知识卡片

普通读书笔记卡 · 6622.b · 生成模型:GPT-6

词元边界不等于概念边界

理解等级:中级

子词可拆开词根、词缀、数字或领域术语,一个完整概念未必只有一个ID。没有专用token也不意味着模型完全不懂该概念;需要结合分词长度、上下文与任务表现判断。

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

      Looking at the vocabulary file in detail, we notice that the first few hundred tokens are all single-character tokens, such as special characters, digits, capital letters, small letters, and accented characters. Longer words appear later on in the vocabulary. A lot of tokens correspond to just a part of a word, called a subword, like “impl,” “inated,” and so on.

    参考来源

    - 位置:《Designing Large Language Model Applications》第03章《词表与分词》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch03.html) - 结论依据:摘录中关于词元边界不等于概念边界的定义、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:Looking at the vocabulary file in detail, we notice that the first few hundred tokens are all single-character tokens, such as special characters, digits, capital letters, small letters, and accented characters. Longer words appear later on in the vocabulary. A lot of tokens correspond to just a part of a word, called a subword, like “impl,” “inated,” and so on.