mayubo2333 / paie Goto Github PK

hello,
how to chage the number of gpu?
It doesn't work when I add ''os.environ['CUDA_VISIBLE_DEVICES'] = "2"", and I didn't find where is the code that you set gpu.

您好，请问这个模型如何解决一个事件参数有多个值的情况？

thanks for sharing! how you get event type?

It may be a silly question but haunts in my mind. As mentioned in the paper, pesudo tokens in soft prompts are learnable. However, I found the training process of PAIE with soft prompt is pretty similar to PAIE with manual prompts. They share the same model architecture which means, when backwarding, the loss object are from the same model. So how to learn pseudo tokens? In the paper, you mentioned that you follow two previous method, but how to implement them in PAIE code?

why the prompt template have "prompt start" and " end" ,which is not show in the paper . Is it get an better result ?

Hello, I don't understand why W should be divided by 2 in line 208 of the code below. I found that this will cause the position of event_trigger to become a negative value, so I think W should not be divided by 2 in line 208. I hope to get your answer. Thanks！

ACE05数据集在LDC网站上好像不是免费下载的，请问您是否可以提供一下这个数据集呢？希望能尽快得到您的回复！

Hello, nice work, and thanks for sharing the code!

I have a question about the RAMS dataset. It looks like some arguments get dropped because of the window size. Were the dropped arguments included in the final test set? Thank you very much!

It may be a silly question but haunts in my mind. As mentioned in the paper, pesudo tokens in soft prompts are learnable. However, I found the training process of PAIE with soft prompt is pretty similar to PAIE with manual prompts. They share the same model architecture which means, when backwarding, the loss object are from the same model. So how to learn pseudo tokens? In the paper, you mentioned that you follow two previous method, but how to implement them in PAIE code?

Hello author,

Thanks for sharing the code. I have one question when reading the paper w.r.t. the result comparison with the baseline BART-Gen. Some results in Table 2 of your paper have mismatches with the results in the paper of BART-Gen. For example, the Arg-I Arg-C Head-C numbers for WIKIEVENTS are 66.8 62.4 65.4, but I can't find exactly the same numbers in either Table 5 or 6 in the paper of BART-Gen. Are these results reproduced by yourself? If so, it seems that they are very different from the numbers in the original paper.
no offense, but just confused about this issue. Did I miss some experimental details?

Hello, thanks for the work you do!
I have some doubts about the trigger-argument distance experiment, can you provide me with more details of the implementation process?

Thanks for your code! I have some questions about how to add a new dataset. I think there are 3 steps to add a new dataset:

1. Place the data in the ./data folder.
2. Add data reading code in DSET_processor class in processor_base.py file. And modify some details such as arg parser and build_processor.
3. Add the description file in ./data/dset_meta folder and the prompt file in ./data/prompts folder.
   Then I can run the code to train/infer on my own dataset. Are there any misunderstandings? Thank you!

您好，请问该指标只用判断 论元的head span正确就可以了吗？是否还要求角色也识别正确？
因为我看BartGen代码里似乎是head word span + role name

I cannot find it in models