Input

[Jiang-X-Pro]

thanks for your excellent work in KGC filed.
i have not seen the code yet and i am confused about the input . the input text of entity is the short version like "Gary Rydstrom" or the long version like "Gary Roger Rydstrom is an American sound designer and director. He has won seven Academy Awards for his work in sound for movies."?

[wangbo9719]

Thanks for your attention.

We use the long one as input.

[Jiang-X-Pro]

Thanks for your response , and could you explain the why we need the steps of "do_train" and "do_eval" for getting a new dev subset? I am not clear with the get_new_dev_dict.py about how it works.Hope for your reply!!!

[wangbo9719]

The original dev set is the same size as the test set. If we use the original dev set in the validation stage, it will take about one hour (sorry, I forgot the accurate time) for one validation. And there will be many validations during the training stage, which is very time-consuming.
Therefore, we develop a new dev subset to mitigate this problem. We apply a basic model to get the ranked top-50 candidate entities that are hard to predict for this triple. And the other candidates far from the gold entity are ignored. The selected entity set of each triple replaces the original candidate entity set containing all entities in the graph. In this way, we speed up the validation stage and maintain its effectiveness.

[Jiang-X-Pro]

thanks for your quick reply , do you mean that for the fact (h,r,{t1,t2,...,tn}) , you keep the top-50 t_i which gets the lower score than other entity t_j by Bert(h,r,t_i) < Bert(h,r,t_j) ?

[wangbo9719]

Sorry, I forgot that the Bert(h,r,t_i) is lower or higher than Bert(h,r,t_j) in the implement. Whatever, Bert(h,r,t_i) is closer than Bert(h,r,t_j) and the "closer " means that the t_i is more likely to be mispredicted by the trained model as a gold entity than other t_j.

[Jiang-X-Pro]

Thanks for your reply!
Could you tell me how much memory you use for the evaluation of fb15-237. I use up it with 50GB.

[wangbo9719]

Sorry, I forgot the detail. But I used a 24G GPU before, which is enough.

[Jiang-X-Pro]

Thanks for your great work.
Could you tell me how to use the new dev file "new_dev.dict"?

[wangbo9719]

Add "--do_eval" to the training command of StAR. For example, section 4.1 in README.

[Jiang-X-Pro]

according to the code

StAR_KGC/StAR/kbc/kb_dataset.py

Line 91 in 890dd54

self.data_path = join(self.data_dir, "{}.tsv".format(self.data_type))

, it will use the dev.tsv instead,should I rename the file "new_dev.dict" as "dev.tsv"?

[wangbo9719]

It has been written directly in

StAR_KGC/StAR/link_prediction/train.py

Line 625 in 890dd54

dev_dict = torch.load("./data/" + args.dataset + "/new_dev.dict")

.

[Jiang-X-Pro]

it went wrong when reproducing the results of NELL, usepacka was not found

[wangbo9719]

Sorry for the wrong command. You can just remove the 'usepacka' parameter. It may be a copy mistake in the readme.md writing. I will revise it.