Edge-Aware Graph Neural Network for Multi-Hop Path Reasoning over Knowledge Base.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Knowledge Representation

Background:

• Multi-hop path reasoning is key for knowledge base question answering (KBQA).
• Existing deep learning models overlook latent relation information and question-relation correlations.
• This limits the effectiveness of simulating human-like multi-hop question-solving.

Purpose of the Study:

• To propose an edge-aware graph neural network for enhanced multi-hop path reasoning in KBQA.
• To improve the integration of question and relation information within graph neural networks.
• To address limitations in current models regarding latent relation information and question-relation correlation.

Main Methods:

• Constructing a 'query graph' by adding a query node to the knowledge base subgraph.
• Integrating question-related relation information into entity node representations during graph updates.
• Employing an attention mechanism to prioritize relevant neighbor node information.

Main Results:

• The proposed edge-aware graph neural network significantly outperformed baseline methods on MetaQA and PathQuestion-Large (PQL) benchmarks.
• Achieved higher Hit@1 and F1 scores, demonstrating superior performance.
• Ablation studies confirmed the positive impact of both the graph construction and update strategies.

Conclusions:

• The edge-aware graph neural network effectively enhances multi-hop path reasoning in KBQA.
• The novel graph construction and update methods are crucial for performance gains.
• This approach offers a promising direction for advancing KBQA research.