老哥，我想引用你成果作为参考文献，请问你有英文链接吗，我在谷歌上没找到，很急，谢谢老哥！

请问一下，我在复现实验的时候，在运行“004_数据预处理（替换英文字符)时出错”，报错信息：FileNotFoundError: [Errno 2] No such file or directory: 'replacement/train.csv'，为什么会找不到replacement文档呢？谢谢！

Hi, I'm deeply inpressed by your work about the sentiment classification competition, and I'm trying to learn it. Howerer, I still have some troubles understanding your model.

You guys concat the pooled_output vector of bert with vectors from the first token's(CLS) representation of last three(or two) layers of bert. But isn't the pooled_output vector itself is the first token's vectorof the last layer of bert?
If so, what's point of concating identical vectors? If not the case, then what dose pooled_out vector mean? Is it the average of the CLS representations across all layers of bert?

Hoping for reply.

你好，最近在做文本分类，参照你的方法我把后三层的cls的输出和pooler output拼接成了一个长向量，然后做softmax，但是在训练的时候有时候会出现”Found Inf or NaN global norm. : Tensor had NaN values“的报错，请问你特征拼接的时候有遇到这个问题吗？我想是不是特征的尺度不一样导致的

你好 伪标签对应的那张图是你们自己画的还是引用的？如果是有原始论文吗？