Solving Crossword Puzzles with AI: - PowerPoint PPT Presentation

1 / 10
About This Presentation
Title:

Solving Crossword Puzzles with AI:

Description:

Proverb was developed in 1999 to solve crosswords puzzles. Works on American puzzles ... 0.010101 ray. Training the modules... – PowerPoint PPT presentation

Number of Views:470
Avg rating:5.0/5.0
Slides: 11
Provided by: rachel47
Category:

less

Transcript and Presenter's Notes

Title: Solving Crossword Puzzles with AI:


1
Solving Crossword Puzzles with AI
  • a look at Proverb

2
What is Proverb?
  • Proverb was developed in 1999 to solve crosswords
    puzzles
  • Works on American puzzles
  • Cannot beat expert humans
  • Solves a puzzle in 15 minutes
  • Gets over 93 of words correct

3
Goal
  • Maximize the number of answers in the crossword
    puzzle that are the same as the solution.
  • So how does it work?

4
Searching
  • Different searches performed depending on the
    category abbr, synonym, kind of, pop culture,
    geography, literature, film
  • Two stage architecture
  • 1. Specific modules that generate candidate
    answers
  • 2. Combines results from the modules

5
Kinds of Modules
  • Database movie, music, geography, literary,
    synonyms, etc
  • Syntactic fill-in-the-blanks, kind of
  • Word list
  • CWDB specific
  • Information retrieval encyclopedia, partial
    match, etc

6
More about modules
  • modules are given the clue and the number of
    letters in the target
  • grid constraints are ignored at this point with
    the exception of word length
  • the module returns anywhere between 0 and 10,000
    possible answers
  • each one has a weighted likelihood or probability
    that it is the correct solution
  • each module also returns a value that represents
    its confidence that the answer is part of its list

7
  • Clue Farrow of Peyton Place (answer Mia)
  • Confidence score 1.0
  • 0.909091 mia
  • 0.010101 tom
  • 0.010101 kip
  • 0.010101 kip
  • 0.010101 peg
  • 0.010101 ray

8
Training the modules
  • 30 modules were evaluated with test data that
    consisted of 5374 clues
  • Measures of performance
  • How often the correct answer was included in the
    candidate list
  • The average length of the candidate list
  • The number of times that the correct answer
    appeared as the 1 candidate
  • Percentage of clues that the module guessed at

9
(No Transcript)
10
Grid Filling
  • Probabilistic Constraint Satisfaction Problem
  • Each box is represented as a variable
  • Must maximize the correct solutions
  • Limitations
  • confused by creativity in clues
  • only solves American puzzles
Write a Comment
User Comments (0)
About PowerShow.com