Scholarlink

<p>The financial statement footnotes provide supplementary explanations for items not presented in the financial statements, and the information carried is equally crucial for decision-makers. However, financial statement footnotes suffer from complexities in presentation, diverse disclosure forms, and scattered argumentation, resulting in limitations in current event extraction methods and publicly available datasets for the financial statement footnote event extraction task. To address these issues, this paper summarizes the event structure in financial statement footnotes, constructs a dataset called FE-FSF for extracting financial events from financial statement footnotes, which provides a data foundation for automatic extraction of financial events from financial statement footnotes. Additionally, this paper proposes a Joint Event Extraction Framework based on Chapter Structure (JEEFCS) for financial statement footnotes of Chinese Listed Companies. By learning the chapter structure of the financial statement footnotes and the event meta-information that incorporates lexical-level, syntactic-level, and semantic-level information of the paragraphs (i.e., each sample), JEEFCS identifies the event categories of the financial events in the financial statement footnotes that have a significant impact on the items listed in the financial statements, and then completes the task of filling the argument roles with the help of the path construction of the directed acyclic graph. Extensive experiments and comprehensive analysis demonstrate that JEEFCS significantly outperforms existing state-of-the-art methods, including DCFEE-O, DCFEE-M, GreedyDec, and Doc2EDAG, in terms of F1 scores across five event categories (EF, ER, EP, IC, MC). Specifically, JEEFCS shows an average improvement of approximately 52.92% in recall (R), 19.18% in precision (P), and 38.41% in micro-F1 score. These improvements indicate that JEEFCS outperforms other models in terms of overall performance, balancing the recognition of positive and negative samples, and event extraction tasks in the context of financial statement footnote event extraction.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

JEEFCS: A Joint Event Extraction Framework based on Chapter Structure for Financial Statement Footnotes of Chinese Listed Companies

  • 期刊论文

  • 发表日期：

    2025年9月3日

查看全文
  • Ding Pan,
  • Yan Li,
  • Xing Zhou

摘要

The financial statement footnotes provide supplementary explanations for items not presented in the financial statements, and the information carried is equally crucial for decision-makers. However, financial statement footnotes suffer from complexities in presentation, diverse disclosure forms, and scattered argumentation, resulting in limitations in current event extraction methods and publicly available datasets for the financial statement footnote event extraction task. To address these issues, this paper summarizes the event structure in financial statement footnotes, constructs a dataset called FE-FSF for extracting financial events from financial statement footnotes, which provides a data foundation for automatic extraction of financial events from financial statement footnotes. Additionally, this paper proposes a Joint Event Extraction Framework based on Chapter Structure (JEEFCS) for financial statement footnotes of Chinese Listed Companies. By learning the chapter structure of the financial statement footnotes and the event meta-information that incorporates lexical-level, syntactic-level, and semantic-level information of the paragraphs (i.e., each sample), JEEFCS identifies the event categories of the financial events in the financial statement footnotes that have a significant impact on the items listed in the financial statements, and then completes the task of filling the argument roles with the help of the path construction of the directed acyclic graph. Extensive experiments and comprehensive analysis demonstrate that JEEFCS significantly outperforms existing state-of-the-art methods, including DCFEE-O, DCFEE-M, GreedyDec, and Doc2EDAG, in terms of F1 scores across five event categories (EF, ER, EP, IC, MC). Specifically, JEEFCS shows an average improvement of approximately 52.92% in recall (R), 19.18% in precision (P), and 38.41% in micro-F1 score. These improvements indicate that JEEFCS outperforms other models in terms of overall performance, balancing the recognition of positive and negative samples, and event extraction tasks in the context of financial statement footnote event extraction.