MaskDGNets: Masked-attention guided dynamic graph aggregation network for event extraction.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Traditional deep learning event extraction methods overlook correlations between word features and sequence information.
• Existing models fail to fully capture intricate associations between events and their attributes.

Purpose of the Study:

• To develop an advanced framework for event extraction that addresses limitations in current deep learning approaches.
• To enhance the understanding of relationships between events and their primary attributes.

Main Methods:

• Introduced a Masked Attention-guided Dynamic Graph Aggregation Network (MaskDGNets).
• Integrated word and character vectors for robust representation.
• Employed a squeeze layer in bidirectional recurrent units for enhanced sequence modeling.
• Utilized a dynamic graph aggregation module for inter-event and event-attribute connections.
• Developed a reconstructed weighted loss function for module supervision.

Main Results:

• MaskDGNets demonstrated robust performance in event extraction tasks.
• Achieved high F1 scores of 81.443% on the DuEE dataset and 87.382% on the CCKS2020 dataset.
• The framework effectively models long-term dependencies and global context.

Conclusions:

• The proposed MaskDGNets framework significantly improves event extraction performance.
• The novel architecture effectively captures complex relationships within event data.
• This approach offers a promising direction for advanced natural language understanding tasks.