Clever Clues Clobber Crossword Computer 70
Hugh Pickens writes "Steve Lohr reports that an impressive crossword-solving computer program called Dr. Fill matched its digital wits against 600 of the nation's best human crossword-solvers, finishing only 141st at the American Crossword Puzzle Tournament in New York. 'I wish it had done better,' says Dr. Matthew Ginsberg, the creator of Dr. Fill and an expert in artificial intelligence. Dr. Fill typically thrives on conventional crosswords, even ones with arcane clues and answers; it solved one of the most difficult puzzles at the tournament perfectly. But the computer does poorly with clever clues based on puns or jokes, because humans and machines solve the crosswords very differently. Humans recognize patterns based on accumulated knowledge and experience, while computers make endless calculations to determine the most statistically probable answer. The computer program is literal minded, and tends to struggle on puzzles with humor, and puzzles with unusual themes or letter arrangements. Take this clue from a 2010 puzzle in The Times: Apollo 11 and 12 (180 degrees). The answer is SNOISSIWNOOW, seemingly gibberish. A clever human could eventually figure out that those letters, when rotated 180 degrees, spell MOON MISSIONS. Humans get the joke, while a literal-minded computer does not. 'Occasionally, Dr. Fill just doesn't get it,' says Ginsberg. 'That's my nightmare.'"
He named it a pun, but it sucks at puns. (Score:2, Funny)
Can it manage ironic clues?
GIGO (Score:2)
My favourite "not real" crossword clues (Score:5, Funny)
I'm such a funny guy...
Oh - another one is to say "seven up is lemonade"...
Re: (Score:2, Funny)
Perhaps unsurprisingly, one down is justifiable homicide.
Re: (Score:3)
He was buried 6 down and 4 across.
^^ you can have that one for free
I'd luck to congratulate (Score:5, Interesting)
I'd luck to congratulate submitter on a clever title. Does not happen very often here.
Re: (Score:2)
*like. If somebody thought that this was some kind of word play and modded up because of that (that's the only explanation of modups I have: my typo somehow made my comment clever beyond my understanding), it was not.
Re: (Score:2)
Computers have no sense of humor... (Score:2)
Re: (Score:2)
Non ex transverso, sed deorsum.
Poor example? (Score:2)
Would the Apollo example really trip up a decently-written program though? I mean, my first thought was "Well, what if it had a fallback routine where it tries anagrams of possible answers?" so I have to imagine someone smarter than me has thought of that. I guess there's some limitation I'm not seeing...
On second thought...what am I still doing awake at 5:48am commenting on a post about crossword-solving computers?!
Re: (Score:2)
Re:Poor example? (Score:4, Funny)
Nor is it a word in any conventional sense, nor comprised of words. The only place it ever appears is in relation to discussion of this particular clue. The correct clever human response to such thing is to punch the setter in the face; they have broken the formal compact of crossword setting by using a non-word/non-phrase as an answer.
Re: (Score:2)
Exactly. I like doing crosswords, but it hate it when they do stupid shit like that. I really hate it when the answer is two words and there's no indication of it.
Re: (Score:3)
There is a fine line between 'subtle clue' and 'inside jok
Re: (Score:2)
Exactly. If not, then what's to stop somebody from using "bunch of random letters" as a clue to a bunch of random letters?
OTOH, without such back-assed clues, a brute force attack would probably be more efficient at solving a crosswords than some AI.
Re: (Score:2)
There is no formal contract. Also, if you look at the information on that particular puzzle that was linked in TFA, it appears the crossword setter pretty much gave the game away as to what the trick was in one of the other clues. [cruciverb.com]
Re: (Score:2)
I don't think it's that simple. Of course you could write the program to include answers made of letters rotated 180 degrees. But how often does an answer like that happen? I'm not a big crossword person, but i'd guess it's pretty rare. Maybe 1 out of 10,000 - if even that much? So while you might get that one clue right, meanwhile you've expanded the set of possible answers to include numerous elements with an extremely low chance of ever being correct. In addition to the added expense of considering them,
Re:Poor example? (Score:4, Interesting)
Here is one that just appeared this week (LA Times, I think):
Clue: Hail Answer: DANTESINFERNO
Clue: Poe Answer: FLATBROKE
Clue: What you need to get the above two answers: SOUTHERNDRAWL
Not sure how you make a routine to come up with those answers.
Re: (Score:1)
Re: (Score:2)
Christ! Those are... Those are just ridiculous... I guess what I didn't understand is what crosswords have become!
Now I get what they mean by requiring some cleverness... I mean, there're PEOPLE who wouldn't get those! It makes a lot more sense now.
Not cryptic though (Score:1)
Re: (Score:2)
Re: (Score:2)
A clue definition of '...V' is acceptable, if it follows on from a previous related question or answer. Even then it may seem related, but isn't. This will throw one off the scent, but the answer will often be part of an overarching theme. A theme often allows weaker definition cluing. Several other clues might include tangential references to science, but there will one clue that is often refered to by clue number only, elsewhere in the puzzle - e.g. 'Amphibian is working by force' (6) - hinting at the ove
Re: (Score:2)
Re: (Score:1)
Re: (Score:2)
Re: (Score:1)
Re: (Score:2)
There is a nasty one inside the crossword app on the Nook Color (not sure if it's in other Nooks, but I'm guessing it is) where the answer is spelled "CENTRE". The problem is the down answers really want "CENTER" to make any sense (one of the down ones was "TENT" which became "TRNT").
Not sure if it was a typo or not. And the puzzles have no identifier so you can point it out.
Re: (Score:2)
Other one AI might struggle with:
Lisping girl of legend (4) - (ans: 'Myth')
How about some Cockney Rhyming slang?:
Beehive in North London? (4, 6) - (ans: 'High Barnet')
Re: (Score:1)
Use of cockney isn't uncommon so it would make sense to include both of those in the definition for "hairstyle".
Re: (Score:2)
It probably has a lot to do with the frequency of letters in a language. I can't think of any reasonable way to fill out all the 4x4 blocks in your puzzle with actual words in my native language, Dutch. I'm guessing the distribution of letters in Finnish is much more skewed.
Looks like (Score:1)
the computer needs help from Joe Piscopo [youtube.com]
My favourites (Score:2)
I'd like to see if Dr Fill manages these two:
HIJKLMNO (5)
___ (2, 3, 4, 1, 4)
Re: (Score:1)
Yes, I remember seeing that one. One of those clues that filled me with delight when I got it.
___ (2, 3, 4, 1, 4)
To Not Have A Clue?
Re: (Score:2)
HIJKLMNO (5)
That reminds me of this one:
ABCDEFGHIJKMNOPQRSTUVWXYZ (4)
Redundancy (Score:4, Insightful)
Not being able to guess a few words might not be a problem, skip it and solve the other ones, once there are enough letters in it a computer can easily look up the available words, and if there are more than one even use a nonlinear approach. Even without any clues, a few words can't be that hard to bruteforce.
Re: (Score:2)
A computer can bruteforce much faster than a human.
Re: (Score:3)
A human can cleverly clobber a crossword computer with brute force.
Re: (Score:2)
Re: (Score:2)
Not necessarily. The computer likely double checks cross clues to verify any individual answer. If one of answers starts to appear to be jibberish, it will throw a large swath of intersecting answers in doubt. If the jibberish answer crosses a large portion of the grid, the doubt can propagate through the entire puzzle.
"That's my nightmare" (Score:2)
Unfortunately for Dr Fill's creator, the problem of how to get the program to work with such unorthodox solutions is the same as getting it to think like a person. At a certain point, all AI questions become the same AI question: this is the very essence of Turing horizon, and all such efforts converge there.
The program he wants to write is, sadly, doomed, as it will be impossible until such time as our species generates a true artificial consciousness with human intelligence, at which point the problem wi
Witty Watson Won (Score:1)
The crossword puzzle guy needs lessons from Watson, who clobbered several Jeopardy human champions [slashdot.org]. That show has clever categories and clues. Watson probably had more impressive computing power, but I doubt that was the issue. The Watson designers clearly had a better grasp of natural language, including humor-filled and storied language.
different processes? (Score:2)
Humans recognize patterns based on accumulated knowledge and experience, while computers make endless calculations to determine the most statistically probable answer.
what's different about that? I've often said, "The more I learn about AI, the more I think it lacks any intelligence at all. The more I learn about psychology, the more I believe that humans think just like an AI."
Humans are also determining the most statistically probable answer. They just have a better algorithm for factoring humor into those statistics.
Re: (Score:2)
Re: (Score:3, Funny)
So Doctor Fill... (Score:2)
So Doctor Fill has the same ability to comprehend humour as Doctor Phil?
SNOISSIWNOOW (Score:1)
I thought Crosswords... (Score:1)
...were supposed to be composed of REAL words.
WTF is "SNOISSIWNOOW"??
I hate crossword puzzles (Score:2)
I can't blame the computer for not doing well on these; a lot of crossword puzzles are a puzzle of "guess what the creator was thinking", and not a puzzle of words and language. Quite frankly, I'm not interested in guessing what someone else happens to be thinking when they write down a clue like "blue, red, and big"; I find that fundamentally uninteresting and of no long term value.
I have the same problem with many Mensa puzzles. A lot of them I can do, but puzzles that require deep and specific informat