Summary: Crossword Puzzles as a Constraint Problem
Anbulagan and Adi Botea
NICTA and Australian National University, Canberra, Australia
Abstract. We present new results in crossword composition, showing
that our program significantly outperforms previous successful tech-
niques in the literature. We emphasize phase transition phenomena, and
identify classes of hard problems. Phase transition is shown to occur
when varying problem parameters, such as the dictionary size and the
number of blocked cells on a grid, of large-size realistic problems.
In this paper we propose new ideas in solving crossword puzzles presented in a
hybrid model with two viewpoints, one containing cell variables and the other
containing word slot variables. We discuss an architecture where search and
nogood learning exploit the strenghts of each viewpoint. Our program solves
more crossword problems than previous successful techniques in the literature.
We present the program performance on a collection of realistic puzzles, which
has been used in previous studies [1,2,3].
We also analyze the behaviour of the crossword domain in detail. We em-
phasize phase transition phenomena and identify classes of hard problems. We