Crosswords in Mathematica using Pattern Matching

An alternative approach is to use SatisfiabilityInstances with constraints specifying that every row and every column must be a valid word. Code below takes 40 seconds to get first 5 solutions using dictionary of 200 three-letter words. You could replace SatisfiabilityInstances with SatisfiabilityCount to get the number of such crosswords.

setupCrossword[wordStrings_] := (
   m = Length[chars];

   words = Characters /@ wordStrings;
   chars = Union@Flatten@words;

   wordMatch[vars_, word_] := And @@ (Thread[{vars, word}]);
   validWord[vars_] := Or @@ (wordMatch[vars, #] & /@ words);
   validCell[{i_, j_}] := 
    BooleanCountingFunction[{1}, {{i, j}, #} & /@ chars];

   row[i_] := {i, #} & /@ Range[n];
   col[i_] := {#, i} & /@ Range[n];
   cells = Flatten[row /@ Range[n], 1];

   rowCons = validWord[row[#]] & /@ Range[n];
   colCons = validWord[col[#]] & /@ Range[n];
   cellCons = validCell /@ cells;
   formula = And @@ (Join[rowCons, colCons, cellCons]);
   vars = 
    Table[{{i, j}, c}, {i, 1, n}, {j, 1, n}, {c, chars}] // 
     Flatten[#, 2] &;
   decodeInstance[instance_] := (
     choices = Extract[vars, Position[instance, True]];
     grid = Table[{i, j}, {i, 1, n}, {j, 1, n}] /. Rule @@@ choices
     )
   );

n = 3;
wordLimit = 200;
wordStrings = 
  Select[DictionaryLookup[], 
   StringLength[#] == n && LowerCaseQ[#] &];
setupCrossword[wordStrings[[;; wordLimit]]];

vals = SatisfiabilityInstances[formula, vars, 5];
Framed@TableForm@decodeInstance@# & /@ vals


(source: yaroslavvb.com)

This approach uses variables like {{i,j},"c"} to indicate the cell {i,j} gets letter "c". Each cell is constrained get exactly one letter with BooleanCountingFunction, every row and column is constrained to make a valid word. For instance, constraint that first row must be either "ace" or "bar" looks like this

{{1,1},"a"}&&{{1,2},"c"}&&{{1,3},"e"}||{{1,1},"b"}&&{{1,2},"a"}&&{{1,3},"r"}

I am not sure if you would consider the following approach pattern based -- but it works, and it can conceivably be extended to many dimensions, although with the all3 dataset, it would probably konk out rather early...

The idea is to start with a blank crossword:

blankCW={{_,_,_},{_,_,_},{_,_,_}};

and then recursively do the following: For a given pattern, look at the rows in turn and (after filling out any with exactly one completion) expand the pattern on the row with the fewest number of matches:

(* Cache the number of matches for a given pattern *)
nmatch[patt_]:=nmatch[Verbatim@patt]=Length@Cases[all3,patt]

(* A helper to fill single matches if needed *)
fixone[ml_,nl_]:=If[FreeQ[ml[[nl]],Verbatim[_]],ml, 
  ReplacePart[ml, nl->First@Cases[all3,ml[[nl]]]]];

findCompletions[m_]:=Module[{nn,ur},
  (* Pattern w/ filled single matches -> ur, ordering by # of matches -> nn *)
  {ur,nn}=NestWhile[{fixone[#[[1]],First@#[[2]]], Rest@#[[2]]}&,
    {m,Ordering[nmatch/@m]},
    (Length[#[[2]]]>0&&nmatch@#[[1,#[[2,1]]]]==1)&];

  (* Expand on the word with the fewest number og matches *)
  If[Length[nn]==0,{ur},
    With[{n=First@nn},ReplacePart[ur,n-> #]&/@Cases[all3,ur[[n]]]]]];

For a given candidate pattern, try out the completion along both dimensions and keep the one that yield the fewest:

findCompletionsOriented[m_]:=Module[{osc},
  osc=findCompletions/@Union[{m,Transpose@m}];
  osc[[First@Ordering[Length/@osc,1]]]]

I do the recursion breadth first to be able to use Union, but depth first might be necessary for bigger problems. Performance is so-so: 8 laptop minutes to find the 116568 matches in the example problem:

Timing[crosswords=FixedPoint[Union[Join@@(findCompletionsOriented/@#)]&,{blankCW}];]
Length@crosswords
TableForm/@Take[crosswords,5]

Out[83]= {472.909,Null}
Out[84]= 116568
          aah aah aah aah aah
Out[86]={ ace ace ace ace ace }
          hem hen hep her hes

In principle, it should be possible to recurse this into higher dimensions, i.e. using the crosswords list instead of the wordlist for dimension 3. If the time to match a pattern against a list is linear in the list-length, this would be quite slow with a 100000+ sized wordlist...