MOGANED

MOGANED

The code is an unofficial implementation of Event Detection with Multi-order Graph Convolution and Aggregated Attention (EMNLP 2019 paper).

Requirements

• tensorflow-gpu==1.10 w/ CUDA 9 (or tensorflow-gpu==1.14 w/ CUDA 10.0)

• stanfordcorenlp (see https://github.com/Lynten/stanford-corenlp for detail)

• numpy

• tqdm

Usage

To run this code, you need to:

1. modify MAVEN dataset path, GloVe file path and stanfordcorenlp path in constant.py
2. Run python train.py --gpu [YOUR_GPU] --mode MOGANED to train.
3. Run python train.py --gpu [YOUR_GPU] --mode MOGANED --eval to get prediction on test set (dumped to results.jsonl).

All hyper-parameters are in constant.py, you can modify them as you wish.

About Preprocessing

When you first run this code, the code will do preprocessing. The preprocessing is quite low and may take a whole night (so run it and you can go to sleep!). This is because getting dependency trees are quite slow.

However, preprocessing will only run once and the preprocessed files will be dumped to the maven dataset path. Next time you run the code the code will read them and won't do any more preprocessing.

Results on MAVEN

We run this code and submit the results to the CodaLab leaderboard (username: wzq016):

Method	Precision	Recall	F1
MOGANED (Paper)	63.4 ± 0.88	64.1 ± 0.90	63.8 ± 0.18
MOGANED (Leaderboard)	64.7 ± 0.05	66.0 ± 0.02	65.3 ± 0.01

P.S. We updated the running environments after the MAVEN paper was published and found that the results magically gone higher.

Note

There are some differences on training strategy between this code and the original MOGANED paper:

1. The code doesn't use BIO schema. This is because trigger words are usually a single word rather than a phrase in ACE05, this won't affect results in ACE05.
2. The code doesn't use L2-norm, only use dropout.
3. The code uses AdamOptimizer rather than AdadeltaOptimizer. During experiments, I found Adadelta can't train a good classifier, however, Adam can.
4. This code sets bias loss lambda to 1 rather than 5 since I found this will make F1 score higher.

Running on ACE 2005

Please refer to this repo.