Conceitos essenciais
The proposed Semantic and Structure-aware KG Entity Typing (SSET) framework effectively integrates semantic and structural knowledge to improve the performance of knowledge graph entity typing.
Resumo
The content discusses the Knowledge Graph Entity Typing (KGET) task, which aims to predict missing type annotations for entities in knowledge graphs. Recent works have focused on utilizing only the structural knowledge in the local neighborhood of entities, disregarding the semantic knowledge in the textual representations of entities, relations, and types that is also crucial for type inference.
The paper proposes a novel SSET framework that consists of three modules:
Semantic Knowledge Encoding (SEM) module: This module encodes factual knowledge in the knowledge graph using a Masked Entity Typing task, which leverages the textual information of entities, relations, and types.
Structural Knowledge Aggregation (SKA) module: This module aggregates knowledge from the multi-hop neighborhood of entities, including 1-hop neighbors, multi-hop neighbors, and known types, to infer missing types.
Unsupervised Type Re-ranking (UTR) module: This module utilizes the inference results from the SEM and SKA modules to generate type predictions that are robust to false-negative samples, by leveraging the agreement between semantic and structural knowledge.
The authors observe that the interaction between semantic and structural knowledge can be utilized to address the false-negative problem, where some plausible type annotations are missing from the knowledge graph. The proposed SSET framework significantly outperforms existing state-of-the-art methods on two widely used datasets, FB15kET and YAGO43kET.
Estatísticas
The statistics on the FB15kET dataset show that 10% of the entities with the type "/music/artist" do not have type "/people/person".
Citações
"We argue that the textual representations of entities, relations, and types provide important semantic knowledge for type inference."
"We observe that semantic and structural knowledge can complement each other to alleviate the false-negative problem."