Abstract
Event extraction (EE) is a critical NLP task that involves identifying event triggers and their participants (arguments). However, high-quality human annotation for EE is expensive and time-consuming. This paper introduces DEGREE (Data-Efficient Generation-Based Event Extraction), a model that reformulates EE as a conditional generation task. By utilizing manually designed prompts and a pre-trained encoder-decoder architecture, DEGREE generates natural language summaries of events. This approach allows the model to leverage label semantics and joint dependencies, leading to significant performance gains in low-resource environments.
Introduction
Traditional event extraction models typically treat the task as a sequence labeling or classification problem. These methods often require large amounts of labeled data and fail to capture the semantic meaning of event roles (e.g., treating "Agent" as an opaque category). Furthermore, pipeline-based models that extract triggers and arguments separately cannot exploit the mutual information between them.
Methodology: DEGREE
DEGREE shifts the paradigm from classification to generation. The model takes a passage and a specific Prompt as input. The prompt provides semantic guidance through three components: an event type description, event keywords, and an end-to-end (E2E) template.
The model uses a pre-trained BART-large transformer. During inference, it generates a sentence following the template, filling placeholders with entities from the passage. A deterministic algorithm then extracts the final predictions from this generated text. This design allows the model to use its pre-trained "common sense"—for instance, knowing that a "Place" role should be filled by a location entity.
Experimental Results
The researchers evaluated DEGREE on the ACE 2005 and ERE-EN benchmarks, focusing on its performance when training data is scarce.
Key Findings
- Data Efficiency: In the 1% data regime, DEGREE outperformed the state-of-the-art OneIE model by 15 F1 points in trigger classification and 5 points in argument extraction.
- Semantic Leveraging: The use of natural language templates proved superior to abstract tags, as the model could better utilize its pre-trained linguistic knowledge.
- Joint Learning: By generating triggers and arguments simultaneously, the model effectively captures the dependencies between them, improving overall extraction accuracy.
Conclusion and Future Work
DEGREE demonstrates that generative models are highly effective for low-resource event extraction. By framing the task as natural language generation, the model overcomes the limitations of traditional classification approaches. Future work will focus on improving inference efficiency—as the current model must iterate through event types—and automating the creation of high-quality templates.
Summary updated