Preserving overlapped information via parallel one-hop and multi-hop neighbor encoding for knowledge graph entity typing

Abstract

Knowledge graph entity typing (KGET) is critical for predicting and completing missing entity types in a knowledge graph (KG). Existing KGET models mainly aggregate local semantic and structural information from multi-hop and one-hop neighbors via weighted aggregation. However, the one-hop neighbor information within the multi-hop neighbor context is often diluted during aggregation, resulting in incomplete information collection and inaccurate type prediction. To preserve this overlapped one-hop neighbor information, we propose a novel framework, the one-hop and multi-hop neighbor parallel encoding framework (OMNPEF), which captures local-to-global semantic and structural information. Specifically, OMNPEF encodes one-hop and multi-hop neighbors in parallel to better preserve the overlapped one-hop neighbor information and integrates local information with global semantic and structural insights, enhancing the model’s capacity to learn from the graph structure. Experimental results on the FB15kET and YAGO43kET datasets demonstrate that OMNPEF outperforms state-of-the-art models, achieving a mean improvement of at least 1.3% in MRR.