Clue to 15-Across: Crossword Puzzle Solver (22 Letters)—Artificial Intelligence

by Jamie Beckett

Now there’s a better way to cheat at crossword puzzles—with artificial intelligence.

Researchers from three universities designed a system that uses a type of AI known as deep learning to help computers understand language more quickly and effectively. As a bonus, the researchers built a web-based tool that’s handy when a particularly baffling crossword clue comes around.

The crossword assistant is just a demo of how deep learning is improving machines’ ability to understand language, said Felix Hill, a Ph.D. candidate at the University of Cambridge and lead author of a paper describing the research. Learning language is challenging for computers because it’s difficult to recreate the rich and diverse information sources available to humans when they learn to speak and read, he said.

Researchers focused on helping computers understand phrases. Although computers are pretty good at understanding words, Hill said, “most of the useful information in the world is communicated in phrases.”

Training Neural Networks

The team turned to deep learning as a way of bridging the gap between machines that understand the meanings of individual words and machines that can understand the meanings of phrases and sentences. In deep learning, researchers train artificial neural networks that learn naturally from massive amounts of data to recreate human abilities. (See “Accelerating AI with GPUs: A New Computing Model.”)

The group trained its neural networks on definitions from six dictionaries and Wikipedia, using our GeForce GTX TITAN Black GPUs for the speed needed to process hundreds of thousands of definitions. That sped up training from several days to roughly 10 hours, Hill said.

“Our system can’t go too far beyond the dictionary data on which it was trained, but the ways in which it can are interesting. And this makes it a surprisingly robust question and answer system — and quite good at solving crossword puzzles,” Hill said.

More to Learn

The deep learning system wasn’t particularly good at puns or abstract questions that require a logical leap. But it was better at answering standard crossword clues in tests against commercial products designed for the task, according to the paper.

So how well does it work? We tossed the online tool a few clues from a puzzle by Merl Reagle, a nationally syndicated crossword creator for more than 30 years. It dispatched some of the straightforward clues with ease: “deterioration” is indeed a 13-letter word for “gradual decline” and Edam is a type of Dutch cheese. But it failed to answer other simple clues like “looked up to” (respected) and “fearful one” (scaredycat).

After training on dictionaries and Wikipedia, the deep learning system isn’t exactly steeped in cultural awareness. Its answer to “epic film of 1979”? The Iliad, Beowulf and other literary epics — but it didn’t come up with “Apocalypse Now.” And “Flying Pan”? The tool responds with names of various birds and fish, but never with that boy from Neverland, Peter.

The next step, said Hill, is to train the neural networks on more data, and eventually incorporate direct interactions with people. Researchers could make improvements to the based on their feedback, he said.