Can computers crack crosswords?
Google Goggles is a tool that allows your smartphone to “see” what’s around you—a landmark, say, or a logo—and to search for that thing without the need for language. I vividly remember the first time I saw the software in action. I opened a newspaper on the sudoku page, pointed my cell phone at the page, and took a snapshot. In a matter of seconds, the screen showed a Google-generated image with the puzzle correctly completed. I stared at the image for a moment or two, utterly unflabbergasted.
These moments of gob-unsmacked, wholly plussed non-wonder at Things Computers Can Now Do But Once Couldn’t are becoming more frequent as technology bounds along—almost part of everyday life. And sudoku is perhaps one of the least incongruous activities a computer might take a crack at.
If you’re a human solving a sudoku, you’re essentially slowly working through algorithms colossally better suited to a central processing unit. It’s the cognitive equivalent of deciding that machines aren’t best placed to handle your e-mail and resolving instead to handwrite a message, walk it across to a Palo Alto server farm, then walk back, before placing it on the desk of your colleague and returning the six feet to your own workstation. Well done, you. Go, human!
With crosswords, it feels different. Solving a crossword makes you feel something. You’ve combined feats of memory with some lateral thinking and teased out the hidden treasure. It feels creative, thoughtful. It’s a little humbling, then, when you see the strides that computers are making as crossword solvers.
Artificial intelligence researchers have long been interested in the crossword as the kind of thing that computers might find challenging—but not necessarily impossible. The most prominent of the current crossword bots is called Dr. Fill (a pun on Phil McGraw’s psychology TV show Dr. Phil). Dr. Fill was created by former artificial intelligence researcher Matt Ginsberg, who runs algorithms for the US Air Force when he isn’t constructing New York Times puzzles.
Ginsberg wanted his machine to be portable, so he loaded up a MacBook Pro laptop with the code he’d created. He added preexisting crossword answers dating back to 1990 and chunks of resources such as Wikipedia and the Internet Movie Database, and took it along to the 2012 American Crossword Puzzle Tournament.
Dr. Fill was not an official entrant: One of the requirements for entry to the tournament is that you be a person. If a person had performed the same as Dr. Fill, he or she would have come in 140th—not a bad placing, and one that can only improve as more time and human thought goes into the software.
Ginsberg, though, doesn’t think that what Dr. Fill does counts as thinking. It is, he says, a serial business of summoning likely answers and seeing whether they fit with one another. “Thinking” or not, that’s a pretty good description of how most humans approach a puzzle. There is, though, something missing from Dr. Fill’s approach, something that should become more apparent as it learns from its mistakes—and that’s a discipline it’s certainly more serious about than many of its biped rivals.
A hint as to what that missing element might be came when the 2012 tournament included an unexpected twist in one of the puzzles, whereby some answers had to be entered unconventionally. A human might wonder for a while what is going on when the answer MOONMISSIONS doesn’t seem to fit the squares for “Apollo 11 and 12 [180 degrees].” But once you twig what “180 degrees” is asking for and see that SNOISSIWNOOW fits with the crossing answers, you can enter it with the kind of confidence that a machine isn’t going to have unless it’s been given a line of code that explicitly says that Ws can be exchanged for Ms under certain conditions.
And while adding such extra possibilities is not an insurmountable technical problem, that moment of realization is where automated solvers currently part company with their flesh-and-blood rivals. For Dr. Fill, this would be merely another device in the armory. For human solvers, there’s something endearingly daft about entering an answer the wrong way up, the words hanging upside down like a spaceman on a moon mission.
Computers will make qualitative progress in tone, speed, range, and so on—but there’s something spooky about giving them clues whose whole purpose is a silliness and humor that they seem eternally unlikely to be able to enjoy.
Perhaps there’s a better job for technology in creating puzzles rather than solving them. While the range of available grids was once determined by the lead blocks from which the squares were printed, that job is now done by the software used by the constructors, editors, and printers of crosswords.
What about the clues? Some take comfort in the assumption that the architectural donkeywork—providing grids and even words to fill them—might be the extent of computers’ involvement. In 2003, the former crossword editor of the London Sunday Times Barbara Hall told the BBC that she thought a further impediment would be the size of the word bank the machine would require “because there are so many different meanings for one word.”
However, in technical terms, there’s barely a difference between a “word bank” of a few thousand words and one of a couple of hundred thousand. It would require no ingenious coding to program a computer to fill grids with answers from a list—of which there are plenty available—then clue each with a database of synonyms.
The resulting puzzle would, technically speaking, be a crossword—indeed, some of the shoddier collections in book form give the impression of having been thus compiled. More interesting is the question: Would it satisfy the solver?
The answer, I suspect, is: by no means all of them. As Stanley Newman has remarked, computer-generated puzzles tend to be full of “junk: foreign phrases, weird abbreviations and obscure words so unfamiliar they don’t even qualify as crosswordese.”
A human constructor is better able to imagine the experience of answering the clues—where the starting points are likely to be and how each solver’s unique journey through the grid might unfold. More importantly, the solver of a decent puzzle in a decent paper does not think of the exercise as an abstract means of whiling away some minutes but as a contest between two people, where the solver knows that the constructor has conceived of the grid as a whole, balanced in terms of tone, subject matter, technique, and difficulty.
More importantly, the constructor’s role is, in a phrase much beloved in the crosswording world, to “lose gracefully”: to guess correctly that the solver will, with enough application, find the wherewithal to topple every clue and fill every square.
Each constructor’s idea of what the solver is likely to know differs, based on hunches and experience—and those differences are as good a way as any of defining the varying personalities among constructors. It’s with constructors’ personalities that solvers make the relationships that keep them coming back. “I beat Fred Piscop today” means something qualitatively different from “I beat Midweek-Bot v2.1.” And once a solver has found the newspaper that suits him, he’s made a relationship with a kind of gang—and a gang of highly characterful and diverse individuals who share an ethos and an editorial guiding hand.
Construction is regarded by solvers as a kind of authorship, and our relationship with an author is dependent on his or her being a person, too. The editors of the London Daily Telegraph discovered this in 1998 when they initiated a scheme to automate crossword production. Humans would still be paid to write clues, but at that point they would enter a database from which each day’s puzzle could be assembled.
The ostensible reason was to reduce the frequency with which some words appeared as answers, although there was also an undeniable financial incentive. A core of the constructors refused to play ball and became known as the “Telegraph Six.” Ruth Crisp (Crispa) gave her withering assessment of the wheeze: “I don’t think a crossword done on a computer can possibly compare with one done individually. I have been compiling crosswords for half a century and I think my judgment can be depended on.”
The invisible symmetries of human arrangement could not be replicated by machine, she insisted, and Roger Squires (Rufus) predicted that the tone of the resulting puzzles would be “like combining the musical styles of Beethoven and Mozart in the same musical movement.” This was all the more telling coming from the constructors for a newspaper that leaves each of its puzzles anonymous.
The paper’s deputy editor, Boris Johnson—who has since become the mayor of London—was forced to agree. “In spite of the advantages the computer possesses, the machine has been condemned for a fatal lack of soul,” he announced. “The crossword will remain a duel of wits between the individual composer and the solver.” Squires responded with a single tart clue—“Submit to pressure and return to base (9)”—and returned to business. The answer? CLIMBDOWN.
(It should be granted, though, that the home of the world’s first programmable electronic digital machine was a place rammed with super-solvers . . .)