Span-based few-shot event detection via aligning external knowledge.

Area of Science:

• Natural Language Processing
• Machine Learning
• Artificial Intelligence

Background:

• Few-shot Event Detection (FSED) is crucial for identifying novel event types in new domains with limited data.
• Existing Prototypical Network (PN)-based methods struggle with learning token-wise label dependencies and creating accurate prototypes.

Purpose of the Study:

• To propose a novel span-based FSED model, SpanFSED, to overcome the limitations of previous PN-based methods.
• To enhance the learning of token-wise label dependency and the accuracy of event prototypes in FSED.

Main Methods:

• SpanFSED decomposes FSED into two subprocesses: a span extractor and an event classifier.
• Span extraction converts sequential labels into a global boundary matrix for precise boundary information acquisition.
• Event classification aligns event types with knowledge bases (e.g., FrameNet) and enhances support sets with trigger information for improved prototyping.

Main Results:

• SpanFSED demonstrates superior performance compared to existing methods.
• Experiments conducted on four diverse datasets (ACE2005, ERE, MAVEN, FewEvent) validate the model's effectiveness.
• The proposed methods significantly improve boundary information acquisition and prototype accuracy.

Conclusions:

• SpanFSED offers a more effective approach to Few-shot Event Detection by addressing key limitations in prior methods.
• The span-based decomposition and enhanced classification strategy contribute to improved performance in identifying novel event types.
• The study provides accessible code and data for reproducibility and further research in FSED.