BenfengXu
BenfengXu
> hey @BenfengXu i am using ssan(https://github.com/PaddlePaddle/Research/tree/master/KG/AAAI2021_SSAN). > I have followed the steps as per readme but still not getting any result.json file. > Following are logs of my trial....
> 想请问论文中的意思是说,对于每种类型的依赖,比如intra+coref,inter+relate,文中的意思是说每种类型共享一个神经网络?这边好像分了6种类型,就是这一块总共有6个独立的神经网络?然后根据sij的值来选择,具体使用哪个网络计算? 6种类型,除NA外有5种类型,每种都有它自己对应的module,module的结构是一样的,module的参数每个s_ij各不相同,一共有5种,比如矩阵A,网络里一共存储了5个,根据sij来决定具体A的参数是五个中的哪一个。
> 还有就是这一块里提到的bl,sij,没有看懂是什么意思,具体而言这个值是怎么得出的呢?不是训练得到的?还是人为给六种依赖设置了六个值? 分别随机初始化,然后经过梯度传播,网络训练,学习得到的。和其他的矩阵module是一样的,只不过bl,sij的参数只有一维,即一个数值。
The pre-training code for distantly supervised data is identical to the one in this repo for the train set data. Just pass the DocRED distant data file, set your training...
> 您好,非常感谢您的工作! 但是想请问一下,如何对更一般的句子做预测,而不是docred的测试集 > > 如只提供这个句子中:A man wear white clothes play cricket 抽取出关系: wear 或者 play? > > 或者需要去构造docred数据集的格式? > > 感谢! 将你的数据转换成DocRED的格式即可,这是最简单的方法。 不过需要注意,如果使用在DocRED上训练好的模型,预测的结果也是和DocRED的schema一致的。如果你的数据分布、schema设定不一致的话,需要在自己的数据上重新训练。
> DocRED的数据格式是json,每条数据包括实体、mention位置等等,这些在数据集里都是已标注的,你也需要提供同样的信息,比如A man, clothes作为实体,然后以同样的格式写入数据的vertexSet, Sents等字段。
> Hi @BenfengXu , I have two follow-up questions. > > 1. If have extracted entities and their positions using NER, does it mean that I can convert to json...
> Hi, may I ask if anyone know what causes this Stop Iteration? > > These codes are inherited from the huggingface transformers implementation, you can find similar issuse there:...
> Hi, > Have you shared your trained models anywhere so that we can use them to reproduce the results on test data? > If not, would you share the...
> Is it possible to share the checkpoint that achieved 65% F1 though? Sry, the distant data pretrained ckpt will not be released as this is a joint work with...