Uncertain Knowledge Graph Embedding Using Auxiliary Information

Uncertain knowledge graphs (UKGs) offer a more realistic representation of knowledge by capturing the uncertainty associated with facts. However, existing UKG embedding methods primarily rely on structural information for confidence score prediction, neglecting other sources of uncertainty. This paper investigates the effectiveness of incorporating auxiliary information into UKG embeddings. We propose two approaches: Text-BEUrRE, which leverages textual information, and UCompGCN, which utilizes topological information. Our extensive experiments demonstrate that both methods successfully integrate these auxiliary data sources. Notably, Text-BEUrRE and UCompGCN outperform state-of-the-art baselines on most metrics in the confidence prediction task. On the CN15K dataset, Text-BEUrRE achieves a 7.39% improvement in Mean Squared Error (MSE) compared to the best existing model, while UCompGCN achieves an 8.27% improvement in Mean Absolute Error (MAE).