How does ULMFiT do the intial word embeddings?

Part 1 (2019)
1

In the beginning of training, how does the model take a string and convert it to a vector?
Is there a particular way the word embedding is done?