Breaking words into smaller units (subwords) during tokenization, which can complicate analysis of word-level representations.