This has led to work showing that models can do better sometimes if you decompose these into their constituent characters, e.g.: https://aclanthology.org/2022.emnlp-main.472.pdf