ReranKGC: A cooperative retrieve-and-rerank framework for multi-modal knowledge graph completion.

Area of Science:

• Artificial Intelligence
• Data Science
• Graph Theory

Background:

• Multi-modal knowledge graph completion (MMKGC) predicts missing links using diverse entity attributes.
• Existing embedding-based methods excel at structural knowledge but underutilize multi-modal data.
• Fine-tune-based (FT-based) methods leverage multi-modal knowledge but struggle with entity ambiguity.

Purpose of the Study:

• To develop an ensemble framework, ReranKGC, that integrates the strengths of embedding-based and FT-based methods for MMKGC.
• To address the limitations of individual methods by creating a synergistic approach for improved link prediction.
• To enhance the accuracy and robustness of MMKGC by effectively utilizing both structural and multi-modal information.

Main Methods:

• ReranKGC employs a retrieve-and-rerank pipeline for knowledge graph completion.
• The retriever utilizes embedding-based methods for initial candidate retrieval.
• The re-ranker incorporates KGC-CLIP, an FT-based method using CLIP for multi-modal attribute analysis and candidate refinement.

Main Results:

• The retriever generates a candidate pool with semantically and structurally related entities.
• The re-ranker refines the ranking within this high-quality pool, improving precision.
• ReranKGC consistently enhances baseline performance and outperforms state-of-the-art models on link prediction tasks.

Conclusions:

• The ReranKGC framework effectively combines complementary methods to overcome individual limitations in MMKGC.
• This cooperative approach leads to superior performance in predicting missing links.
• ReranKGC demonstrates significant improvements, offering a more comprehensive solution for multi-modal knowledge graph completion.