JointRel: Joint semantic embedding with relational message passing for knowledge graph completion

Area of Science:

• Artificial Intelligence
• Data Science

Background:

• Knowledge graph completion (KGC) traditionally focuses on entity semantics, often overlooking crucial edge features.
• This imbalance limits the semantic expressiveness and practical applications of knowledge graphs.

Purpose of the Study:

• To develop a method that explicitly learns both node and edge features for a more balanced semantic representation in KGC.
• To improve the accuracy, stability, and robustness of knowledge graph completion.

Main Methods:

• Proposed JointRel, a dual-channel graph augmentation network for joint semantic embedding with relational message passing.
• Employed node-level and edge-level graph learning to capture neighbor information for entity embeddings.
• Utilized graph attention and relational context to enhance edge feature representations.

Main Results:

• JointRel demonstrated superiority on four KGC datasets, achieving significant Mean Reciprocal Rank (MRR) improvements over state-of-the-art methods.
• Observed MRR gains of 4.2%, 0.2%, 3.4%, and 24.8% across different datasets.
• The method provides a more balanced semantic framework for entities and relations.

Conclusions:

• JointRel offers a robust framework for KGC, enhancing semantic representation for both entities and relations.
• The improved stability and robustness benefit downstream applications such as question answering and recommendation systems.