Validate LitEm Across Larger Graphs, Higher Dimensions, and Model-Specific Tuning

Determine whether the performance patterns reported for LitEm and its co-training framework hold at higher embedding dimensionalities, on larger knowledge graphs, or under model-specific hyperparameter tuning.

Background

The study evaluates LitEm and its co-training framework using embeddings of at most 64 dimensions, 32 dimensions for the literal-awareness evaluation, and knowledge graphs containing approximately 12,000–15,000 entities. Each experiment type uses a single fixed hyperparameter setting. Consequently, the reported comparative patterns may not generalize to larger-scale knowledge graphs, higher-dimensional embeddings, or settings in which hyperparameters are optimized separately for each model. The paper explicitly leaves this empirical question unresolved.

References

Whether the reported patterns hold at higher dimensionality, on larger graphs, or under model-specific tuning therefore remains untested.

— Neural Regression with Embeddings for Numerical Attribute Prediction in Knowledge Graphs  (2608.26729 - Sapkota et al., 27 Aug 2026) in Section Discussion, subsection “Limitation and Future Work”