#subword-regularization

Question 1: Single choice

Kudo (2018) introduced the Unigram language-model tokenizer to support subword regularization. What does a Unigram tokenizer provide for this that BPE does not?

Probabilities over all segmentations of a sentence, so several can be sampled1
A smaller vocabulary for the same corpus, so the embedding layer has fewer parameters2
Tokens that may cross word boundaries, so multi-word phrases become single tokens3
A base vocabulary of 256 bytes, so that no input ever produces an unknown token4

#subword-regularization

Timed test