91av

AI beats a thousand of the world’s best human crossword solvers

A crossword-solving AI has beaten more than a thousand of the best human solvers at the prestigious American Crossword Puzzle Tournament
2J2KT94 Stamford, Connecticut, USA. 02nd Apr, 2022. Solvers warm up before the start of play at the 44th Annual American Crossword Puzzle Tournament. The ACPT, founded in 1978 by New York Times puzzle master Will Shortz, is back this year as an in-person event after a cancellation in 2020 and a virtual competition last year.(Credit Image: ? Brian Cahn/ZUMA Press Wire)
Solvers warm up before the start of play at the 2022 American Crossword Puzzle Tournament
ZUMA Press, Inc./Alamy

AI has beaten more than a thousand of the best human crossword solvers at a prestigious competition and can complete The New York Times crossword more than 80 per cent of the time.

at the University of California, Berkeley, and his colleagues created a two-part AI model called the as a collaborative project during the covid-19 lockdown.

One component analysed individual crossword clues and created a range of solutions that could fit. The model was trained on a data set of 6 million pairs of clues and correct answers from a range of crosswords including The New York Times and the Los Angeles Times. The second component took these individual guesses and used them to complete the puzzle, selecting those that both correctly answered the question and fit into the grid and other solutions.

“Crosswords are a nice mix of this very human sort of creative language problem, on one hand,” says Klein. “And at the same time, they’re on a grid, and that grid has exactly the same kinds of configurational rules that other games like chess or Go have. There’s this mix of this very human problem and a very analytical reasoning problem.”

Klein and his collaborators designed both components of the AI but found that their question-answering module outperformed their puzzle-completion model. So, they borrowed code from the . The researchers obtained special permission to enter the AI in the 2021 American Crossword Puzzle Tournament, where it earned first place.

The American Crossword Puzzle Tournament, which attracts more than a thousand of the best human solvers, uses American crosswords, which employ wordplay, allusion and riddles, but aren’t as complex or convoluted as cryptic crosswords. Klein says it is difficult to know whether cryptic crosswords would be harder or easier for AI to solve. “Cryptics have more going on; you find the word, there’s an anagram, they have to rhyme… But that might be easy for a machine. I don’t know,” he says.

The researchers have since made improvements to their own puzzle-completion component. In tests, it now outperforms the system they took to the competition. Dr. Fill was previously the state-of-the-art AI crossword solver to beat, managing to solve 57 per cent of The New York Times crosswords, but the Berkley Crossword Solver has now bested that to complete 82 per cent. However, Klein says he and his team are unlikely to return to competitions in the future now that pandemic restrictions have eased and they are able to work on other projects again.

Reference:

Topics: AI