Reproduce “SemStruct: Contextualizing Semantic Embeddings with Structural Information for Schema Matching” 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
Inwon Kang, Kavitha Srinivas, Nandana Mihindukulasooriya, Sola Shirai, Parikshit Ram, Horst Samulowitz, Oshani Seneviratne
SemStruct is a schema matching framework that integrates semantic embeddings from frozen Pre-trained Language Models (PLMs) with structural information captured by Graph Neural Networks (GNNs). This approach explicitly models row-level value co-occurrences, achieving statistically significant improvements in accuracy on Valentine and SOTAB-SM benchmarks without fine-tuning large language models.
Key insights
- Explicitly modeling row-level value co-occurrences through a heterogeneous graph and GNN message passing is more effective for disambiguating column semantics than simple averaging or direct PLM-based methods.
- Row nodes in the heterogeneous graph function most effectively as 'structural conduits' for information propagation (best with zero-initialization), rather than requiring complex semantic content themselves.
- Combining strong semantic priors from frozen PLMs with a lightweight GNN for structural context enables high accuracy and computational efficiency, avoiding the burden of fine-tuning large language models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.