EZ-Crossword
Solves newspaper crosswords and generates new ones with transformer QA models
EZ-Crossword started as my final year project and turned into something I still like showing people. You hand it a crossword from a newspaper and it solves it. Or you hand it a theme and it builds one.
How it works
Solving is a question answering problem with a twist: every clue has a known answer length and shares letters with its neighbours. A transformer QA model produces candidate answers for each clue, and a constraint solver picks the combination that keeps the grid consistent. When the model is unsure, the intersections usually settle it.
Generation runs the same pieces backwards. Given a set of answer words, the generator lays out a valid grid, then asks the model to write clues that are specific enough to be fair.
What I learned
- Constraint propagation does most of the heavy lifting. The model only needs to be good at ranking, not at being right on the first try.
- Async job queues matter even for a demo. Solving takes seconds, so the web app polls a job status endpoint instead of holding the request open.
- Clue quality is the hard part of generation and it is where better language models made the biggest difference.