LongDocBench: New Benchmark for TOC Hierarchy and Contextual Relationship Recovery in Long Documents
A new benchmark called LongDocBench has been launched by researchers to assess the recovery of document-level structures in lengthy texts. It emphasizes two primary tasks: recovering the Table-of-Contents (TOC) hierarchy and understanding contextual relationships. These tasks require the reconstruction of TOC hierarchies that span multiple pages and the identification of connections from tables and figures to their respective captions, notes, and sources, often in a one-to-many format. While existing benchmarks mainly address page-level element recognition and related tasks, they do not fully encompass these document-level challenges. LongDocBench includes 85 authentic financial reports, textbooks, and academic papers, totaling 2,582 pages, with individual documents reaching up to 105 pages. This initiative aims to fill the evaluation gap for these important tasks. The details were shared on arXiv under the identifier 2608.15064.
Key facts
- LongDocBench is a new benchmark for document-level structure recovery.
- It evaluates two tasks: TOC Hierarchy Recovery and Contextual Relationship Recovery.
- The benchmark includes 85 real-world documents: financial reports, textbooks, and academic papers.
- The documents span 2,582 pages, with up to 105 pages per document.
- Existing benchmarks focus on page-level tasks, not document-level.
- The tasks involve reconstructing cross-page TOC hierarchies and identifying links from tables/figures to captions, notes, and sources.
- The announcement was made on arXiv with identifier 2608.15064.
- The benchmark addresses a gap in evaluating document-level structure recovery.
Entities
—