A Joint Learning Framework for Document-Level Event Extraction.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Information Extraction

Background:

• Document-level event extraction (DEE) is complex due to scattered arguments forming variable-length event lists.
• Sequential autoregressive methods limit information interaction and propagate classification errors.

Purpose of the Study:

• To develop a DEE method that enhances interaction between global event and local argument information.
• To reduce error propagation from event type classification to argument extraction.
• To support the extraction of variable-length event lists across multiple sentences.

Main Methods:

• Proposed an event- and argument-aware attention mechanism to mitigate error propagation.
• Introduced a joint learning framework (JLF) for improved event and argument information interaction.
• Designed a complete event topology decomposition (ETD) for handling variable-length event lists.

Main Results:

• Achieved new state-of-the-art performance on three public datasets.
• Demonstrated significant improvements: 10.6% on ChFinAnn, 5.6% on DuEE-Fin, and 14.8% on FNDEE.
• The proposed method effectively addresses challenges in DEE.

Conclusions:

• The joint learning framework with event- and argument-aware attention and ETD significantly advances DEE.
• The method shows superior performance and robustness across different financial datasets.
• This work provides a more effective approach for extracting complex event structures from documents.