In the late 1950s Noam Chomsky proposed that we have a natural capacity to process grammar and thus to use language – in essence that our brain is hard-wired to use language.
It was, and is, a controversial theory (though Chomsky would not agree), but this week new evidence has been published to support it – and, as outlined in the New Scientist, you can even conduct a thought experiment on yourself to test it.
Writing in the Proceedings of the National Academy of Sciences (the US National Academy that is), Jennifer Culbertson and David Adger consider whether language learners pick up language patterns by observing statistical patterns from existing speakers/users or – as the Chomskian theory would suggest – apply some form of “hard wired” rule to process grammar.
To do this they presented subjects (English speakers) with a “new” limited language based on common words in English. The subjects were then asked to judge whether a new phrase in this “language” – made by combining elements of the limited language they had already seen – would be correct in one of two forms. If they picked one form then they would likely be using some form of statistical inference – picking a form that looked closest to the forms they had already seen – if they picked another they were likely using an internal grammar machine in their brains.
And this is where you can test yourself … (shamelsssly nicked from the New Scientist as this example does not appear to be in the article itself):
Here are two phrases in the new language:
- shoes blue
- shoes two
So which of the following phrases is correct in this language:
- shoes two blue
- shoes blue two
If, as I did, you picked “shoes blue two” and not “shoes two blue” then you are favouring a semantic hierarchy and not a frequency based approach – in English two usually precedes blue, but blue is a stronger modifier of the noun than two.
In fact people chose the semantic hierarchy about 75% of the time – strongly suggesting that we do have a internal grammar engine running inside our heads.
(Chomsky himself appears to be dismissive of the study, despite it appearing confirm his work – “like adding a toothpick to mountain”. Tells you quite a lot about him, I think.)
What are the practical implications? I think it points to a limit to the effectiveness of things like big data based machine translation, if all that relies on is statistical inference. Inside a decade big data has made machine translation much more practical than the previous 50 years of AI research, but the quest for a way to compute grammar is still going to matter.