Source-linked AI summary
A Scalable Cross-Domain Event Extraction System via a Unified Generative Training Framework
Siting Liang, Omar Adjali, Omair Shahzad Bhatti, Daniel Sonntag
TL;DR
Event extraction remains difficult to scale across separate subtasks, domains, and evolving label spaces. The paper presents a unified T5-based generative framework trained jointly across diverse datasets, supporting pipeline and end-to-end extraction, and demonstrates it in a schema-aware web application. The reported system shows robust performance across datasets, configurations, and ontology sizes while enabling practical, reproducible workflows.
Problem
Existing event extraction methods often separate subtasks or depend on dataset-specific schemas, limiting scalability across domains and large, evolving label spaces.
Method
A unified T5-based sequence-to-sequence framework jointly models event extraction across six datasets and supports both pipeline and end-to-end configurations.
Results
The system demonstrates robust performance across datasets, extraction configurations, and ontology sizes in a complete web-based event extraction workflow.
Takeaways & Limitations
The open-source application supports scalable adaptation, schema-aware extraction, and reproducible analysis without extensive dataset-specific redesign.
Abstract
from arXiv · showhide
Event extraction is fundamental to information extraction. Prior approaches often separate event detection and argument extraction or depend on dataset-specific designs, limiting scalability and cross-domain generalization. We propose a unified generative sequence-to-sequence framework that performs event extraction subtasks jointly and supports both pipeline and end-to-end configurations. We fine-tune pretrained language models on multiple event datasets across diverse domains, enabling a single model to retain domain-specific semantics while generalizing over large and evolving label spaces. We demonstrate these capabilities through a web-based application tailored for researchers and practitioners. The platform supports document upload, schema-aware event extraction, visualization of triggers and arguments, and comparison of different extraction configurations across domains.
1 Introduction
Event extraction converts unstructured text into structured events but remains difficult to scale across domains and changing schemas. The paper addresses this gap with a unified, cross-domain generative system supporting multiple extraction configurations.
- Task background: Event extraction identifies triggers and event types, then extracts arguments and assigns semantic roles for structured event representations.These representations support knowledge base construction.
- Research gap: Prior approaches often separate event subtasks and rely on domain- or schema-specific designs, requiring substantial re-engineering for new datasets.Unified generative methods still face limitations with large label spaces and computational scaling.
- Proposed system: The proposed T5-based framework jointly models event extraction across six datasets and diverse domains while supporting pipeline and end-to-end modes.It is implemented as a web-based system for practitioners.
2 Technology and Innovation
The framework unifies event extraction as sequence-to-sequence generation while preserving pipeline flexibility and supporting cross-domain, evolving label spaces. Task and domain conditioning allow one jointly trained model to handle diverse schemas with limited redesign.
- Unified generative modeling: The framework formulates trigger identification, event type classification, and argument extraction as one unified sequence-to-sequence task.Task codes support input changes without modifying the core architecture.
- Design rationale: Joint modeling reduces dependence on separate models and supports consistent training across datasets, including low-resource scenarios.The framework is designed to balance scalability, flexibility, and domain adaptation.
- Pipeline and E2E modes: Pipeline mode exposes separate extraction stages, whereas end-to-end mode jointly performs trigger identification, event typing, and argument extraction.The two modes use distinct input and target patterns during inference.
- Cross-domain conditioning: Adding dataset and domain identifiers yields the input format {dataset name}+{task prefix}+{text}, guiding schema selection and cross-domain learning.The design accommodates diverse and evolving schemas while retaining domain-specific semantics.
3 Application Features
The application provides configurable, schema-aware event extraction with two-stage and full end-to-end workflows. Its interface supports visual comparison, multi-schema annotation, and reusable structured outputs.
- Application overview: The application supports JSON uploads, document previews, six event schemas, and both domain-specific and cross-domain T5-base models.The models are fine-tuned from a pretrained model with approximately 220 million parameters.
- Two-stage extraction: In the two-stage workflow, Stage 1 detects and merges triggers from pipeline and end-to-end modes before Stage 2 extracts conditioned arguments.Users can configure model selection in both stages.
- Result comparison: Visualization uses offsets, aligned highlights, and event cards to compare configurations and inspect how detection choices propagate to arguments.Results can be examined per document and saved for side-by-side comparison.
- E2E extraction: Full end-to-end extraction runs all event extraction subtasks in one step for efficient large-scale processing.The interface presents all modes in a consistent layout.
- Multi-dataset annotation: Schema-aware annotation applies multiple event schemas to the same document while preserving dataset labels, provenance, and hierarchical JSON outputs.These reusable artifacts support downstream analysis and comparison.
4 Conclusion
The work presents a unified generative approach for cross-domain event extraction, demonstrating robust performance across datasets, extraction configurations, and ontology sizes while supporting practical deployment.
- The system demonstrates robust performance across datasets, extraction configurations, and ontology sizes.
- An open-source application provides a workflow from document ingestion to structured event outputs.
- The application supports both pipeline and end-to-end extraction configurations.
- The system exposes domain-specific and cross-domain models for flexible adaptation.
- The design prioritizes deployment, adaptation to large label spaces, and reproducible analysis without extensive dataset-specific customization.
Ethical statement
The authors used GitHub Copilot during code development and reviewed and validated the generated code, retaining responsibility for the publication.
- GitHub Copilot assisted with code development, after which the authors reviewed and validated the generated code.
- The authors take responsibility for the accuracy, integrity, and final content of the publication.