Reproduce “Convolutional 2D Knowledge Graph Embeddings” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tim Dettmers, Pasquale Minervini, Pontus Stenetorp, Sebastian Riedel
Researchers from Università della Svizzera italiana and University College London developed ConvE, a multi-layer convolutional network for knowledge graph link prediction that achieved state-of-the-art performance with up to 17 times fewer parameters than previous methods. The study also quantified severe test set leakage in popular benchmarks, leading to the creation of robust datasets for more accurate evaluations.
Key insights
- 2D convolutional layers can effectively capture complex interactions between entity and relation embeddings, offering expressive power with high parameter efficiency.
- A significant portion of reported state-of-the-art results on original WN18 and FB15k datasets were due to simple rule-based exploitation of inverse relations, not advanced relational understanding.
- Deeper models like ConvE provide increasing performance advantages over shallower models as the complexity of the knowledge graph (measured by indegree and PageRank) increases.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.