← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 6656.c · 生成模型:GPT-6

字节对编码按相邻符号迭代合并

理解等级:中级

BPE从基础符号序列开始,统计相邻符号对,选择高频对合并为新符号,再更新后续统计,直到达到词表或合并停止条件。合并后才会形成更长片段;它不是一次统计所有固定字符二元组,也不等于找到最长字符串的全局最优算法。

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

      You can think of the BPE tokenizer algorithm as a matchmaker in a social network of friends. BPE pairs up characters that commonly appear next to each other a lot, and then it creates a new token for these character combinations. BPE can then pair up the multicharacter tokens whenever those token pairings are common in your text. It keeps doing this until it has

    参考来源

    - 位置:《Natural Language Processing in Action, Second Edition》第02章《思维的词元》(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-2.html) - 结论依据:摘录中关于字节对编码按相邻符号迭代合并的定义、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:You can think of the BPE tokenizer algorithm as a matchmaker in a social network of friends. BPE pairs up characters that commonly appear next to each other a lot, and then it creates a new token for these character combinations. BPE can then pair up the multicharacter tokens whenever those token pairings are common in your text. It keeps doing this until it has