Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

关于results of ablation study #3

Closed
AnWang-AI opened this issue Oct 23, 2021 · 5 comments
Closed

关于results of ablation study #3

AnWang-AI opened this issue Oct 23, 2021 · 5 comments

Comments

@AnWang-AI
Copy link

您好,非常感谢您非常优秀的工作成果,让我重新认识了relation识别对entity识别有帮助的可能。在阅读paper时,有一些疑问,希望能够得到解答。

  1. 文章中说到在SciERC上进行了ablation study实验,但是不同ablation的结果都很相近,请问下您的结果是单轮的训练结果吗?我运行了您的代码,发现SciERC上的结果非常不稳定,在不同random seed上结果变化较大,您是否能够给出显著性的值呢?或者在其他数据集上ablation study的结果和SciERC数据集上结论相同吗?
  2. ablation study实验结果中只显示了relation extraction的结果,请问不同方式对entity extraction结果有怎样的影响呢?是否有导致relation extraction结果高但是entity extraction结果低的情况呢?
  3. 您使用了partition操作但是并没有验证这个操作的效果,请问在使用filter操作的情况下,使用partition操作和其他方式的实验结果区别是怎样的呢?
    非常感谢!
@Coopercoppers
Copy link
Owner

您好,我来回答一下您的问题

  1. 是单轮的训练结果。全部实验都是random seed = 0,因为encoding scheme比较重要,所以我在webnlg上也是做过测试的,sequential和parallel大概比joint低个0.5-1.5点左右。SciERC结果不稳定,想能问下浮动能大到多少呢?浮动大的原因可能是超参没有怎么调,可以试试一些稳定训练的trick,比如说warmup等。
  2. entity extraction确实会出现您说的这种情况,这部分可能是RE部分没有cover实体尾部的结果,可以试试PFN-nested。
  3. 这个没有做过。因为没有partition,也就得不到三个分区了,这样的话也就不知道是对哪个部分进行filter了。

@AnWang-AI
Copy link
Author

  1. 我使用了您的代码,重新运行了在seed = [0,1,2,3,4]的情况。NER的结果在65.67-68.59,RE的结果在36.23到37.86。方差还是比较大的。
  2. 如果方便的话,请问您是否可以在ablation study添加上entity extraction的结果呢?看到您的代码是同时输出NER和RE两个结果的。

非常感谢您的回答。希望您科研顺利。

@Coopercoppers
Copy link
Owner

Coopercoppers commented Oct 23, 2021

RE结果和我这差的有点多。我跑了下1,2,3,4的seed,NER的结果是66.9, 66.4, 68.0, 67.7, RE的结果是36.9, 36.3, 38.9, 37.7。我在readme上更新了cuda版本要求,貌似这个也是会有影响的。

ablation我这里只存了sequential和parallel,sequential是68.7, parallel是67.0,scierc上确实波动很大,这部分我觉得可能是因为训练数据集本身的原因导致的。webnlg上NER sequential和parallel比joint差个0.3-0.5左右。

由于论文是最终稿,emnlp不让改了,所以可能没法添加结果。希望上边几个数值对您有用。

@AnWang-AI
Copy link
Author

好的,谢谢您的回复。论文中的内容虽然无法额外变动,但是希望能够在github项目的readme上更新实验结果。同时建议您的ablation study能够在5-10个random seed上取NER和RE的平均结果和方差。单次的RE结果恐怕不能够使得您论文中ablation study的结论令人信服。

@Coopercoppers
Copy link
Owner

Coopercoppers commented Oct 24, 2021

嗯,谢谢您的建议,ablation的多次结果之后会考虑加上的。

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants