Confusion about argument detection evaluation

[xixy]

Confusion about argument detection evaluation

[wzq016]

For second case, in our original experiments, we used tuple, i.e. (event type, role), to do evaluation, and classify N/A class only and if only event type is N/A or role is N/A. This code is fully rewritten because the origin code is quite dirty and I apologized that I missed this point in this rewritten code. Now I fixed this point , please check commit history to find out the modification, I also tested the modified codes and the results remained same as the paper.

Just a reminder, once you use tuple to do classification, since the evaluation is more strict, you need to do some more process in dataset to reach 53.5 in DMCNN, such as cutting sentences if its length is more than a threshold, please do such process by yourself. Also, from the experiments results, once DMCNN reaches 53.5 using these two evaluation metrics separately, then other models will perform similarly using these two evaluation metrics separately, which means it is just a trick in data preprocess, and this is the one of main disadvantages of ACE2005.

For first case, when a trigger is classified with N/A, it is not necessarily to do EAE stage with this trigger because all entities are N/A in such situation and are totally noises when testing EAE. Also, actually all models are tested in this way, if you want to see what will happen if you take the first case into consideration, you can simply comment line 167,168 in models.py,i.e.
dev = [np.take(d,dev_slices,axis=0) for d in dev]
test = [np.take(d,test_slices,axis=0) for d in test]
You will see that there will be a dramatically drop in performance and DMCNN will have no way to reach 53.5 since there are too many meaningless noises. From this perspective, hope you can understand what happened here.

[xixy]

Thank you for your response.

I'm okay with the second case and thank you for modification.

For the first case, if we mislabel a trigger as None type, the trigger and its arguments are not included in EAE stage for performance calculation. I got your point but it doesn't looks like a right way.

Thank you again for modification.

[xixy]

Another question is what event detection model is used when you report HMEAE(CNN) and HMEAE(BERT) in Table 3/4? Is DMCNN used as ED model? The experimental setting is confusing for me since you used two event detection models:

As our work does not involve the event detection stage, we conduct the ar- gument role classification based on the event detection models in Chen et al. (2015) and Wang et al. (2019) for HMEAE (CNN) and HMEAE (BERT) respectively.

Or HMEAE(CNN) used DMCNN as ED model and HMEAE(BERT) used model from Wang et al. (2019) as ED model?

Thank you.

[wzq016]

Second one, i.e. DMCNN as ED + HMEAE(CNN) as EAE, DMBERT as ED + HMEAE(BERT) as EAE.

[xixy]

Got it and Thanks!