RSCF: Relation-Semantics Consistent Filter for Entity Embedding of Knowledge Graph

Anonymous

Abstract: In knowledge graph embedding, leveraging relation-specific entity-transformation has markedly enhanced performance. However, this approach lacks assurance for consistent changes in relation and entity embeddings due to the disconnected entity-transformation representation, missing valuable inductive bias among semantically similar relations. Furthermore, a generalized plug-in approach as a SFBR disrupts this consistency through excessive concentration of entity embeddings under entity-based regularization, generating indistinguishable score distributions among relations. To tackle these challenges, we introduce Relation-Semantics Consistent Filter (RSCF), characterized by three features: 1) shared affine transformation of relation embeddings across all relations, 2) change-based entity-transformation that adds an entity embedding to its change represented by the transformed vector, and 3) normalization of the change to prevent scale reduction. In knowledge graph completion tasks with distance-based and tensor decomposition models, RSCF notably enhances performance across all relations, particularly in rare relations in long-tailed distribution.

Paper Type: long

Research Area: Information Retrieval and Text Mining

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Approaches to low-resource settings

Languages Studied: English