Unlike the long, beautifully written dialogue in most books and movies, the language of our everyday ieraction is often messy and incomplete, full of mis-seences, false starts, and inappropriate pauses. This includes any type of conversation, from casual chats between friends, argumes between siblings, to formal discussions in the boardroom. Therefore, given the random and experieial nature of language, it seemed miraculous that anyone could learn language by experience alone. For this reason, many scieists in the field of linguistics, including Noam Chomsky, the founder of modern linguistics, believe that language learners need some kind of glue to corol the unruly nature of everyday language, and that is nothing but grammar (a system of rules for production of grammatical seences) is not.
From this poi of view, children should also have a grammatical pattern in their brain that helps them to overcome the limitations of their language experience. For example, the pattern may coain a specific super-rule that dictates how new sections are added to existing expressions; After that, the child’s brain checks whether his native language matches what he made.
In this way, according to the same basic pattern, the English-speaking child knows that the verb is placed before the object (such as: I eat sushi); While the same child, if he has Japanese nationality, he knows according to his super law that the verb comes after the object (in Japanese, the structure of the same seence is: I sushi eat).
But the new insight io how to learn a language comes from a special source that you probably haven’t even thought of, and that’s none other than artificial ielligence! In fact, a new generation of AI models can write newspaper articles, poetry, and computer code after being exposed to large amous of linguistic data as input. The most amazing part of the story is that all these tasks are done for artificial ielligence without the help of grammar.
Expressing seences correctly without using grammar
Even if AI models’ word choices are sometimes weird, nonsensical, or coain racist, sexist, and other biases, one thing is crystal clear about their phrasing; And that is that the vast majority of the output of these AI language models is grammatically correct! While they are not given any grammar patterns or rules, and models rely only on linguistic experience to provide correct expressions as output.
The model is GPT-3, which is a deep learning neural network with 175 billion parameters, and it has a little brother called GPT-2!
Among these, one of the most well-known artificial ielligence models is GPT-3, which is a very gigaic deep learning neural network with 175 billion parameters. During the training process, the AI was given hundreds of billions of words from the Iernet, books, and Wikipedia as input, and the AI was asked to use what it learned to predict the next word in a seence. In this regard, when artificial ielligence made a wrong prediction, its parameters were adjusted using an automatic learning algorithm so that it could predict the next word with less error!
Ierestingly, GPT-3 can produce believable text in response to seences such as “A synopsis of the latest Fast and Furious movie…” or “An Emily Dickinson-esque poem.” In addition, the GPT-3 can answer SAT-level comparisons, reading comprehension questions, and even simple math problems, all learned using next-word predictive learning!
An AI model and a human brain may produce the same output, but do they do it in the same way?
Comparison of artificial ielligence models with human brain
However, the similarities of the output expressions of artificial ielligence models with human language do not stop here, and research published in the famous journal “Nature Neuroscience” shows that these deep learning artificial networks use the same computational principles as the human brain. The researchers in this field, led by neuroscieist Uri Hasson, first compared the predictions of the next words of a story from the podcast “This American Life” between humans and the artificial ielligence model GPT-2 (the little brother of the GPT-2 model). 3) They paid. During these investigations, it was found that the human brain and artificial ielligence predicted exactly the same word in the story almost 50% of the time.
The researchers recorded the volueers’ brain activity as they listened to the story, and the best explanation for the patterns of activation observed in the volueers’ brains was that the brain, like GPT-2, does not use only one or two previous words when predicting. Rather, it relies on the accumulated coext of the previous 100 words. Therefore, overall, the authors concluded that their findings of predictive and spoaneous neural signals while participas listened to natural test speech suggest that active prediction may underlie lifelong language learning in humans. Provide.
Of course, you should know that a possible concern in this experime is that the AI models use big data (the GPT-3 AI model was trained with 20,000 years of language experience!) as input. But another preliminary study, which has yet to be fully peer-reviewed, shows that GPT-2 can still model next-word predictions and brain activation even when trained with only 100 million words. It should be noted that this amou of language input is equivale to the average words of children during the first 10 years of life!
It should be emphasized that I am not claiming in this article that the GPT-3 or GPT-2 AI models learn language exactly the same way as children; Because one of the importa aspects of considering such similarity is understanding, and artificial ielligence does not understand much of what it says.
It should be emphasized that I am not claiming in this article that the GPT-3 or GPT-2 AI models learn language exactly the same way as children; Because one of the importa aspects of considering such a similarity is understanding, and artificial ielligence does not understand much of what it says, while understanding is one of the basic elemes in human language. What these models do prove, however, is that a learner (albeit made of silicon!) can produce language sufficiely through exposure to perfectly correct grammatical seences, and do so in a manner similar to Human brain processing.
Redefining the language learning debate
For years, many linguists believed that language learning was impossible without an iernal grammatical pattern. But as we meioned throughout this article, the new artificial ielligence proved the opposite. In fact, artificial ielligence models show that the ability to produce grammatically correct seences can be learned only through language experience. Therefore, it can be claimed that probably children do not need to have an innate grammar and match it with what they hear to learn a language! Instead, children should be involved in as many differe conversations as possible in order to develop their language skills. Because language experience and not grammar is the key to becoming a person with correct and expressive expression.
main source
Source: arsTECHNICA
