Noise-Robust Financial Numerical Entity Attribute Tagging
The paper introduces a new method called NORA for improving the understanding of financial numerical entities in reports. It addresses limitations in existing studies related to noisy labels and underrepresented attributes. The results demonstrate that NORA outperforms current methods in accuracy and robustness against label noise.
- ▪NORA uses instance-specific weighting to mitigate the impact of noisy labels during training.
- ▪The method includes a Neighborhood Prior-adjusted KNN filtering for reliable evaluation on noisy test sets.
- ▪NORA achieved the best performance in concept name and time-relation prediction compared to state-of-the-art methods.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.24910 (cs) [Submitted on 24 May 2026] Title:Noise-Robust Financial Numerical Entity Attribute Tagging Authors:Hsin-Min Lu, Chen-Yang Lai, Yi-Jhen Li, Ju-Chun Yen View a PDF of the paper titled Noise-Robust Financial Numerical Entity Attribute Tagging, by Hsin-Min Lu and 3 other authors View PDF HTML (experimental) Abstract:Financial Numerical Entity (FNE) understanding aims to recover the meaning of numerical mentions in financial reports. Existing studies primarily focus on concept name prediction and face two important limitations. First, labels derived from inline XBRL may contain errors because filings are usually prepared manually.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
