I don't think that's the difference. I split on separator characters when forming the array @words but I negate the character class when doing the s{ ... }{ ...}xeg to add the annotation. Thus, like you, I am pulling out words but by capturing one or more non-separator characters.

Cheers,

JohnGG

Update: I substituted your pattern

      ([A-Za-z']+)(?{++ $found{lc $1}})
[download]

for my pattern

      ([^.,;:?! \n]+)(?{++ $found{lc $1}})
[download]

and the results were identical.