You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Bert的预训练主要包含两个任务,MLM和NSP,Masked Language Model任务可以理解为完形填空,随机mask每一个句子中15%的词,用其上下文来做预测;Next Sentence Prediction任务选择一些句子对A与B,其中50%的数据B是A的下一条句子,剩余50%的数据B是语料库中随机选择的,学习其中的相关性。BERT 预训练阶段实际上是将上述两个任务结合起来,同时进行,然后将所有的 Loss 相加