Text this: An Analysis of the Training Data Impact for Domain-Adapted Tokenizer Performances—The Case of Serbian Legal Domain Adaptation