Reproduce “How Can Graph Neural Networks Help Document Retrieval: A Case Study on CORD19 with Concept Map Generation” 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
Hejie Cui, Jiaying Lu, Yao Ge, Carl Yang
Researchers at Emory University investigated how Graph Neural Networks (GNNs) can enhance document retrieval by representing documents as concept maps, demonstrating that novel semantics-oriented graph functions (E-Pool and RW-Pool) improved NDCG@20 by 11-12% over traditional BM25 on the CORD-19 dataset, outperforming more complex GNN architectures.
Key insights
- Semantically rich concept maps, generated via constituency parsing, effectively reflect document content and provide discriminative signals for document relevance.
- For textual document retrieval, simpler, semantics-oriented graph functions focusing on semantic units and their interactions are more effective and stable than complex, structure-oriented GNNs.
- Considering simple higher-order interactions, such as those captured by edges (E-Pool) or short random walks (RW-Pool), significantly improves retrieval performance compared to processing isolated nodes (N-Pool).
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.