Optimizing Knowledge Graph Embeddings for Enhanced Question Answering Performance
Akshar Patel · 2024
This study investigates the impact of various negative sampling techniques on the performance of Knowledge Graph Embeddings (KGEs) in Knowledge Graph Question Answering (KGQA) systems. I evaluate six KGE variations using three metrics: Mean Rank, Mean Reciprocal Rank (MRR), and Hits@k. Our analysis reveals that the ComplEx model with Random Corrupt Negative Sampling outperforms other configurations, offering superior accuracy by effectively capturing complex relational patterns in knowledge graphs. Furthermore, I explore the performance of twelve Embed-KGQA variations, demon-strating that the RoBERTa model excels as a question encoder, especially when paired with ComplEx embed dings and Random Corrupt Negative Sampling. Comparative results between the EmbedKGQA pipeline and cosine similarity methods highlight EmbedKGQA's significantly higher success rate in answering questions. This performance gain underscores the advantage of integrating advanced natural language processing models like RoBERTa and selecting optimal negative sampling strategies to enhance the accuracy and efficiency of KGQA systems.