A neural network created by a group at the University of California has beaten 1,300 participants in an American crossword puzzle tournament for the first time. The model was trained on 6 million crosswords already solved and used the Wikipedia content to answer the questions.
The Dr. Fill algorithm was originally developed by Oxford astrophysicist Matt Ginsberg. The principle of its operation was similar to that used in the Deep Blue chess supercomputer, based on brute-force methods, and did not contain deep learning methods. For the first time, Dr. Fill was used at the championships in 2012, and then took 141st place. So far, the best result – the 11th place-the algorithm showed in 2017. Two factors allowed us to win in 2021. First, Dr. Fill was initialized not on a regular laptop, but on a 64-core processor and two video cards. Second, a research team at the University of California, performing research in the field of natural language processing, joined in improving the algorithm.
Video recording of Dr. Fill’s solution of six crossword puzzles of the tournament:
The researchers integrated a neural network into the algorithm, similar to those used in voice assistants such as Siri and Alexa. To train Dr. Fill, we used a database of solved crosswords consisting of 6 million hint-answer pairs. To answer the questions, the entire content of Wikipedia was uploaded to the neural network. During the tournaments, Dr. Fill was disconnected from the Internet.
The development of the new version of Dr. Fill took two weeks. While the original algorithm performed the mathematical analysis of the grid, searching and placing the answers, the neural network performed the recognition of hidden semantic structures in the hints. The neural network made three mistakes during the tournament, but showed the best response rate – 49 seconds to solve the crossword puzzle. The second-placed player solved a similar crossword puzzle without errors, but in three minutes, which allowed Dr. Fill to win by a margin of 15 points.