Reproduce “Tailoring Table Retrieval from a Field-aware Hybrid Matching Perspective” 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
Da Li, Keping Bi, Jiafeng Guo, Xueqi Cheng
Researchers at the Chinese Academy of Sciences developed THYME, a table-tailored hybrid matching retriever that achieves state-of-the-art performance on table retrieval benchmarks by adaptively combining dense semantic and sparse lexical matching, tailored to specific table fields like titles or cells. This approach improved NDCG@5 by over 2.4 points on NQ-TABLES and 1.3 points on OTT-QA compared to prior best hybrid methods.
Key insights
- Table titles and captions benefit more from dense (semantic) matching, while table cells and headers require fine-grained sparse (lexical) matching for precise relevance.
- Applying distinct pooling strategies tailored to different table fields (e.g., titles, headers, cells) and dynamically aggregating them via a Mixture of Field Experts (MoFE) mechanism are crucial for effective table retrieval.
- A training strategy that employs score dropout, forcing both semantic and lexical paths to learn robustly, is effective in optimizing a hybrid retrieval model for tables.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.