[ericjang]

Fair enough, I agree that if we really examine the comment "word as a discrete unit of meaning", the edge cases start to accumulate and the semantics rapidly break down. But barring things like prefixes/suffixes/modifiers/composite word characters in traditional Chinese, words are fairly discrete and generally regarded as the primary layer for expressing singular units of "meaning"

[solarmist]

They are, but only because we don't have better language to express them. Similar to a lot of the problems with Chomsky's works the composability of language is only a subset of the whole breadth of what is expressable in a given language.

Or in other words, I believe the surface area of "edge cases" has a similar surface area as the rest of the language. The difference being they aren't invoked nearly as often because they require more effort and creativity.

Just look at the rise of words like "hangry". There are types of mashups that show up in creative uses of language that defy nearly any rule for any language you can come up with. In many languages, if you choose any of those supposed rules you can probably construct an algorithm to generate odd, but understandable words that defy that rule.

[Mezzie]

> Or in other words, I believe the surface area of "edge cases" has a similar surface area as the rest of the language. The difference being they aren't invoked nearly as often because they require more effort and creativity.

Edge cases or exceptions do tend towards being highly used; this is because language is more likely to change the more it's used, so the most highly used words/phrases/sentences/etc tend to accumulate changes. One example of this is that if a language has verb conjugation and irregular verbs, then odds are some of its most common verbs will be irregular.

> Just look at the rise of words like "hangry". There are types of mashups that show up in creative uses of language that defy nearly any rule for any language you can come up with. In many languages, if you choose any of those supposed rules you can probably construct an algorithm to generate odd, but understandable words that defy that rule.

There are rules for that that would work, weirdly enough. There are just a ton of them.

[solarmist]

My only point here is that any framework for generalization needs to be able to account for and incorporate these kinds of "exception-seeking" cases. Similar to the same way that mathematics uses counter-examples to strengthen and reinforce the definitions chosen.

[ericjang]

I agree with your comment "In many languages, if you choose any of those supposed rules you can probably construct an algorithm to generate odd, but understandable words that defy that rule." - it comes many forms, from Goodhart's Law to the "hot dog vs. sandwich" debate.

I do mention this in my blog post - although I think Generalization is Language, I don't think it's possible to create a formal framework of language, for precisely because of "adversarial examples" that can be supplied for any formal definition.

Natural language itself, ignorant of formality, is able to account for these exceptions insofar as language is sufficient for people to convey a bare minimum of meaning. I am proposing to define language and generalization via the implicit understanding of large language models, in the same way you might use an image classifier to define "cat images" or "hot dogs"

[solarmist]

Hmm, I can understand the motivation. However, I feel it either won't work or will be very fragile because it's already part of the model because they're trained using natural language.

DL is already far from formal models, that's why deep learning “works.” And even at the current level of DL models, those exceptions are represented to some extent.

So ultimately, your idea is to push the models toward further generality, which in my option, will bake these “exceptions” deeper into the model.

And my question is, what does that mean for your idea? In my mind trying to exclude them would break what works. On the other hand, ignoring them means you can't direct development towards your goal because there’s no map from language to generalizations, so that you would be relying on random chance for progress.

If this is off in left field, let me know, but that's what I can see from your description.