MLSFF: Multi-level structural features fusion for multi-modal knowledge graph completion.

Area of Science:

• Artificial Intelligence
• Data Science

Background:

• Multi-modal knowledge graphs are crucial for storing human knowledge.
• Knowledge graph incompleteness is a significant challenge, necessitating knowledge graph completion (KGC).
• Existing KGC methods lack interpretability and struggle with multi-hop reasoning.

Purpose of the Study:

• To propose a novel method for multi-modal knowledge graph completion.
• To enhance reasoning accuracy and interpretability in multi-modal knowledge graphs.
• To address the limitations of current embedding-based and graph neural network approaches.

Main Methods:

• Utilizes Transformer architecture for separate image and text modality representation learning.
• Employs multimodal gating units for information filtering and feature fusion.
• Extracts multi-level path features and applies graph convolutional operations for structural learning.

Main Results:

• Achieved 1.8% improvement in Hits@1 on FB15K-237-IMG dataset.
• Achieved 0.7% improvement in Hits@1 on WN18-IMG dataset.
• Demonstrated effectiveness in multi-modal knowledge graph completion tasks.

Conclusions:

• The proposed method effectively learns multi-level graph structural features.
• The approach improves reasoning accuracy for multi-modal knowledge graph completion.
• This work offers a promising direction for enhancing knowledge graph completeness and utility.