Research on Financial Event Extraction and Type Expansion Methods

Abstract

To tackle the challenges posed by scattered event arguments and event overlaps in financial event extraction task, In this study, we introduce an FMTL-UIE framework that integrates Universal Information Extraction with multi-task learning. This approach integrates FinBERT, a finance-oriented pre-trained language model, with the UIE framework, establishes a multi-task learning architecture for identifying event types and extracting event arguments, while introducing an adaptive prompting mechanism to enhance extraction performance in complex financial contexts. To overcome the limited diversity of event categories and the poor generalization to unseen events in current datasets, this study develops a clustering-driven event ontology augmentation framework with the support of large language models for efficient annotation. Experimental results show that FMTL-UIE achieves F1 scores of 86.9%, 94.2%, and 98.6% for event type detection on FewFC, DuEE-Fin, and a self-built FinEE dataset, respectively, and F1 scores of 75.3%, 76.1%, and 80.8% for event argument extraction, respectively, confirming the effectiveness and robustness of the proposed approach.