Documenting Knowledge Graph Embedding and Link Prediction using Knowledge Graphs

Abstract

In recent years, sub-symbolic learning, i.e., Knowledge Graph Embedding (KGE) incorporated with Knowledge Graphs (KGs) has gained significant attention in various downstream tasks (e.g., Link Prediction (LP)). These techniques learn a latent vector representation of KG's semantical structure to infer missing links. Nonetheless, the KGE models remain a black box, and the decision-making process behind them is not clear. Thus, the trustability and reliability of the model's outcomes have been challenged. While many state-of-the-art approaches provide data-driven frameworks to address these issues, they do not always provide a complete understanding, and the interpretations are not machine-readable. That is why, in this work, we extend a hybrid interpretable framework, InterpretME, in the field of the KGE models, especially for translation distance models, which include TransE, TransH, TransR, and TransD. The experimental evaluation on various benchmark KGs supports the validity of this approach, which we term Trace KGE. Trace KGE, in particular, contributes to increased interpretability and understanding of the perplexing KGE model's behavior.