The models of crossword play my lab and I are developing involve trying to understand how our memory works. Puzzles provide a great analog to memory-based decision making, because they require you to identify your action (your answer) from all the possible knowledge you have, unlike many studies of decision making which are really studies of choosing between given options. Our work is looking at how we use the cues in front of us to search memory and find the correct answer.
Much of our knowledge can be thought of in terms of associations. Good crossword clues are ones in which the association from the clue to the answer is weak, but the association, or at least the shared meanings, are strong. One of the things that makes good players good is that (because they have very good memories of some sort), those associations that are weak for most of us are strong for them, and they can thus access and generate the right answer.
What this means is that along with trying to solve a puzzle, you are also learning about future puzzles you haven't solved yet. Words appear over and over again in puzzles, and often have similar (or identical) clues, especially for the computer-generated puzzles found in some of the books you buy at airport newsstands.
I was reminded of this when on Tues;day, March 25 2014 NYT puzzle by David Wolf included the somewhat familiar clue:
47D Strand, in a way.
This was familiar because I had seen it recently in the Friday, March 21 WSJ 2014 puzzle by Maryanne Lemot:
56A Place apart
and also in the March 7 2014 NYT by Matt Ginsberg
6D Put on a key?
The answer, ENISLE, is the sort of word you only see in crossword puzzles and scrabble games, but it is somehow amusing; it seems like if this is legal, mock-words like BESHORE and ENRESORT could also work. See also: Put in a vehicle? BECAR. Deck the hall? ENCORRIDOR. Get Married? ENAISLE
The first time I saw this, I realized the trick, but could only think of 'MAROON'. The second time, 'place apart' didn't bring the previous clue to mind, and it was also a challenge, but getting the answer left a stronger memory for the previous time. In each case, because I didn't figure out the answer until I had most of the letters, I didn't get the residual benefit of of retrieval and did not form a strong association between 'isolate' and ENISLE (although the reverse is strong). But by the third time, and perhaps because it was recent, I remembered that the current clue was the same as the recent clue (put on a key), and got it right away. This happened even though I didn't remember that Ginsberg's clue (Put on a key?) was any different from Wolf's (strand, in a way).
So, the third time, I was reminded of the previous case or cases. By then, I had formed a strong association from the maroon/isolate clue and the ENISLE answer, and no longer had to carefully search for possible words. This may be episodic memory at play (a memory for an event), or some specific context-associations.
Part of our theories of memory access involve association-based cues. The clue activates a bunch of words that are related in one way or another (typically by meaning), and the player must search through these until they can find one that works. On simple non-theme puzzles, our models can achieve performance with these strategies 90% as good as experts, and better than all of our novice players. But to get the extra 10%, we need some more specific types of memory; specific associations between clues and answers.
So, part of what crossword players learn is the specific mappings between clues and answers. This is part of what expertise is in many domains. A mechanic has learned that a specific noise coming from a Honda Odyssey is likely to have some specific cause, even though that noise can generally mean many different things on many different vehicles; similarly a doctor diagnoses mappings between symptoms, risks, and diseases, but many symptoms overlap and are not specific to any particular disease, but they still can map the specific combination to the right outcome.