LollipopE: Bi-centered lollipop embedding for complex logic query on knowledge graph.

Area of Science:

• Artificial Intelligence
• Data Science
• Computer Science

Background:

• Complex First-Order Logic (FOL) queries are crucial for multi-hop knowledge graph (KG) reasoning.
• Geometric methods offer a promising approach for KG query embedding (QE).
• Existing geometric QE models face challenges including underutilized embedding space, overreliance on angular information, and failure to capture hierarchy.

Purpose of the Study:

• To address limitations in current geometric query embedding models.
• To propose a novel method, LollipopE, for improved KG reasoning.
• To enhance the utilization of embedding space and capture hierarchical structures in KGs.

Main Methods:

• Introduced LollipopE, a lollipop-like bi-centered query embedding method.
• Employed learnable centroid positions to maximize embedding space utilization.
• Developed a combined angular and centroid-based metric for answer selection.
• Incorporated dynamic moduli to represent entity hierarchy within the KG.

Main Results:

• LollipopE significantly outperforms existing state-of-the-art geometric methods.
• The proposed method shows the most substantial improvements on datasets with complex hierarchical structures.
• LollipopE effectively addresses the underutilization of embedding space and overreliance on angular metrics.

Conclusions:

• LollipopE offers a superior approach to geometric query embedding for multi-hop KG reasoning.
• The method's ability to leverage embedding space and capture hierarchy leads to enhanced reasoning performance.
• LollipopE represents a significant advancement in addressing complex query answering in knowledge graphs.