Tag
This paper investigates how character-level transformer models generalize to irregular verb subtypes in Japanese past-tense inflection. Controlled experiments show that including irregular examples can improve generalization, challenging the assumption that regularity simplifies learning.