知识卡片
大小写归一是信息取舍
保留大小写可区分专名和普通词,但增加需要学习的符号差异;转小写能合并某些变体,也可能抹去语义线索。应与模型训练配置一致,不把“有大小写词表总更好”当所有语料下成立的规律。
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
- · 词表与分词
原书出处 · 1 处
- · 查看原书依据
You can see that the tokens are case-sensitive: there are separate tokens for “office” and “Office.” Most models these days have case-sensitive vocabularies. Back in the day, the BERT model was released with both a cased and an uncased version.
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:摘录中关于大小写归一是信息取舍的定义、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 see that the tokens are case-sensitive: there are separate tokens for “office” and “Office.” Most models these days have case-sensitive vocabularies. Back in the day, the BERT model was released with both a cased and an uncased version.