Reproduce “Fonduer: Knowledge Base Construction from Richly Formatted Data” 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
Sen Wu, Luke Hsiao, Xiao Cheng, Braden Hancock, Theodoros Rekatsinas, Philip Levis, Christopher Ré
Fonduer is a knowledge base construction system designed to extract relations from richly formatted documents by leveraging a multimodal deep learning architecture and a weak supervision programming model. The system demonstrated an average 41 F1 point improvement over existing methods and extracted up to 1.87 times more correct entries compared to expert-curated knowledge bases.
Key insights
- Document-level context is critical for achieving high-quality relation extraction from richly formatted data, providing up to a 71 F1 point improvement over sentence-level contexts in certain domains.
- The integration of multimodal features (textual, structural, tabular, visual) is essential, as each modality contributes uniquely and complementarily to the overall extraction quality.
- A weak supervision approach with multimodal Labeling Functions enables non-domain experts to construct high-quality knowledge bases more efficiently, leading to 1.9 times higher F1 scores compared to manual annotation in a user study.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.