Celebrating National Crossword Puzzle Day
At BasisTech, we love words so we are celebrating National Crossword Puzzle Day (December 21) with a little bit of crossword history, the application of AI to puzzle solving, and a puzzle of our own.
The first crossword puzzle was created by journalist Arthur Wynne and published on December 21, 1913 in the New York World newspaper. Crossword puzzles gained popularity in the US in the 1920s and most major newspapers have had some form of crossword since then.
In a twist of irony, The New York Times, one of the current leading purveyors of crossword puzzles, complained in 1924 of the “sinful waste in the utterly futile finding of words the letters of which will fit into a prearranged pattern, more or less complex. This is not a game at all, and it hardly can be called a sport … [solvers] get nothing out of it except a primitive form of mental exercise, and success or failure in any given attempt is equally irrelevant to mental development.” The NYT only began publishing crossword puzzles in 1942.
In fact, crossword puzzles have long been studied for their effect on mental acuity, especially in aging adults with early signs of dementia. In a recent paper, researchers found that in key measurements — cognitive decline scores, functional skills, and brain volume changes — the regular crossword players fared better than players of other computer games.
Studies have also shown the positive effect of crossword puzzles on vocabulary retention, which explains why they are a favorite tool of teachers seeking to build students’ word power. The challenging nature of crossword puzzles encourages students to pay more attention to the words as they fill in the boxes. This subsequently enhances their retention of vocabulary, builds positive perceptions, and develops cognition in their linguistic domains.
Crossword puzzles require some distinctly human skills – the first is the ability to process semantic information. We need to understand the meaning of words. We also need broad knowledge of the subject matter – in crossword puzzles this covers a range of history, art, literature, culture, and language. A puzzle solver also needs the ability to abstract their answers from often obtuse clues. Finally, those answers need to fit within the orthographic constraints of the puzzle.
We know that people of all skill levels can do crossword puzzles, but what about computers? Many people have asked this same question, notably Matt Ginsberg, the creator of the AI puzzle-solver “Dr. Fill.” Although Dr. Fill performed admirably in the American Crossword Puzzle Tournament for years, it wasn’t until 2021 that it took first prize. That was the year Dr. Fill’s more traditional AI was augmented by a neural network developed and described in a paper by the Berkeley NLP Group.
Ginsberg originally programmed Dr. Fill with the entire contents of Wikipedia along with some specialized databases of crossword clues and answers. During competition, Dr. Fill would rapidly search its knowledge base for answers and calculate the probability that the answer was correct based on the constraints of the puzzle grid. The Berkeley NLP group trained an AI model using 6 million clues and answers, giving Dr. Fill the ability to apply the generalizations and abstractions needed to handle the more “human language” side of the clues.
Not to be outdone, a collaboration between Expert.ai and the University of Siena resulted in WebCrow 2.0 in 2022. This AI uses knowledge graphs, ontologies, the self-updating repository of information found on the web, and the experience of solving previous crosswords to apply a range of strategies to solving new puzzles. And, it can do it across languages.
While it may seem frivolous to train an AI to solve crossword puzzles, the unique nature of these games provides an important testbed for AI capabilities. This is because solving crossword clues goes beyond typical question and answer frameworks, and beyond search. Crosswords present a diverse set of challenges that can evolve our understanding and development of AI models for solving abstract problems.
As you celebrate National Crossword Puzzle day, test your knowledge of AI and linguistic vocabulary with our own puzzle offering. [Use full screen option for the best experience.]