290 likes | 411 Vues
Combining KR and search: Crossword puzzles. Next: Logic representations Reading: C. 7.4-7.8. Changes in Homework. Mar 4 th : Hand in written design, planned code for all modules Mar 9 th : midterm Mar 25 th : Fully running system due Mar 30 th : Tournament begins. Changes in Homework.
E N D
Combining KR and search: Crossword puzzles Next: Logic representations Reading: C. 7.4-7.8
Changes in Homework • Mar 4th: Hand in written design, planned code for all modules • Mar 9th: midterm • Mar 25th: Fully running system due • Mar 30th: Tournament begins
Changes in Homework • Dictionary • Use dictionary provided; do not use your own • Start with 300 words only • Switch to larger set by time of tournament • Representation of dictionary is important to reducing search time • Using knowledge to generate word candidates could also help
Midterm Survey • Start after 9AM Friday and finish by Thursday, Mar. 4th • Your answers are important: they will affect remaining class structure
Crossword Puzzle Solver • Proverb: Michael Litman, Duke Univ • Developed by his AI class • Combines knowledge from multiple sources to solve clues (clue/target) • Uses constraint propogation in combination with probabilities to select best target
Algorithm Overview • Independent programs specialize in different types of clues – knowledge experts • Information retrieval, database search, machine learning • Each expert module generates a candidate list (with probabilities) • Centralized solver • Merges the candidates lists for each clue • Places candidates on the puzzle grid
Performance • Averages 95.3% words correct and 98.1% letters correct • Under 15 minutes/puzzle • Tested on a sample of 370 NYT puzzles • Misses roughly 3 words or 4 letters on a daily 15X15 puzzle
Questions • Is this approach any more intelligent than the chess playing programs? • Does the use of knowledge correspond to intelligence? • Do any of the techniques for generating words apply to Scrabble?
To begin: research style • Study of existing puzzles • How hard? • What are the clues like? • What sources of knowledge might be helpful? • Crossword Puzzle database (CWDB) • 350,000 clue-target pairs • >250,000 unique pairs • = # of puzzles seen over 14 years at rate of one puzzle/day
How novel are crossword puzzles? • Given complete database and a new puzzle, expect to have seen • 91% of targets • 50% of clues • 34% of clue target pairs • 96% of individual words in clues
Categories of clues • Fill in the blank: • 28D: Nothing ____: less • Trailing question mark • 4D: The end of Plato?: • Abbreviations • 55D: Key abbr: maj
Expert Categories • Synonyms • 40D Meadowsweet: spiraea • Kind-of • 27D Kind of coal or coat: pea • “pea coal” and “pea coat” standard phrases • Movies • 50D Princess in Woolf’s “Orlando”: sasha • Geography • 59A North Sea port: aberdeen • Music • 2D “Hold Me” country Grammay winner, 1988: oslin • Literature • 53A Playwright/novelist Capek: karel • Information retrieval • 6D Mountain known locally as Chomolungma: everest
Candidate generator • Farrow of “Peyton Place”: mia • Movie module returns: • 0.909091 mia • 0.010101 tom • 0.010101 kip • 0.010101 ben • 0.010101 peg • 0.010101 ray
Ablation tests • Removed each module one at a time, rerunning all training puzzles • No single module changed overall percent correct by more than 1% • Removing all modules that relied on CWDB • 94.8% to 27.1% correct • Using only the modules that relied exclusively on CWDB • 87.6% correct
Word list modules • WordList, WordListBig • Ignore their clues and return all words of correct length • WordList • 655,000 terms • WordListBig • WordList plus constructed terms • First and last names, adjacent words from clues • 2.1 million terms, all weighted equally • 5D 10,000 words, perhaps: novelette • Wordlist-CWDB • 58,000 unique targets • Returns all targets of appropriate length • Weights with estimates of their “prior” probabilities as targets of arbitrary clues • Examine frequency in crossword puzzles and normalize to account for bias caused by letters intersecting across and down terms
CWDB-specific modules • Exact Match • Returns all targets of the correct length associated with the clue • Example error: it returns eeyore for 19A Pal of Pooh: tigger • Transformations • Learns transformations to clue-target pairs • Single-word substitution, remove one phrase from beginning or end and add another, depluralizing a word in clue, pluralize word in target • Nice X <-> X in France • X for short <-> X abbr. • X start <-> Prefix with X • X city <-> X capital • 51D: Bugs chaser: elmer, solved by Bugs pursuer: elmer and the transformation rule X pursuer <-> X chaser • http://www.oneacross.com
Information retrieval modules • Encyclopedia • For each query term, compute distribution of terms “close” to query • Counted 10-k times every times it apears at a distance of k<10 from query term • Extremely common terms (as, and) are ignored • Partial match • For a clue c, find all clues in CWDB that share words • For each such clue, give its target a weight • LSI-Ency, LSI-CWDB • Latent semantic indexing (LSI) identifies correlations between words: synonyms • Return closest word for each word in the clue
Database Modules • Movie • www.imdb.com • Looks for patterns in the clue and formulates query to database • Quoted titles: 56D “The Thief of Baghdad” role: abu • Boolean operations: Cary or Lee: grant • Music, literary, geography • Simple pattern matching of clue (keywords “city”, “author”, “band”, etc) to formulate query • 15A “Foundation of Trilogy” author: asimov • Geography database: Getty Information Institute
Synonyms • WordNet • Look for root forms of words in the clue • Then find variety of related words • 49D Chop-chop: apace • Synonyms of synonyms • Forms of related words converted to forms of clue word (number, tense) • 18A Stymied: thwarted • Is this relevant to Scrabble?
Syntactic Modules • Fill-in-the-blanks • >5% clues • Search databases (music, geography, literary and quotes) to find clue patterns • 36A Yerby’s “A Rose for _ _ _ Maria”: ana • Pattern: for _ _ _ Maria • Allow any 3 characters to fill the blanks • Kindof • Pattern matching over short phrases • 50 clues of this type • “type of” (A type of jacket: nehru) • “starter for” (Starter for saxon: anglo) • “suffix with” (Suffix with switch or sock: eroo
Implicit Distribution Modules • Some targets not included in any database, but more probable than random • Schaeffer vs. srhffeeca • Bigram module • Generates all possible letter sequences of the given length by returning a letter bigram distribution over all possible strings, learned from CWDB • Lowest probability clue-target, but higher probability than random sequence of letters • Honolulu wear: hawaiianmuumuu • How could this be used for Scrabble?
Questions • Is this approach any more intelligent than the chess playing programs? • Does the use of knowledge correspond to intelligence? • Do any of the techniques for generating words apply to Scrabble?