Reinforcement learning improves event extraction from text
EAGER: Enhancing Generative Event Extraction via Reinforcement Learning with Verifiable Rewards
Computation and Language
Summary
Extracting detailed events from text is hard because it requires spotting event triggers, categorizing event types, and pulling out important parts accurately. The authors present a method called EAGER that uses a special kind of learning rewarding detailed, verifiable accuracy at several levels to improve extraction. Their approach balances different goals like accuracy and avoiding extra unrelated info, making the system better at understanding events. Tests on seven datasets show EAGER beats other recent methods consistently.
What this means in practice
- •For information extraction engineers: Build automated systems that extract structured event information from documents more accurately using fine-grained reward-guided training.
- •For legal technology developers: Improve software that identifies key events and related details in legal texts for case analysis and summarization using advanced event extraction.
Authors
Omar Adjali, Siting Liang, Omair Shahzad Bhatti, Daniel Sonntag
Abstract
End-to-end event extraction remains challenging for large language models as it requires simultaneous identification of event triggers, classification of event types, and extraction of schema-grounded argument spans. We present EAGER, a reinforcement learning framework for generative event extraction that combines fine-grained verifiable rewards with Schema-Contrastive Advantage Estimation to alleviate advantage collapse under sparse binary rewards. Our reward design explicitly targets structural validity, extraction accuracy, groundedness, coverage, over-generation, and span precision. Experiments across seven benchmark datasets show that EAGER consistently outperforms prompting, supervised fine-tuning, and prior reinforcement learning baselines, achieving a substantial improvement over the strongest prior method. Results demonstrate that task-aligned verifiable rewards and contrastive advantage estimation substantially improve structured extraction.