≡ Menu

Wittgenstein


Here’s a story right up my blogging alley. I’ve written quite a bit in the past on translation (about Horace and ESL/film), as well as bit on technology and language. I wrote about how Google used the insights of Wittgenstein to overcome the problem of polysemy in search, but ended questioning whether Google could ever overcome the complexities of poetry. Turns out Google has been laboring away at creating a machine translator of poetry.

If I understand it correctly, the poetry translator basically layers several poetic constraints on top of the standard translator: line length, rhyme, meter, etc. Google’s translator uses what Jaron Lanier calls a “brute force” approach to translation. That is, it doesn’t know the rules of grammar—it doesn’t even really have a dictionary. Rather, it scours its database and determines statistical correlation between translations of pages. Put another way, it imitates by means of statistical analysis.

Meta-lord of the cloud-lords of meta of!

Questions of quality aside (i.e., let’s assume Google can be completely successful and create passable—even good poetry translations), would you really prefer Google’s translations Rimbaud over, say, Ashbery’s? Aside from needing a translation in a pinch, I can only imagine an interest in Google’s translation that is analogous to the Turing test: an interest that asks the question “If I didn’t know—could I tell the difference between the results of computer and human translation?”

I have been reading Jaron Lanier’s book You Are Not a Gadget over the last few weeks. He makes a convincing point that Turing’s test is essentially the wrong question. Part of the function of asking “can it fool us?” is a desire to find a computer that can. As a result, we’re essentially willing to dumb down our expectations of what it means to be human in hopes we’ve created machines that think. Ironically, it’s our very human desires that make the Turing test fail. The real judge of the Turing test should be a computer with a merciless set of criteria. No doubt somebody, somewhere has already realized this, and there is a computer slaving away at creating and judging its own intelligence.

Which brings me back to the question: why do we want to read Ashbery’s translations of Rimbaud? I see two motivations: the first is to read Rimbaud without learning French; the second is to read Ashbery reading Rimbaud. Google doesn’t read. To say that it does would actually change the definition of reading, wouldn’t it? Reading implies not a functional end (e.g., Ashbery produces a translation of Rimbaud), since it can exist without a functional end (e.g., Ashbery reads Rimbaud in French).

Perhaps more importantly, Google doesn’t even use language in a way that we recognize as language. Some animals use what we would rightly be called protolanguage. They can acquire a vocabulary, and perhaps even use it in creative ways (I heard a story once about an ape that put two words together to ask for a watermelon: “candy water” or something along those lines). At best, though, animals can only mash together vocabulary, without what we could refer to as “syntax.” Syntax is the ability not only to acquire vocabulary, but to manipulate it according to a deeper intelligence that categorizes vocabulary. It’s the difference between “Micah smile” and “Micah smiles.” The latter indicates not only the fact that I have associated one thing with another (the action of smiling with the word “smile”), but that I can categorize it as a verb and thus deploy it in a sentence (oh the difference an “s” makes). This syntactic ability expands when we think about relative clauses, which nest and hierarchize ideas. We even have words for pure functions of language (e.g., articles). Animals are unable to do this (unless, of course, you’re teaching a gorilla that it will die someday—perhaps death is the motivator of syntax!). Google uses statistical analysis to achieve a kind of protolanguage at best. At best, it “learns” (a word also worth an essay) to associate certain phrases with one another. But, unlike animals, it has no will to use them.

All this is to say that there is something uniquely motivating about a person doing something. A Google poetry translation will never make me reconsider my life, except in a purely serendipitous (i.e., accidental) way.

I suppose deep down I am a personalist, believing there is something utterly unique and irreducible about persons. And I worry sometimes that the whole preoccupation with AI actually takes away from the real achievements of Google’s poetry translator: we clever people have found a way to essentially use an on-off switch (0s and 1s) to do something as complex as creating a passable translation of a poem. But as we are humans wont to do, we get distracted, venerating our creation rather than marveling at the deep mystery inside us which motivated us to create it in the first place.

Here’s a quick overview of the project if you’re interested in reading more about it. Here’s an interview about the project from CBC radio (scroll down to “Do Not Go Gentle Into That Good Digital Night”—below that one, there’s also another very interesting interview with a Canadian student who created a computer program to analyze rap lyrics).

Canada, it’s spring-break time. We’ve already got trees budding. Actually, many schools have gotten the whole Winter Olympics off for two or three weeks of extended spring break drunkenness. I’ve been glued to CTV for the last week or so, watching my new favorite sport: curling. No joke, this game is intense. It’s the skill of bowling with the strategy of chess. Not to mention, curling and hockey are two of the few things that get Canadians riled up (they go from passive agressive to just plain aggressive).

In the spirit of breaks from routine (are two weeks of blog posts long enough to be considered a routine?), I figured it would be good to take a break from the Grossman inspired posts and do a little reflection on a recent article in Wired magazine: “How Google’s Algorithm Rules the Web.” And in the spirit of Spring, I want to see if I can connect it to W.C. Williams.

Ever since I read an article on cloud computing and Google’s ability to translate web pages based upon its database alone (that is, nobody programmed the various language rules in it; it literally translates via algorithm), I’ve been interested in Google’s relationship with language. Now, any of you who have used Google Translation know it’s pretty awful, but the idea alone is impressive, and there’s no telling where improvements will take it.

This particular article goes under the hood of Google’s search engine, and we find out the real difficulties lies not so much in web crawlers, page ranking, or any of the stuff Google is known for (although, that is certainly a feat), but rather interpreting the desires of the Googler:

“We discovered a nifty thing very early on,” Singhal says. “People change words in their queries. So someone would say, ‘pictures of dogs,’ and then they’d say, ‘pictures of puppies.’ So that told us that maybe ‘dogs’ and ‘puppies’ were interchangeable. We also learned that when you boil water, it’s hot water. We were relearning semantics from humans, and that was a great advance.”

But there were obstacles. Google’s synonym system understood that a dog was similar to a puppy and that boiling water was hot. But it also concluded that a hot dog was the same as a boiling puppy. The problem was fixed in late 2002 by a breakthrough based on philosopher Ludwig Wittgenstein’s theories about how words are defined by context. As Google crawled and archived billions of documents and Web pages, it analyzed what words were close to each other. “Hot dog” would be found in searches that also contained “bread” and “mustard” and “baseball games” — not poached pooches. That helped the algorithm understand what “hot dog” — and millions of other terms — meant. “Today, if you type ‘Gandhi bio,’ we know that bio means biography,” Singhal says. “And if you type ‘bio warfare,’ it means biological.”

Did you catch that? Google uses Wittgenstein…Holy Mother of all snake-eating-its-own-Postmodern-tail!

Google, of course, is probably at the forefront of product innovation. Google has created an environment where failed ideas are OK, a sort of decentralized and messily creative workplaces that capitalizes on an excess of time and resources. Knowledge workers rejoice! (See here and here)

But that wasn’t what caught my attention most of all. It was this:

One unsuccessful search became a legend: Sometime in 2001, Singhal learned of poor results when people typed the name “audrey fino” into the search box. Google kept returning Italian sites praising Audrey Hepburn. (Fino means fine in Italian.) “We realized that this is actually a person’s name,” Singhal says. “But we didn’t have the smarts in the system.”

The Audrey Fino failure led Singhal on a multiyear quest to improve the way the system deals with names — which account for 8 percent of all searches. To crack it, he had to master the black art of “bi-gram breakage” — that is, separating multiple words into discrete units. For instance, “new york” represents two words that go together (a bi-gram). But so would the three words in “new york times,” which clearly indicate a different kind of search. And everything changes when the query is “new york times square.” Humans can make these distinctions instantly, but Google does not have a Brazil-like back room with hundreds of thousands of cubicle jockeys. It relies on algorithms.

The Mike Siwek query illustrates how Google accomplishes this. When Singhal types in a command to expose a layer of code underneath each search result, it’s clear which signals determine the selection of the top links: a bi-gram connection to figure it’s a name; a synonym; a geographic location. “Deconstruct this query from an engineer’s point of view,” Singhal explains. “We say, ‘Aha! We can break this here!’ We figure that lawyer is not a last name and Siwek is not a middle name. And by the way, lawyer is not a town in Michigan. A lawyer is an attorney.”

This is the hard-won realization from inside the Google search engine, culled from the data generated by billions of searches: a rock is a rock. It’s also a stone, and it could be a boulder. Spell it “rokc” and it’s still a rock. But put “little” in front of it and it’s the capital of Arkansas. Which is not an ark. Unless Noah is around. “The holy grail of search is to understand what the user wants,” Singhal says. “Then you are not matching words; you are actually trying to match meaning.”

My current job is teaching upper level writing to ESL students who are entering graduate school. Most of them will go on to do MBAs, but I try to give them a heavy dose of the liberal arts, which many of the students (especially ones from China) are lacking. It’s an incredibly frustrating process, at first, but it has turned into the best kind of reward, as I get a glimpse of my own language and system of thought from an outside (alienated?) perspective. For as many discernible, overarching truths and rules about the language, I often find the same number of beguiling nooks and crannies, particularities that indicate a long history of human choice and situation enshrined in our very words. Almost everyone knows this about language, but to actually encounter it on a regular basis is a bizarre experience.

I think my experience teaching is the same experience that Google’s engineers must deal with: Why do we associate certain things, and what complex process takes place in our brain that allows us to instantly recognize them? The question of lines in poetry adds a layer of complexity to this question. Grossman says that “lineation” is one of the defining qualities of poetry (even prose poetry is defined by its lack of lines, isn’t it?). As I was attempting to write poetry for the first time in high school, I remember obsessing over my lines. I could never understand why I wanted to break a line here one day and there on another day. As I write, I feel pretty confident about cutting my lines. That isn’t to say I still don’t play with line breaks, but I have a pretty intuitive sense of when to hit the Enter button. Sometimes I break a line for the sake of a playful slight rhythm, but usually it comes as a sense of whim–it just seems right. Is this an intuitive sense that has a core? Or is it really just whim?

This leads me to my question: How would Google parse a poem like W.C. William’s “The Red Wheelbarrow?” It reminds me of a passage from a paper I wrote in the beginning of grad school:

Even though this poem is essentially a sentence, each image is carefully isolated by means of juxtaposition. Each stanza contains images that are a juxtaposition within itself according to the line breaks: “depends” versus “upon” (two directional words going in opposite direction), the particulars of the wheel barrow (its redness and wheel[ness]) versus the wheel barrow in its wholeness, the glaze of rain versus the rainwater, the chickens versus their own whiteness. The details of the things are pulled apart and highlighted, bringing out a rich multi-faceted view of each object. Williams’ accomplishment is almost that of the cubists, allowing the reader to see these objects in many different ways, from the different angles of detail. Yet despite this almost excessive juxtaposition, the poem has a unity. It does not communicate the same fractured nature that a painting like Duchamp’s Nude Descending a Staircase. Rather, this poem explores the unity of objects, their interconnectedness, while also evoking the particularities of objects and, in some sense, how they vie with one another. What is even more striking about this poem is how commonplace it is. These are objects that many Americans in Williams’ time could see on a regular basis. For Williams to find so much juxtaposition and still unity, to be so common and yet absolutely metaphysical is a feat. More important here, we can see the way he perceives language. Each word is isolated either visually or by juxtaposition in the same way each imagistic object in the poem is isolated. This is one thing Williams does often in his poetry: isolate each word visually, either through an extreme sparseness of form or by simply leaving a word on a line by itself. What would today be considered gimmicky by most MFA students, Williams accomplishes with verve in a way that is not gimmicky in the least. This is because Williams largely helped pioneer this technique, but also because the reader senses the whole power of idiom behind Williams’ language. Its commonality is the source of its power. The idiom arises from the commonplace here. And more importantly, Williams communicates this idiom through objectness.

That passage contains perhaps one of the most absurd phrases I’ve ever written: “the chickens versus their own whiteness.” (No post-colonial analysis, that!) But no matter what you think of my analysis, the question of where these words are “broken” from one another and why is a question that I suspect Google could never answer in the form of an accurate search. Google can tell when “little rock” means the capital and not a small pebble (or some sort of midget music genre), it’s true, but humans are capable of something even more complex than that: breaking things apart and still recognizing the relationship. As the passage from my paper indicates, I believe this ability stems out of idiom (in William’s case, the American Idiom).

Ironically, Google sees everything as fractured. When you search johnny cash hurt, Boolean logic looks for johnny and cash and hurt. Google uses modified Boolean logic and has accomplished the ability to tell when certain words probably go together.

Now for the second order language intelligence of poetry: can Google understand line breaks? Could Google help us become better writers?