Due to the licence issues, we cannot share the source ACE2005 dataset or the preprocessed data.
So we specify the data format as follows and you can preprocess the data follow the format.
The data used in our model should be preprocessed to follow files (Tag means dataset tag, such as ACE_Train, ACE_Test, DS_Train):
Word embedding indices for input sentences, each sentence's length must be equal to SenLen in constant.py. The padding index is 0.
Its shape should be (BatchSize, SenLen).
[[1,2,3,4,5,6],[2,4,2,1,4,0]]
Relative position embedding indices to the candidate trigger for each instance. For instance, for a token in the tokenized sentence, its position embedding
index is TokenPosition-TriggerPosition+SenLen. The position embedding index for padding token is 0.
Its shape should be (BatchSize, SenLen).
[[4,5,6,7,8,9],[2,3,4,5,6,0]]
Lexical feature for DMCNN. It is the word embedding indices for the words whose distance to the candidate trigger is less than LocalLen in constant.py.
Its shape should be (BatchSize, 2*LocalLen+1).
[[2,3,4],[1,4,0]]
Event type label for each instance, the NA is 0.
Its shape should be (BatchSize).
[1,0]
Mask matrix for the words on the left of the candidate trigger.
Its shape should be (BatchSize,SenLen).
[[1.,1.,1.,0.,0.,0.],[1.,1.,1.,1.,1.,0.]]
Mask matrix for the words on the right of the candidate trigger except the padding tokens.
Its shape should be (BatchSize,SenLen).
[[0.,0.,0.,1.,1.,1.],[0.,0.,0.,0.,0.,0.]]
The input mask for BERT, which includes all words except the padding tokens.
Its shape should be (BatchSize,SenLen).
[[1.,1.,1.,1.,1.,1.],[1.,1.,1.,1.,1.,0.]]
The overall index for each instance, which is used to find corresponding original data item. (after random shuffle)
If you want it, uncomment corresponding lines in dataset.py.
Its shape should be (BatchSize).
[1,2]