Comment by ComposedPattern

14 days ago

I'm not that convinced by this paper. The "impossible languages" are all English with some sort of transformation applied, such as shuffling the word order. It seems like learning such languages would require first learning English and then learning the transformation. It's not surprising that systems would be worse at learning such languages than just learning English on its own. But I don't think these sorts of languages are what Chomsky is talking about. When Chomsky says "impossible languages," he means languages that have a coherent and learnable structure but which aren't compatible with what he thinks are innate grammatical facilities of the human mind. So for instance, x86 assembly language is reasonably structured and can express anything that C++ can, but unlike C++, it doesn't have a recursive tree-based syntax. Chomsky believes that any natural language you find will be structured more like C++ than like assembly language, because he thinks humans have an innate mental facility for using tree-based languages. I actually think a better test of whether LLMs learn languages like humans would be to see if they learn assembly as well as C++. That would be incomplete of course, but it would be getting at what Chomsky's talking about.

Also, GPT-2 actually seems to do quite well on some of the tested languages, including word-hop, partial reverse, and local-shuffle. It doesn't do quite as well as plain English, but GPT-2 was designed to learn English, so it's not surprising that it would do a little better. For instance, they tokenization seems biased towards English. They show "bookshelf" becoming the tokens "book", "sh", and "lf" – which in many of the languages get spread throughout a sentence. I don't think a system designed to learn shuffled-English would tokenize this way!