Reproduce “Benchmarking Table Extraction from Heterogeneous Scientific PDF Documents” 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
Marijan Soric, Cécile Gracianne, Ioana Manolescu, Pierre Senellart
Soric et al. introduced a comprehensive end-to-end benchmark for evaluating table extraction from heterogeneous scientific PDF documents, utilizing novel metrics and two newly developed datasets. The study evaluates fifteen diverse methods, revealing limitations in model generalization across varied data and highlighting the trade-offs between extraction performance and computational cost.
Key insights
- End-to-end table extraction from heterogeneous scientific PDF documents remains a challenging problem, with no single method achieving perfect performance across all aspects.
- Model performance is highly sensitive to data heterogeneity; specialized or heuristic methods can excel on specific table types (e.g., bordered tables) but struggle with others (e.g., LaTeX-typeset tables).
- General Vision Language Models often suffer from significant precision drops due to false positives on pages without tables and exhibit content hallucination, generally performing worse than specialized TE models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.