Zwaarddijk wrote:Basically, what you get is a kind of randomized algorithm that introduces - as already noted - a small extra level of redundancy. The extra level of redundancy given by a gender system, although not 100% helpful, is apparently helpful enough to compensate for the extra learning it requires. (Extra redundancy isn't of course the only thing it provides - it also provides some of the benefits of having two distinct third persons (generally called proximate and obviate) without requiring all the complications that such a system brings along. This extra distinct third-person system won't work all the time - you'll run into contexts where all the things you need to distinguish are of the same gender, but it's a sort of randomized solution, and it works often enough to be worth maintaining, for less of a ~cost than developing a proximate-obviate system of native north American style would entail for most languages that currently do have gender systems)

Randomized solutions that only help some of the time really are a surprisingly efficient and cheap method of going about this kind of thing!

What makes the benefits even more obvious is the fact that we don't actually hear every sound the speaker we listen to makes (nor do we read every letter in every word on the page or screen in front of us!) - and having extra clues here and there makes it way more likely that the utterance is reconstituted into the right string of morphemes in our mind.

Zwaarddijk wrote:Think of the set of words in a language. If this language has noun classes or genders, some subset of the words will belong to one class, some will belong to another class ...

Now, as the language develops, there's no knowledge of what utterances will be made. Words, by a variety of approaches, fall into some class (some, in some languages, fall into several; etc etc etc - there's myriads of ways this can happen).
The choice of what class a word belongs to may diachronically be clear (although words do change gender on occasion through a language's history) but synchronically it may approximate randomness pretty well.

Zwaarddijk wrote:An actual statement will have words we can't entirely predict, and these will belong to classes we wouldn't necessarily have been able to predict without the specific knowledge that a given noun belongs to such-and-such a gender.


Does this make sense to you?

Zwaarddijk wrote:I am not trying to be very accurate here. I am mostly trying to sort of compare this to an analogous mechanism.

But yeah, many languages don't have any obvious rule for how a word has been assigned gender historically - in Swedish, German, Russian, Latin, etc there's a few pairs (or triplets) of words that are synonymous but assigned to different genders. E.g. en tallrik (common gender)/ ett fat (neuter) (both meaning 'a plate', so the assignment is not semantically conditioned, it's not phonologically conditioned - ett hus, en mus - there's even homophones that differ in gender, etc. Taking this into account, the distribution is close enough to random-like for this purpose)