MOGANED

An unofficial pytorch code reproduction of EMNLP-19 paper "Event Detection with Multi-Order Graph Convolution and Aggregated Attention"

Prerequisites

1. Prepare ACE 2005 dataset.(You can get ACE2005 dataset here: https://catalog.ldc.upenn.edu/LDC2006T06)

2. Use nlpcl-lab/ace2005-preprocessing to preprocess ACE 2005 dataset in the same format as the data/sample.json.

Usage

Preparations

1、put the processed data into ./data, or you can modify path in constant.py.

2、put word embedding file into ./data, or you can modify path in constant.py. (You can download GloVe embedding here: https://nlp.stanford.edu/projects/glove/)

Train

python train.py

All network and training parameters are in constant.py. You can modify them in your own way.

About the word embedding, we found that wordemb in the way (train the word embedding using Skip-gram algorithm on the NYT corpus) got better performance than the glove.6B.100d. So we choose 100.utf8 (you can get it here https://github.com/yubochen/NBTNGMA4ED) as our word embedding vector.

Result

Performance

Method	Trigger Classification (%)
	Precision	Recall	F1
MOGANED(original paper)	79.5	72.3	75.7
MOGANED(this code)	78.8	72.3	75.4

Note

In many cases, the trigger is a phrase. Therefore, we treat consecutive tokens which share the same predicted label as a whole trigger. So we don't use BIO schema for trigger word. This strategy comes from "Exploring Pre-trained Language Models for Event Extraction and Generation" (ACL 2019), Yang et al. [paper]

Reference

• Event Detection with Multi-Order Graph Convolution and Aggregated Attention (EMNLP 2019), Yan et al. [paper]
• Nlpcl-lab's bert_event_extraction repository [github]
• lx865712528's EMNLP2018-JMEE repository [github]