Hi, I'm curious about the MLM task's masking strategy in your paper (section 2.2.2)
"When a grapheme is selected, its phonemes tokens are replaced with a MASK token 80% of the time, are replaced with random phonemes token 10% of the time, and stay unchanged 10% of the time."
when you replace with random phonemes, will it be related to an real word? For example, will insane phoneme sequence such as 'abcde' be possible?
Yes, it is completely random phonemes, not a word, because eventually we want to predict the correct phonemes from random phonemes, and the number of phonemes for a word must match that of the target word.
Hi, I'm curious about the MLM task's masking strategy in your paper (section 2.2.2)
"When a grapheme is selected, its phonemes tokens are replaced with a MASK token 80% of the time, are replaced with random phonemes token 10% of the time, and stay unchanged 10% of the time."
when you replace with random phonemes, will it be related to an real word? For example, will insane phoneme sequence such as 'abcde' be possible?