Skip to content
“ I was terrible with the crosswords, so I built an AI to do them ”

Matt Ginsberg is good at many things – he’s an AI scientist, author, playwright, magician, and stunt plane pilot. But he’s not very good at crosswords.

In fact, despite writing them for The New York Times, he says when they are published, he often fails to resolve his own.

So, while he was sitting in a hotel ballroom and once again losing in a big crossword puzzle contest in the United States, he decided to do something about it.

“I was with 700 people who were really good at solving crosswords and it bothered me for being so terrible, so I decided to write a computer program that would be profitable on my behalf,” he said. he told the BBC.

And finally he did. After 10 unsuccessful attempts, Dr Fill – as the program is known – has just won his first competition.

He came out on top in the American Crossword Tournament, the premier crossword puzzle competition in the United States.

Dr Fill has been trained on a mass of data, including a giant database of crossword clues and answers pulled from the web.

She was taught to quickly search for possible word locations in a crossword puzzle. It was, Dr. Ginsberg admits, a fairly “primitive” system.

This year he got help.

“A few weeks before the event, I was contacted by people working at Berkeley who had built a crossword response system. We realized pretty quickly that we could combine the two.”

Professor Dan Klein, who heads the Natural Language Processing group at Berkeley College, University of California, told the BBC he was looking for something to bring the team together during the lockdown – and they got the idea to create a crossword solver.

When he heard about Dr Fill, he thought the two systems would be a good partnership.

Crosswords can fool even the smartest humans, so how would a machine do?

“Our system brought a broader understanding of language, and Dr. Fill was well aware of how the answers combine with other clues. They are very different techniques, but they spoke a common language of probability.”

Poisoned penne

Crosswords might seem like an odd thing for AI to solve, but they actually represent a very fertile playground for machine learning.

Basic crosswords that simply require someone to know the answer to the prompt are extremely easy for an AI, which will have been programmed with large amounts of information coming from the web from sources such as Wikipedia.

Cryptic crosswords, which a UK audience might be more familiar with, are also fairly easy on a machine, as they contain very defined rules and pointers to things like anagrams.

American-style crosswords, on the other hand, require both knowledge and a certain degree of lateral thinking.

One question Professor Klein is particularly proud to have got right of Dr. Fill was: “The pasta dish at the center of a murder mystery.”

The answer was poisoned penne. “It’s not on Wikipedia,” said Professor Klein.

Dr. Ginsberg agrees that American-style crosswords can be “brutally difficult” for computers to understand.

Dr. Fill made just three mistakes in the overall competition, even though he ended up winning only by a small margin.

Dr Ginsberg did not receive the $ 3,000 (£ 2,100) prize, which he said had been agreed in advance and was “the right decision”.

He admitted that it is delicate for the organizers of competitions, if humans and machines are involved.

Fortunately, he says, the crossword community is a “wonderful bunch.”

While competitors may claim to hate his AI rival – booing Dr. Fill when he does well and applauding if he hurts – deep down, he thinks the competitors “were really cheering me on.”

He doesn’t know for sure, as this year’s event was virtual – which meant he couldn’t see any of the contestants.

This meant, however, that Dr Fill could benefit from additional computing power, which would not normally have been transportable.

This achievement was praised by DeepMind, a leading AI research firm that is not used to winning games – including beating a world-class Go player in 2016.

Michael Bowling, Principal Investigator at DeepMind and Professor of Computer Science at the University of Alberta, said of the victory: “Congratulations to Dr. Ginsburg and the team at Berkeley. and see powerful building blocks of research and learning AI used together.

“Knowing that there is another better crossword puzzle than I am will not change my enjoyment of wrestling a Tuesday puzzle.”

Learn differently

The move to what’s called general-purpose AI, where a machine can perform a series of tasks rather than just being in one domain, is currently a long way off, but a lot of progress has already been made.

Natural language processing – Professor Klein’s specialization – has already recorded achievements in real-world scenarios as diverse as translation, speech recognition and enabling the everyday conversations we have with voice assistants.

But, said Professor Klein, we are only at the beginning of our understanding of how machines learn.

“Our understanding of what is easy and what is difficult for computers is a moving target. People were amazed that a computer could compete in chess, but now we think it is amazing that a human could compete with a chess machine.

The way a computer decides which movement to make in this game is a combination of “math, logic and looking to the future”, which is unlikely to be exactly the same way a human would view the same movement. , did he declare.

Dr Ginsberg agrees that humans and machines approach problems from different angles.

“Dr. Fill solves these puzzles very differently from the way we do. He does a giant search for all possible answers.”

This diversity is, he said, a “good harbinger” for the future.

“We will solve more problems with them on our side than we can on our own. We will team up with machines to our mutual benefit.”

But he has no plans for world domination yet.

“Dr. Fill is just a crossword puzzle program and I agree with that.”

Source link