Using glossaries, indexes and other alphabetically ordered word listings to leverage the explicitly learned spellings in order to deduce beginnings of other words – e.g. if you knew how to spell the token ‘the’, and you kept seeing the token ‘this’ listed shortly after the token ‘the’ in alphabetic listings, you could reasonably guess that ‘this’ begins with a T, its second letter could well be H, and if so, its third letter comes from the set {E, F, …, Z}. By spending an astronomical amount of time attempting to solve something akin to a 50,000-dimensional Sudoku puzzle, you might be able to achieve high confidence for your guesses as to the first three or four letters of most whole-word token strings.
I would hazard a guess this is the most significant mechanism here. Crossword or Scrabble dictionaries, lists of “animals/foods/whatever starting with [letter]”, “words without E”, “words ending in [letter]”, anagrams, rhyming dictionaries, hyphenation dictionaries, palindromes, lists where both a word and its reverse are valid words, word snakes and other language games… It probably ingested a lot of those. And when it learns something about “mayonnaise” from those lists, that knowledge is useful for any words containing the tokens ‘may’, ‘onna’ and ‘ise’.
I would hazard a guess this is the most significant mechanism here. Crossword or Scrabble dictionaries, lists of “animals/foods/whatever starting with [letter]”, “words without E”, “words ending in [letter]”, anagrams, rhyming dictionaries, hyphenation dictionaries, palindromes, lists where both a word and its reverse are valid words, word snakes and other language games… It probably ingested a lot of those. And when it learns something about “mayonnaise” from those lists, that knowledge is useful for any words containing the tokens ‘may’, ‘onna’ and ‘ise’.