Abstract
Traditional event extraction (EE) methods have transitioned from classification-based pipelines to generative frameworks. However, these generative models often rely on static, manually engineered prompts that fail to capture sentence-specific context. This paper discusses GTEE-DYNPREF, a generative model that utilizes dynamic, learnable prefixes to steer the extraction process. By integrating sentence-level context and employing an irrelevance classifier, GTEE-DYNPREF achieves state-of-the-art performance on ERE benchmarks and demonstrates superior portability to unseen event types.
1. Introduction
Event Extraction (EE) is a core task in Natural Language Processing (NLP) aimed at identifying structured information—specifically, triggers and their associated arguments—from unstructured text. Traditionally, this was approached as a sequence of classification tasks (e.g., identifying triggers and actors separately).
Recent advancements have shifted the paradigm toward generative methods, where models are prompted to fill templates (e.g., "The [Person] was arrested by [Agent]"). Despite their promise, these methods are often hindered by two major hurdles: suboptimal, static prompts that are difficult to optimize, and a failure to account for the broader context of the input sentence or the relationships between different event types.
2. Methodology: GTEE-DYNPREF
The proposed GTEE-DYNPREF (Generative Template-based Event Extraction with Dynamic Prefix) framework is built upon the BART encoder-decoder architecture. It introduces three primary innovations to overcome the limitations of static prompting.
2.1 Continuous Prefix-Tuning
Rather than relying on discrete text-based prompts, which are difficult to optimize and often brittle, GTEE-DYNPREF utilizes continuous prefix-tuning. This involves prepending sequences of tunable continuous vectors to the transformer’s hidden states. These prefixes serve as learnable parameters that "steer" the model toward the specific requirements of the event extraction task more efficiently than human-engineered text.
2.2 Dynamic Context Integration
A key contribution of this work is the "dynamic" nature of the prefixes. The model employs a BERT-based context encoder to process the entire input sentence. Through a multi-head attention mechanism, this context is fused with various event-type prefixes. This ensures that the prompt is not static but is instead uniquely tailored to the specific nuances and relationships present in the current sentence, such as the association between a "Conflict" event and a subsequent "Arrest."
2.3 Irrelevance Classifier (IC)
Data imbalance is a significant challenge in EE, as over 70% of sentences in standard datasets do not contain any events. GTEE-DYNPREF incorporates a binary Irrelevance Classifier to filter out these "empty" sentences before they reach the generative stage. This pre-processing step prevents the generative model from attempting to extract information from noise, thereby significantly boosting precision.
3. Evaluation and Results
The model was evaluated on two gold-standard benchmarks: ACE 2005 (ACE05-E and ACE05-E+) and ERE (ERE-EN). These datasets represent the standard for measuring performance in trigger and argument extraction across dozens of event types.
3.1 Performance Metrics
GTEE-DYNPREF established a new state-of-the-art on the ERE-EN dataset, achieving an F1 score of 66.9% for trigger classification and 55.1% for argument classification, outperforming previous high-performing models like ONEIE. On the ACE 2005 benchmark, the model achieved a 72.6% F1 score for triggers, making it the top-performing generative model and highly competitive with specialized classification-based systems.
3.2 Transfer Learning and Portability
The model's ability to generalize was tested in "new event" scenarios, where the model is trained on certain event types and tested on others. GTEE-DYNPREF showed a significant performance advantage over baselines in these scenarios, suggesting that learnable continuous prefixes capture more transferable features and require less data to adapt to new domains than rigid text prompts.
4. Discussion and Conclusion
The success of GTEE-DYNPREF highlights the importance of context-aware steering in generative models. By moving from discrete, static prompts to continuous, dynamic prefixes, the model effectively captures the "signature" of events within their specific linguistic surroundings. Furthermore, the inclusion of an Irrelevance Classifier demonstrates that for generative models, knowing when to remain silent is as critical as the generation itself.
While the prefixes are dynamic, the underlying event templates remain based on predefined ontologies. Future research will likely focus on automated template induction—learning the structure of events themselves—and further refinement of the filtering mechanisms to maximize noise reduction.
Summary updated