Event Extraction with BERT on RAMS

Event trigger detection and event-type classification using BERT-based models, evaluated on the RAMS (Roles Across Multiple Sentences) dataset from Johns Hopkins University.

This repository supports my doctoral research on information extraction at UESTC, Chengdu.

---

Task

Event extraction is decomposed into two sub-tasks:

1. Trigger detection - identify the token(s) in a sentence that evoke an event.
2. Event-type classification - assign the correct event type to each detected trigger.

Dataset

RAMS (Roles Across Multiple Sentences) - https://nlp.jhu.edu/rams/

RAMS annotates events together with argument roles that may span multiple sentences, which makes it a harder and more realistic benchmark than single-sentence event corpora.

The dataset is not redistributed in this repository. Download it from the official link above and place it under data/ before running the notebooks.

Repository contents

Notebook	Purpose
trigger_detection_bert2.ipynb	BERT-based trigger detection and event-type classification
trigger_detection_2_gpt.ipynb	Comparative approach using a generative model

Getting started

git clone https://github.com/farhadmohmand66/Event_Extraction.git
cd Event_Extraction
pip install -r requirements.txt

Download RAMS from https://nlp.jhu.edu/rams/ into data/, then open the notebooks with Jupyter:

jupyter notebook

A GPU is recommended for fine-tuning.

Approach

• Pre-trained BERT encoder fine-tuned for token-level trigger classification
• Event-type prediction over the detected trigger spans
• Comparison against a generative (GPT-style) baseline

Results

Task	Model	Precision	Recall	F1
Trigger detection	BERT	TBD	TBD	TBD
Event-type classification	BERT	TBD	TBD	TBD
Trigger detection	GPT baseline	TBD	TBD	TBD

Roadmap

• Publish evaluation metrics and confusion matrices
• Refactor notebook code into a reusable src/ package
• Add requirements.txt with pinned versions
• Extend from trigger detection to full argument-role extraction

Citation

If you use this work, please cite the RAMS dataset:

@inproceedings{ebner-etal-2020-multi,
  title     = {Multi-Sentence Argument Linking},
  author    = {Ebner, Seth and Xia, Patrick and Culkin, Ryan and Rawlins, Kyle and Van Durme, Benjamin},
  booktitle = {Proceedings of ACL},
  year      = {2020}
}

Author

Farhad Khan - PhD student in Computer Science, UESTC, Chengdu

GitHub · LinkedIn

License

Released under the MIT License. See LICENSE.