Closed xiongfy closed 9 months ago
Hi! As mentioned in the paper, we didn't use specific action-level and object-level words during pre-training since we found that they could result in leakage for masked language modeling (MLM). They are added in the fine-tuning stage.
好的好的,感谢回复
您好,想请问一下在预训练reverie时有使用到词性划分的特征吗,我在pretrain_src的代码中好像没有找到和PickSpecificWords()相关的部分