The Berkeley Crossword Solver – The Berkeley Synthetic Intelligence Analysis Weblog



We not too long ago revealed the Berkeley Crossword Solver (BCS), the present state-of-the-art for fixing American-style crossword puzzles. The BCS combines neural query answering and probabilistic inference to realize near-perfect efficiency on most American-style crossword puzzles, just like the one proven beneath:

Determine 1: Instance American-style crossword puzzle

An earlier model of the BCS, together with Dr.Fill, was the primary pc program to outscore all human opponents on the earth’s prime crossword match. The latest model is the present top-performing system on crossword puzzles from The New York Instances, reaching 99.7% letter accuracy (see the technical paper, net demo, and code launch).

Crosswords are difficult for people and computer systems alike. Many clues are obscure or underspecified and may’t be answered till crossing constraints are taken into consideration. Whereas some clues are just like factoid query answering, others require relational reasoning or understanding tough wordplay.

Listed here are a handful of instance clues from our dataset (solutions on the backside of this submit):

  • They’re given out at Berkeley’s HAAS College (4)
  • Winter hrs. in Berkeley (3)
  • Area ender that UC Berkeley was one of many first faculties to undertake (3)
  • Angeleno at Berkeley, say (8)

The BCS makes use of a two-step course of to resolve crossword puzzles. First, it generates a likelihood distribution over doable solutions to every clue utilizing a query answering (QA) mannequin; second, it makes use of probabilistic inference, mixed with native search and a generative language mannequin, to deal with conflicts between proposed intersecting solutions.

Determine 2: Structure diagram of the Berkeley Crossword Solver

The BCS’s query answering mannequin is predicated on DPR (Karpukhin et al., 2020), which is a bi-encoder mannequin sometimes used to retrieve passages which are related to a given query. Moderately than passages, nonetheless, our strategy maps each questions and solutions right into a shared embedding house and finds solutions immediately. In comparison with the earlier state-of-the-art methodology for answering crossword clues, this strategy obtained a 13.4% absolute enchancment in top-1000 QA accuracy. We performed a guide error evaluation and located that our QA mannequin sometimes carried out properly on questions involving information, commonsense reasoning, and definitions, but it surely typically struggled to know wordplay or theme-related clues.

After operating the QA mannequin on every clue, the BCS runs crazy perception propagation to iteratively replace the reply possibilities within the grid. This enables data from excessive confidence predictions to propagate to tougher clues. After perception propagation converges, the BCS obtains an preliminary puzzle answer by greedily taking the best probability reply at every place.

The BCS then refines this answer utilizing an area search that tries to exchange low confidence characters within the grid. Native search works by utilizing a guided proposal distribution wherein characters that had decrease marginal possibilities throughout perception propagation are iteratively changed till a domestically optimum answer is discovered. We rating these alternate characters utilizing a character-level language mannequin (ByT5, Xue et al., 2022), that handles novel solutions higher than our closed-book QA mannequin.

Determine 3: Instance modifications made by our native search process

We evaluated the BCS on puzzles from 5 main crossword publishers, together with The New York Instances. Our system obtains 99.7% letter accuracy on common, which jumps to 99.9% when you ignore puzzles that contain uncommon themes. It solves 81.7% of puzzles and not using a single mistake, which is a 24.8% enchancment over the earlier state-of-the-art system.

Determine 4: Outcomes in comparison with earlier state-of-the-art Dr.Fill

The American Crossword Puzzle Event (ACPT) is the most important and longest-running crossword match and is organized by Will Shortz, the New York Instances crossword editor. Two prior approaches to pc crossword fixing gained mainstream consideration and competed within the ACPT: Proverb and Dr.Fill. Proverb is a 1998 system that ranked 213th out of 252 opponents within the match. Dr.Fill’s first competitors was in ACPT 2012, and it ranked 141st out of 650 opponents. We teamed up with Dr.Fill’s creator Matt Ginsberg and mixed an early model of our QA system with Dr.Fill’s search process to outscore all 1033 human opponents within the 2021 ACPT. Our joint submission solved all seven puzzles in below a minute, lacking simply three letters throughout two puzzles.

Determine 5: Outcomes from the 2021 American Crossword Puzzle Event (ACPT)

We’re actually excited in regards to the challenges that stay in crosswords, together with dealing with tough themes and extra advanced wordplay. To encourage future work, we’re releasing a dataset of 6.4M query reply clues, a demo of the Berkeley Crossword Solver, and our code at

Solutions to clues: MBAS, PST, EDU, INSTATER