Unified Item Segmentation for 10-Q and 10-K Filings Using Item-Aware Document-Level Auxiliary Tasks
Journal
Proceedings of the 6th ACM International Conference on AI in Finance
Start Page
942
End Page
950
ISBN
9798400722202
Date Issued
2025-11-14
Author(s)
Abstract
Public company filings, including 10-Qs and 10-Ks, contain critical items such as Risk Factors and Management's Discussion and Analysis that are essential for financial analysis and regulatory compliance. However, heterogeneity in formatting and evolving regulatory mandates pose significant challenges for automatic item segmentation. To address these issues, we propose a unified segmentation pipeline that jointly segments items in 10-Q and 10-K filings, enabling effective knowledge transfer between them. Our approach incorporates two document-level auxiliary tasks, Topic-aware Document Structure Objective (TDSO) and Joint Contrastive Semantics Learning (JCSL), and dynamically balances their contributions during training. These objectives enhance both structural boundary detection and contextual understanding in long-form financial texts. We also introduce IS10QF, a new benchmark dataset for evaluating item segmentation on quarterly filings. Experimental results demonstrate that our method consistently improves segmentation accuracy across various training conditions, particularly in low-resource scenarios, thereby reducing reliance on expensive manual annotation. By enabling scalable, high-quality item segmentation, our framework facilitates downstream analysis and better adaptation to changing disclosure requirements.
Event(s)
ICAIF 2025 - 6th ACM International Conference on AI in Finance
Subjects
Auxiliary task
Contrastive learning
Financial NLP
Item segmentation
Multi-task learning
NLP datasets
Text segmentation
Publisher
ACM
Type
conference paper