The crossword puzzle has entranced players for more than a hundred years, testing their knowledge of vocabulary, trivia, and wits. While AI agents have been able to solve puzzles better than most human players, they have been stumped by certain linguistic challenges. That was until a team from UC Berkeley entered the picture; now, AI agents surpass even the most experienced players.
The researchers from UC Berkeley and Matthew Ginsberg LLC have recently published a paper called Automated Crossword Solving. It features the Berkeley Crossword Solver (BCS), a revolutionary, cutting-edge system capable of automatically completing complicated crossword puzzles. The system was triumphant and earned first place in the American Crossword Puzzle Tournament.
The BCS team adopts a multi-stage approach when tackling the problem-solving process. Initially, each question is addressed independently, and the resulting answers are integrated into the puzzle. Afterward, any answers that were initially uncertain are re-examined, considering the predicted limitations on the letters. This allows the team to more effectively tackle complex queries that may not be solvable without these restrictions.
The BCS pipeline includes three phases: question answering (QA), loopy belief propagation, and local search. To build their QA model, the team implements a bi-encoder architecture which generates a list of possible answers with associated probabilities for each question. Subsequently, loopy belief propagation is used to enhance the probability scores, maximizing the number of correctly placed words and letters in the solution. Finally, the grid is populated with a greedy search, then particular areas of uncertainty are improved through iterative local search.
In their experiment, the team examined the efficiency of the proposed system when tackling puzzles with themes and without. They measured the results against those of the established Dr. Fill system. During their study, the researchers compared the performance of CrossGAP against the established Dr. Fill system, evaluating their efficiency in solving puzzles with themes and without. The results demonstrated CrossGAP’s superior performance, achieving a higher accuracy rate and showcasing its potential for future developments in automated problem-solving tasks.
Last spring, a groundbreaking development for artificial intelligence occurred as an early version of the BCS triumphed over all the best human competitors at the highly respected American Crossword Puzzle Tournament. This wake-up call to crossword puzzle aficionados demonstrated AI’s capability to outperform the most skilled human players.
The research team highlighted CrossGAP’s capacity to automatically complete intricate, problem-solving processes that need expertise and logical thinking. Additionally, the system gives a glimpse into how humans decipher patterns and comprehend language. As they progress, they plan to improve CrossGAP’s efficiency and increase its usage for solving crossword puzzles in various languages.