知识卡片
子词方案不是架构必备条件
分词粒度影响词表、序列长和覆盖,BPE是可选方案之一,BERT又用WordPiece。Transformer的注意力结构并不强制采用某一种分词,也没有固定5000项的通用词表;数据与模型配置需要共同保存。
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
Three innovations came together to give transformers their power to imitate human conversation: Byte pair encoding—Tokenizing words based on character sequence statistics rather than spaces and punctuation Attention—Connecting important word patterns together across long stretches of text, using a connection matrix (attention) Positional encoding—Keeping track of where each token or pattern is located within the token sequence Byte pair encoding (BPE) is an often-overlooked enhancement
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:摘录中关于子词方案不是架构必备条件的定义、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:Three innovations came together to give transformers their power to imitate human conversation: Byte pair encoding—Tokenizing words based on character sequence statistics rather than spaces and punctuation Attention—Connecting important word patterns together across long stretches of text, using a connection matrix (attention) Positional encoding—Keeping track of where each token or pattern is located within the token sequence Byte pair encoding (BPE) is an often-overlooked enhancement