How Good is Your Tokenizer? On the Monolingual Performance of Multilingual Language Models

8625

Knowledge Triples