知识卡片
字节与字符是不同的基础单位
字节级BPE从有限字节集合构造片段,字符级算法从选定字符表构造片段。一个Unicode字符可能跨多个字节,组合后词元也可能不是完整字符;给定可逆编码与覆盖规则时才讨论还原,不能混称所有字符二元组都是字节对。
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
- · 思维的词元
原书出处 · 1 处
- · 查看原书依据
You may think that Unicode characters are the smallest packet of meaning in natural language text. This is true for a human, maybe, but for a machine, no way. Just as the BPE name suggests, characters don’t have to be your fundamental atom of meaning for your base vocabulary. You can split characters into 8-bit bytes. GPT-2 uses a byte-level BPE tokenizer to naturally compose
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:摘录中关于字节与字符是不同的基础单位的定义、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 may think that Unicode characters are the smallest packet of meaning in natural language text. This is true for a human, maybe, but for a machine, no way. Just as the BPE name suggests, characters don’t have to be your fundamental atom of meaning for your base vocabulary. You can split characters into 8-bit bytes. GPT-2 uses a byte-level BPE tokenizer to naturally compose