The labyrinths of meaning

A recent study found that regardless how thoroughly AI-powered chatbots are trained with real, sensible text, they still have a hard time recognizing passages that are nonsense.

Given pairs of sentences, one of which makes semantic sense and the other of which clearly doesn't -- in the latter category, "Someone versed in circumference of high school I rambled" was one example -- a significant fraction of large language models struggled with telling the difference.

In case you needed another reason to be suspicious of what AI chatbots say to you.

As a linguist, though, I can confirm how hard it is to detect and analyze semantic or syntactic weirdness.  Noam Chomsky's famous example "Colorless green ideas sleep furiously" is syntactically well-formed, but has multiple problems with semantics -- something can't be both colorless and green, ideas don't sleep, you can't "sleep furiously," and so on.  How about the sentence, "My brother opened the window the maid the janitor Uncle Bill had hired had married had closed"?  This one is both syntactically well-formed and semantically meaningful, but there's definitely something... off about it.

The problem here is called "center embedding," which is when there are nested clauses, and the result is not so much wrong as it is confusing and difficult to parse.  It's the kind of thing I look for when I'm editing someone's manuscript -- one of those, "Well, I knew what I meant at the time" kind of moments.  (That this one actually does make sense can be demonstrated by breaking it up into two sentences -- "My brother opened the window the maid had closed.  She was the one who had married the janitor Uncle Bill had hired.")

Then there are "garden-path sentences" -- named for the expression "to lead (someone) down the garden path," to trick them or mislead them -- when you think you know where the sentence is going, then it takes a hard left turn, often based on a semantic ambiguity in one or more words.  Usually the shift leaves you with something that does make sense, but only if you re-evaluate where you thought the sentence was headed to start with.  There's the famous example, "Time flies like an arrow; fruit flies like a banana."  But I like even better "The old man the boat," because it only has five words, and still makes you pull up sharp.

The water gets even deeper than that, though.  Consider the strange sentence, "More people have been to Berlin than I have."

This sort of thing is called a comparative illusion, but I like the nickname "Escher sentences" better because it captures the sense of the problem.  You've seen the famous work by M. C. Escher, "Ascending and Descending," yes?

The issue both with Escher's staircase and the statement about Berlin is if you look at smaller pieces of it, everything looks fine; the problem only comes about when you put the whole thing together.  And like Escher's trudging monks, it's hard to pinpoint exactly where the problem occurs.

I remember a student of mine indignantly telling a classmate, "I'm way smarter than you're not."  And it's easy to laugh, but even the ordinarily brilliant and articulate Dan Rather slipped into this trap when he tweeted in 2020, "I think there are more candidates on stage who speak Spanish more fluently than our president speaks English."

It seems to make sense, and then suddenly you go, "... wait, what?"

An additional problem is that words frequently have multiple meanings and nuances -- which is the basis of wordplay, but would be really difficult to program into a large language model.  Take, for example, the anecdote about the redoubtable Dorothy Parker, who was cornered at a party by an insufferable bore.  "To sum up," the man said archly at the end of a long diatribe, "I simply can't bear fools."

"Odd," Parker shot back.  "Your mother obviously could."

A great many of Parker's best quips rely on a combination of semantic ambiguity and idiom.  Her review of a stage actress that "she runs the gamut of emotions from A to B" is one example, but to me, the best is her stinging jab at a writer -- "His work is both good and original.  But the parts that are good are not original, and the parts that are original are not good."

Then there's the riposte from John Wilkes, a famously witty British Member of Parliament in the last half of the eighteenth century.  Another MP, John Montagu, 4th Earl of Sandwich, was infuriated by something Wilkes had said, and sputtered out, "I predict you will die either on the gallows or else of some loathsome disease!"  And Wilkes calmly responded, "Which it will be, my dear sir, depends entirely on whether I embrace your principles or your mistress."

All of this adds up to the fact that languages contain labyrinths of meaning and structure, and we have a long way to go before AI will master them.  (Given my opinion about the current use of AI -- which I've made abundantly clear in previous posts -- I'm inclined to think this is a good thing.)  It's hard enough for human native speakers to use and understand language well; capturing that capacity in software is, I think, going to be a long time coming.

It'll be interesting to see at what point a large language model can parse correctly something like "Buffalo buffalo Buffalo buffalo buffalo buffalo Buffalo buffalo."  Which is both syntactically well-formed and semantically meaningful.  

Have fun piecing together what exactly it does mean.

****************************************