parameter storage caused by Knowledge Graph Embedding (KGE), this study proposes an innovative entity-agnostic representation learning method. Traditional knowledge graph embedding techniques map each element (covering entities and relations) in the knowledge graph to a continuous vector space by assigning specific embeddings (i.e. vectorized representations). However, this approach leads to a linear growth trend of embedding parameters as the size of the knowledge graph increases. On this basis, we propose a model called Entity-Agnostic Representation Learning (EARL). The model learns the embedding of only a small subset of entities, which are known as reserved entities. To obtain embeddings for the full set of entities, we skillfully combine information from relationships, K-nearest neighbor preserved entities, and multi-hop neighbors to encode the unique characteristics of these retained entities. By learning a generic and entity-agnostic encoder, we efficiently translate these features into entity embeddings. Compared with the traditional knowledge graph embedding technology, this innovative method enables the proposed EARL model to have fewer parameters while maintaining high efficiency, so as to show higher staticity. The experimental results fully verify the excellent performance of EARL in the link prediction task, and show significant advantages in parameter efficiency, which further highlights its effectiveness in reducing the use of parameters.