MOGANED-Implementation

The code is an unofficial implementation of Event Detection with Multi-order Graph Convolution and Aggregated Attention (EMNLP19 paper). (The official code is not released, and this code is for a reference.)

Requirments

tensorflow-gpu==1.10

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

numpy

tqdm

Usage

To run this code, you need to:

1. put English folder of ACE05 dataset into ./, or you can modify path in constant.py. (You can get ACE2005 dataset here: https://catalog.ldc.upenn.edu/LDC2006T06)
2. put stanford language model into ./, or you can modify path in constant.py. (You can download here: https://stanfordnlp.github.io/CoreNLP/history.html)
3. put GloVe embedding file into ./glove folder, or you can modify path in constant.py. (You can download GloVe embedding here: https://nlp.stanford.edu/projects/glove/)
4. Run python train.py --gpu 0 --mode MOGANED to run with MOGANED model. Run python train.py --gpu 0 --mode DMCNN to run with DMCNN model.

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

To run it on MAVEN Dataset, see MAVEN dataset repo.

Dataset

Due to license limitation, we can't distribute datasets directly, please download the dataset by yourself. The download link is given in Usage part.

The code will automatically extract information of ACE2005 dataset and dumps them into json format(train.json ,dev.json and test.json) into path ACE_DUMP in constant.py. This is implented in class Extractor in utils.py.

Each file is composed of a list, which elements are instances with following format:

{
    "tokens": XX,           #tokens of a sentence, a list with string elements
    "start": XX,            #starting offsets of the sentence in original files, an integer
    "end": XX,              #ending offsets of the sentence in original files, an integer
    "offsets":XX,           #offsets of each tokens, a list with tuple elements
    "trigger_tokens":XX,    #tokens of trigger words, a list with string elements
    "trigger_start":XX,     #start index of trigger words of tokens, an integer
    "trigger_end":XX,       #end index of trigger words of tokens, an integer
    "trigger_offsets":XX,   #offsets of trigger words, a list with tuple elements
    "event_type":XX,        #event type of tokens with given triggers, a string
    "ner":XX,               #ner tag of each token, a list
    "pos":XX,               #pos tag of each token, a list
    "dependency":XX,        #dependency parsing results of tokens with StanfordCoreNLP format, a list
    "file":XX,              #file name without suffix
    "dir":XX,               #dir name
    "entities":XX           #entitie in this sentencem, a list with entity elements
}

Each entity is a dictionary with following format:

{
    "token":XX,             #tokens of the entity, a list with string elements
    "role":XX,              #role of the entity when trigger is given, a string
    "offsets":XX,           #offsets of entity, a list with tuple elements
    "start":XX,             #start offset of entity, an integer
    "end":XX,               #snd offset of entity, an integer
    "idx_start":XX,         #start index in tokens, an integer
    "idx_end":XX            #end index in tokens, an integer
}

Results

Depends on the split of train/dev/test, results will have some change accordingly, but won't change much. I use this split and get following results:

Method	Precision	Recall	F1
MOGANED(Paper)	79.5	72.3	75.7
MOGANED(This code)	72.4	71.0	71.7

Note

There are some differences on training strategy between this code and the original 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. From my personal experience, this won't affect results much.
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.

TODO

The code structure is quite like another repo, I will merge these codes if I have time in future.