Event Extraction: Overview, Methods, and the OmniEvent Toolkit

Event extraction aims to identify structured event information (type, time, location, participants) from raw text, a crucial component for enriching knowledge graphs and supporting downstream applications such as question answering, intelligence mining, and drug side‑effect analysis.

The task originated from DARPA and has attracted substantial funding; modern knowledge graphs contain millions of entities but only hundreds of thousands of events, highlighting the need for richer event representations.

Four dominant paradigms have emerged for event extraction:

Sequence labeling – formulates extraction as a token‑level tagging problem.

Token classification – classifies each token into event or argument categories.

Machine Reading Comprehension (MRC) – poses extraction as answering event‑related questions over a passage.

Sequence‑to‑sequence generation – treats extraction as a text‑generation task.

Each paradigm has representative works and distinct trade‑offs in complexity and performance.

The OmniEvent toolkit provides a unified implementation of these paradigms, offering:

Standardized data formats and preprocessing scripts for fair benchmarking across datasets.

Comprehensive algorithm coverage, including both Transformer‑based and traditional CNN/LSTM models, supporting Chinese and English.

Modular architecture that allows users to mix and match components or add custom modules.

Support for large‑scale model training (e.g., T5‑11B) via BMTrain.

Simple three‑line API for quick inference.

Future work includes extending OmniEvent to document‑level event extraction, few‑shot and semi‑supervised settings, event induction, and event relation extraction to capture temporal and causal links between events.

A short Q&A addresses practical concerns such as which paradigm to start with (recommendation: sequence labeling for rapid prototyping) and the roadmap for new task settings.