NORA: Noise-Robust Tagging for Financial Numerical Entity Attributes

Researchers propose NORA (Noise-Robust Tagging for Rich Financial Numerical Entity Attributes) to improve financial numerical entity understanding. Existing methods focus on concept name prediction but suffer from noisy labels in inline XBRL filings and neglect attributes like reporting-time relation, measurement scale, and accounting sign. NORA uses task-aware instance-specific weighting to reduce noisy label influence during training and introduces Neighborhood Prior-adjusted KNN (NPK) filtering for reliable evaluation on real-world noisy test sets. A large-scale benchmark with 6.6 million instances and multi-attribute annotations is constructed. The paper is published on arXiv with ID 2605.24910.