金融數值實體理解：任務、資料集與方法

Abstract

隨著 inline XBRL (iXBRL) 在財務報告中被廣泛與強制使用，能夠提升效率、一致性及法規遵循的自動化標記需求日益增加。人工標記不僅耗時費力，且容易出錯。然而，現有資料集在屬性涵蓋範圍及文本資訊上仍有不足，限制了模型效能及其在真實應用場景中的適用性。為解決上述問題，本研究從美國證券交易委員會（SEC）申報文件中建構出一套大規模 iXBRL 標記資料集，包含 660 萬筆句子。每筆資料包含多個屬性，如標籤名稱（tag name）、財務數值（fact value）、時間（time）等，並附有財務報表資訊及句子上下文資訊。本研究亦將標記任務重新定義為多屬性預測問題，以更真實地反映實際財務標記的複雜性。我們提出三階段任務設計，提升資料集在不同模型與應用場景中的靈活性與可用性。本研究比較兩種方法：基於 BERT 架構的多任務學習（multi-task learning, MTL）模型，以及大型語言模型（LLMs）進行的 few-shot 提示學習（prompting）。MTL 模型在多個屬性上展現優異的預測能力，於標籤、時間、數量級與正負屬性的加權 F1 分數分別達到 0.82、0.90 及 0.99。相較之下，few-shot LLMs 表現則相對較差。這一結果凸顯了大型語言模型和few-shot在財務領域的限制。此外，本研究所建資料集亦能協助發現過往申報文件中的標記不一致或潛在錯誤，展現其作為訓練資源與稽核工具的價值。從管理實務觀點而言，本研究提出的資料集與自動標記框架可大幅降低人工工作量，提升財務揭露的可靠性，並促進後續應用，例如對股東會資料、財報摘要，或在 XBRL 強制採用前發布之歷史財報的結構化分析。

The growing adoption of inline XBRL (iXBRL) in financial reporting has increased the need for automation to improve efficiency, consistency, and compliance. Manual tagging is often labor-intensive and error-prone, especially in corporate settings where financial filings must meet strict regulatory standards. While automated iXBRL tagging has shown promise, existing datasets lack comprehensive attribute coverage and contextual information, limiting model performance and real-world applicability. To address these limitations, this study constructs a large-scale iXBRL tagging dataset containing 6.6 million sentences from SEC filings. Each instance is annotated with multiple attributes, such as tag name, fact value, and time, as well as document metadata and surrounding sentence context. We reformulate the tagging task as a multi-attribute prediction problem, which better reflects the complexity of real-world financial reporting. A three-stage task design is proposed to improve the flexibility and usability of the dataset for various modeling approaches and applications. To establish performance baselines, we evaluate two methods: a multi-task learning (MTL) model based on a BERT architecture and few-shot prompting using large language models (LLMs). The MTL model demonstrates strong predictive capabilities across attributes, achieving weighted F1 scores of 0.82 for tag, 0.90 for time, and 0.99 for both scale and sign. By contrast, few-shot LLMs perform relatively worse. These findings reveal the current limitations of prompt-based approaches and highlight opportunities for future improvement through domain-adaptive pretraining and advanced prompting strategies. Additionally, the dataset helped uncover annotation inconsistencies and potential errors in previous filings, highlighting its value as both a training resource and an auditing aid. From a managerial perspective, the proposed dataset and automated tagging framework can significantly reduce human workload, enhance the reliability of financial disclosures, and enable downstream applications, such as financial analysis of shareholder meeting materials or financial reports published before the mandatory XBRL adoption.