Giter VIP home page Giter VIP logo

Comments (11)

lijiaoyang avatar lijiaoyang commented on July 26, 2024

关系抽取排名网站上,REDN的F1值变为0,并且标注FAKE

REDN

from redn.

loujie0822 avatar loujie0822 commented on July 26, 2024

其实移除就可以了,还给标了fake。

from redn.

ARDUJS avatar ARDUJS commented on July 26, 2024

现在咋又上去了,还有HBT之前是89.5现在咋变成了87.5,这是啥原因,求大佬解答
image

from redn.

transformerzhou avatar transformerzhou commented on July 26, 2024

现在咋又上去了,还有HBT之前是89.5现在咋变成了87.5,这是啥原因,求大佬解答
image

同求,懵逼,不知道什么情况

from redn.

pangjiajia avatar pangjiajia commented on July 26, 2024

论文里提到本文不关注实体抽取的方法,我看了代码,作者有提供add_subject_loss的参数,可以选择增加/不增加实体抽取部分。不过不清楚他最终提交的测试结果是否有加上。

from redn.

pangjiajia avatar pangjiajia commented on July 26, 2024

直接下载下来的代码默认add_subject_loss参数是false,不知道我理解的对不对

from redn.

pangjiajia avatar pangjiajia commented on July 26, 2024

仔细看了他的测试结果评价部分,确实没有考虑实体抽取的准确率

from redn.

slczgwh avatar slczgwh commented on July 26, 2024

要不是今天收到公司信息说我学术造假,我还没注意到有这样的issue。先回答issue:
1.关系抽取利用实体gold。我们从来就没有说过我们是联合抽取,文章中也只是提了可以做,而我们关注就是关系抽取本身.而关系抽取向来就是实体识别的下游任务,以实体作为已知结果非常自然。比如我们引用的nyt这个数据集联合抽取在paperwithcode上有另一个榜单:联合抽取榜单,关注联合抽取请移步这里。
2.计算分数代码的问题。这段代码是按照同样的逻辑从清华opennre中改造的。因为我们用了一个矩阵的关系,没法直接复用代码,就只能按同样逻辑改造。代码链接在这里,注释逻辑非常清晰,sentenceRE的入参就包含头尾实体,自己去读一下就知道逻辑是否一致了。(还有这一段test代码,看看传进去的是什么。)另外负样本在nyt中的使用方法已经在文章中说的很清晰,代码逻辑也很清晰,这里不多说了。
最后补充一下,关系分类和关系抽取只是名字的问题,怎么定义也没有很清晰的说法。很多已发表文章都把对句子中实体对的关系分类叫关系抽取,这件事情并不是我们的发明创造(不然sentenceRE这个名字从何而来?干嘛不叫sentenceRC?)。倒是联合抽取有比较明确的说法,实体跟关系同时抽取出来就算联合,否则就是pipeline形式。而且很显然,关系抽取必然会利用实体信息,否则pipeline的级联误差从何而来?另外,这个讨论是用于讨论本文的代码的,关心其他模型的请移步其他模型的开源代码区。如果关注所谓的比较公平性,可以发邮件给作者说明这个问题,不要在git上讨论这个(尤其是中文讨论),这也太业余了。

from redn.

sidney1994 avatar sidney1994 commented on July 26, 2024

作者您好,我看了您的回复。作者回复中提到了“我们从来就没有说过我们是联合抽取”,但作者论文中比较的baseline中, GraphRel2p和HBT均为联合抽取模型,这是不公平比较,希望作者修正论文,并撤下paperwithcode下的结果。

from redn.

slczgwh avatar slczgwh commented on July 26, 2024

@sidney1994 paperwithcode上面的结果已经不知道被谁删除了。此外,我们所挂的榜单并没有问题,在被删除之前一直挂在RE下面。倒是HBT需要check一下看看是否挂错了地方。目前我看到NYT数据集下,HBT还挂在正常的RE下面,没有挂在联合抽取下。WebNLG也是一样的。文章中的联合抽取的问题确实需要修改,我们会更新一下我们把实体抽取模块增加上后获得的分数,再用于与基线对比。近期会完成这个修改。

from redn.

pvcastro avatar pvcastro commented on July 26, 2024

I kind of agree with @slczgwh here:

Relation extraction has always been a downstream task of entity recognition, and it is natural to use entities as known results.

The purpose of the benchmarks are to predict relationship between entities. I think Joint extraction is a way to get there, not necessarily a benchmark by itself...authors claim to use it not because they wish to do it "the hard way", but because they think (empirically, of course) it's best option to have a model learn to classify relations. I believe pretty much all the papers I have read so far regarding joint relation extraction mention this. If I'm wrong about this, please point me to a reference that states otherwise.

And I also do agree that this is a very confusing downstream task to benchmark, because there isn't a clear standard regarding the source of all diferent datasets (I have seen many different versions of NYT so far, with different data, annotations, criteria, etc), annotation criteria, evaluation, etc. In practice, I think it's hard to compare any model with each other I have seen so far.

from redn.

Related Issues (20)

Recommend Projects

  • React photo React

    A declarative, efficient, and flexible JavaScript library for building user interfaces.

  • Vue.js photo Vue.js

    🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.

  • Typescript photo Typescript

    TypeScript is a superset of JavaScript that compiles to clean JavaScript output.

  • TensorFlow photo TensorFlow

    An Open Source Machine Learning Framework for Everyone

  • Django photo Django

    The Web framework for perfectionists with deadlines.

  • D3 photo D3

    Bring data to life with SVG, Canvas and HTML. 📊📈🎉

Recommend Topics

  • javascript

    JavaScript (JS) is a lightweight interpreted programming language with first-class functions.

  • web

    Some thing interesting about web. New door for the world.

  • server

    A server is a program made to process requests and deliver data to clients.

  • Machine learning

    Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.

  • Game

    Some thing interesting about game, make everyone happy.

Recommend Org

  • Facebook photo Facebook

    We are working to build community through open source technology. NB: members must have two-factor auth.

  • Microsoft photo Microsoft

    Open source projects and samples from Microsoft.

  • Google photo Google

    Google ❤️ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.