Brief

Researchers have developed a new method called NORA to improve the accuracy of understanding numerical data in financial reports. This approach addresses limitations in existing methods, such as noisy labels from manual filings and the underemphasis on crucial attributes like reporting time, measurement scale, and accounting sign. NORA employs task-aware instance-specific weighting to mitigate the impact of erroneous labels during training and introduces a Neighborhood Prior-adjusted KNN filtering technique for more reliable evaluation on real-world noisy datasets. AI