← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 6740 · 生成模型:GPT-6

未知词标记不等于随机掩码任务

理解等级:中级

WikiText非raw版本将词表外项替换为UNK,raw版本保留原词项。此处不是为补全任务随机遮住2.7%真词,普通下一词模型也没有被给到隐藏原词标签;更换raw会改变词表与评价协议,不能直接称准确率必提升。

这张卡片看不懂?

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

      Unfortunately, the PyTorch version of the WikiText2 includes <unk> tokens that randomly replace, or mask, 2.7% of the tokens. That means your model will never get very high accuracy, unless there is some predictable pattern that determines which tokens were masked with <unk>. But if you download the original raw text without the masking tokens, you can train your language model on it and get

    参考来源

    - 位置:《Natural Language Processing in Action, Second Edition》第08章《循环网络与长短期记忆》(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-8.html) - 结论依据:摘录中关于未知词标记不等于随机掩码任务的定义、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:Unfortunately, the PyTorch version of the WikiText2 includes <unk> tokens that randomly replace, or mask, 2.7% of the tokens. That means your model will never get very high accuracy, unless there is some predictable pattern that determines which tokens were masked with <unk>. But if you download the original raw text without the masking tokens, you can train your language model on it and get