Recent research by five computational linguists has tested Noam Chomsky's assertion regarding language models and their ability to learn 'impossible' languages. Chomsky had argued that these models could easily master languages governed by rules distinct from any known human language.
The study, titled 'Mission: Impossible Language Models,' received the best paper award at the 2024 conference of the Association for Computational Linguistics. The authors found that language models struggled more with these impossible languages compared to standard English.
Adele Goldberg, a linguist from Princeton University, praised the paper as 'absolutely timely and important,' suggesting that such models could aid researchers in understanding infant babbling.
Chomsky's theories, which gained prominence in the mid-20th century, proposed that humans possess an innate mental mechanism for language processing, explaining the absence of certain grammatical rules across known languages. He argued that if language learning were akin to other learning types, it would not favor specific grammatical rules over others.
In contrast, the recent experiments by Mitchell and Bowers in 2020 showed that language models could learn impossible languages with surprising accuracy. However, Kallini, a graduate student at Stanford University, sought to further investigate Chomsky's claims using modern transformer models, which are prevalent in current language processing.
Kallini's team created a dozen impossible languages by manipulating English texts, observing that models trained on these languages improved over time but still faced challenges compared to those trained on standard English.
This research indicates that language models exhibit preferences in learning certain linguistic patterns, paralleling human tendencies, but not necessarily mirroring them. The findings suggest a complex interplay between neural networks and human language acquisition.
As researchers continue to explore these dynamics, the study opens avenues for further investigation into the nature of language learning and the capabilities of artificial intelligence in processing complex linguistic structures.