Researchers have introduced LongDocBench, a new benchmark designed to evaluate the recovery of structural information in long documents. This benchmark focuses on two key tasks: reconstructing Table-of-Contents (TOC) hierarchies and identifying contextual relationships between elements like tables and figures and their associated text. LongDocBench comprises 85 real-world documents, including financial reports and textbooks, with annotations for over 3,900 heading nodes and 3,200 contextual relationships. Experiments show that improved TOC hierarchies and contextual relationships enhance reasoning in long-document question-answering tasks, while current document parsers struggle with these specific recovery tasks despite strong page-level performance. AI
IMPACT This benchmark could drive improvements in AI's ability to understand and process complex, long-form documents, enhancing applications like research analysis and information retrieval.
RANK_REASON New academic paper introducing a benchmark. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
