知识卡片
掩码位置选择与替换方式分两层
原BERT先选15%词位计入MLM目标,再在这些位置中80%换MASK、10%换随机词、10%保留原词。原词保留也仍受监督,不能说只有实际MASK位置训练,更不能与skip-gram词对预测等同。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
原书出处 · 1 处
- · 查看原书依据
During pretraining, a percentage of input tokens are masked randomly (with a [MASK] token), and the model predicts the actual token IDs for those masked tokens. In practice, 15% of the WordPiece tokens were selected to be masked for training; however, a downside of this is that during fine-tuning, there is no [MASK] token. To work around this, the authors came up with a formula
参考来源
- 位置:《Natural Language Processing in Action, Second Edition》第09章《可堆叠的Transformer》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Natural Language Processing in Action, Second Edition/_epub-src/OEBPS/Text/chapter-9.html)
- 结论依据:摘录中关于掩码位置选择与替换方式分两层的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。
- 原始内容:During pretraining, a percentage of input tokens are masked randomly (with a [MASK] token), and the model predicts the actual token IDs for those masked tokens. In practice, 15% of the WordPiece tokens were selected to be masked for training; however, a downside of this is that during fine-tuning, there is no [MASK] token. To work around this, the authors came up with a formula