An improved GPT2-based joint event extraction method with position expansion and knowledge augmentation.

Area of Science:

• Natural Language Processing
• Information Retrieval
• Artificial Intelligence

Background:

• The proliferation of internet and social media generates vast unstructured data, increasing the need for efficient information extraction.
• Financial texts present unique challenges including length, redundancy, and complexity, hindering traditional event extraction methods.

Purpose of the Study:

• To develop an advanced model for financial event extraction that overcomes the limitations of existing methods.
• To improve the accuracy and robustness of identifying event types, triggers, and arguments in financial documents.

Main Methods:

• Proposed PosEKE-GPT2, a GPT2-based model reframing event extraction as a text generation task.
• Implemented a joint identification strategy for event types, triggers, and arguments using structured canonical text and sub-task extraction.
• Introduced an expanded positional encoding mechanism for better long-text representation and a knowledge augmentation module for dynamic external knowledge integration.

Main Results:

• PosEKE-GPT2 achieved an average F1-score of 90.61 on the DuEE-Fin dataset and 88.85 on the FewFC dataset.
• The model significantly outperformed baseline models in financial event extraction tasks.
• Ablation studies confirmed the effectiveness of the positional encoding and knowledge augmentation modules.

Conclusions:

• PosEKE-GPT2 demonstrates superior performance and robustness in financial event extraction.
• The proposed methods, including expanded positional encoding and knowledge augmentation, are effective for handling complex financial data.
• The model is well-suited for real-world financial information retrieval applications.