KBDR: document retrieval based on graph matching with knowledge enhancement

Area of Science:

• Information Science
• Computer Science
• Artificial Intelligence

Background:

• Traditional document retrieval struggles with vocabulary synonymy, leading to inaccuracies.
• Existing matching algorithms often fail to capture semantic relevance beyond exact keyword matches.

Purpose of the Study:

• To introduce Knowledge Based Document Retrieval (KBDR) for enhanced semantic understanding.
• To address limitations in current document retrieval systems by incorporating external knowledge.

Main Methods:

• Constructing document graphs using named entity recognition and relation extraction.
• Expanding entities with external knowledge bases (HowNet, ConceptNet) for richer semantic connections.
• Applying a Graph Attention Network (GAT) to model inter-entity dependencies for relevance ranking.

Main Results:

• KBDR demonstrated superior performance compared to baseline methods across five datasets (CIPP, CNSE, CNSS, AAN, OC).
• Achieved the highest Recall@1 of 25.88% on the CNSE dataset, indicating significant retrieval improvement.
• The proposed method enhances efficiency and robustness in document retrieval tasks.

Conclusions:

• KBDR effectively overcomes conceptual deviations in document retrieval.
• The graph-based knowledge expansion approach advances the development of more sophisticated retrieval systems.