To be clear, the alternative to having a pre-determined structure for language is that any sound or gesture you see/hear coming from ypur parents as a small child could potentially cary semantic meaning. Also, any characteristic of that spund or gesture (pitch, length, order, frequency, loudness, correlation with other events, time of day etc) could encode that meaning. Even once you somehow identify words as being the relevant part, you still have to identify how words relate to each other - the way order of wrods matters, the way order of sentences matters etc.

All of this ends up with a huge combinatory explosion of possible ways to assign meaning to language, requiring a huge amount of examples and counter examples to actually learn a human language. Just look at the training corpuses of something like GPT-X to see the amount of information that is required to learn elements of a human langauge if you start without any assumptions.

However, we know for a fact that human children get many orders of magnitude less examples of languahe use before learning a language - far too little to actually rule out all other possible structures if they had started with 0 assumptions.

Adding to this, there are certain elements of language structure that are not commonly thought about that happen to work identically in all natural languages that have been studied, sucha s the fact that order of words in a sentence matters, or the way certain words are omitted when constructing embedded sentences (e.g. 'can eagles that fly swim?' always having 'can' refer to 'swim', never 'fly').