Changing the mlm_probability argument wont give you the result you need,
but I think you can create a sub class of DataCollatorForLanguageModeling that does the emoji masking.
You can find the source code for DataCollatorForLanguageModeling here.