OmniEvent is a comprehensive, unified, and modular toolkit for event extraction (EE), designed for researchers and practitioners in Natural Language Processing. It addresses the challenges of inconsistent evaluation and diverse modeling paradigms in EE by providing standardized data preprocessing, output formats, and evaluation metrics, enabling fair comparisons across different approaches.

How It Works

OmniEvent adopts a modular architecture, decomposing models into Input Engineering, Backbone, Aggregation, and Output Head modules. This allows for flexible combination and reimplementation of various components, supporting diverse paradigms like Token Classification, Sequence Labeling, MRC (QA), and Seq2Seq, as well as both Transformer-based and classical models. It also supports both English and Chinese languages.

Quick Start & Requirements

• Install via pip: pip install OmniEvent
• Tested on Python 3.9+, PyTorch 1.12.1+.
• Official documentation and examples are available.

Highlighted Details

• Supports event detection and event argument extraction, either jointly or independently.
• Offers consistent evaluation across different paradigms through standardized preprocessing and output conversion.
• Implements various paradigms including Token Classification, Sequence Labeling, MRC, and Seq2Seq.
• Compatible with 🤗 Transformers and its Trainer for easy integration and customization.

Maintenance & Community

• The project is actively maintained, with a roadmap indicating future development of an agent system based on LLMs for IE tasks.
• The team has released LLMs (ADELIE) for information extraction.

Licensing & Compatibility

• The repository does not explicitly state a license in the provided README.

Limitations & Caveats

• The project will transition to serving only as introductory code for EE, with major updates ceasing in Spring 2024, focusing on a new agent system.