Reproduce “Open Domain Question Answering over Tables via Dense Retrieval” 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
Jonathan Herzig, Thomas Müller, Syrine Krichene, Julian Martin Eisenschlos
A system for open-domain question answering over tables leverages a dense retriever adapted for structured data and a table reader, achieving improved end-to-end performance compared to text-based retrieval methods. The work also introduces NQ-TABLES, a new dataset specifically for this task.
Key insights
- Dense retrieval significantly outperforms traditional methods (BM25) for table retrieval, especially when leveraging tabular-specific encoders like TAPAS.
- Large-scale pre-training on text-table pairs and incorporating mined hard negatives are crucial for developing effective and robust dense table retrievers.
- The table reader component still faces challenges in extracting answers from tables requiring complex reasoning or common-sense knowledge.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.