Reproduce “Concept2Box: Joint Geometric Embeddings for Learning Two-View Knowledge Graphs” 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
Zijie Huang, Daheng Wang, Binxuan Huang, Chenwei Zhang, Jingbo Shang, Yan Liang, Zhengyang Wang, Xian Li and 3 more
Concept2Box introduces a joint geometric embedding model for knowledge graphs comprising distinct ontology and instance views, representing concepts as probabilistic Gumbel boxes and entities as vectors. The model leverages a novel vector-to-box distance metric to bridge these views, demonstrating improved knowledge graph completion and concept linking performance on both public DBpedia and a newly created industrial recipe dataset.
Key insights
- Dual geometric representations, where distinct node types (concepts and entities) are modeled by appropriate geometric objects (boxes and vectors, respectively), are more effective for heterogeneous two-view KGs than uniform embedding approaches.
- Probabilistic box embeddings provide an interpretable semantic meaning for concept granularity, where the volume of a box directly correlates with the broadness of the concept it represents.
- Jointly encoding concept and relation text with attention mechanisms, rather than simple concatenation or separate parameterization, enhances the semantic richness of transformed concept boxes and improves model performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.