Word Play Generation

1
Word Play Generation Vivaek Shivakumar Computer
Systems Lab 2010
2
Background and Purpose

• Computational humor applications
• Human-computer interaction
• Language/educational aid
• General amusement
• No formal model/theories of humor or puns exist
• Not much work done in sophisticated generation of
  palindromes, anagrams, etc.

3
Background and Purpose

• Projects that have implemented puns
• JAPE, STANDUP (Ritchie et al., 2007)
• HAHAcronym
• Humor recognition
• Purpose to investigate methods for word-play
  generation of various types

4
Development

• Punning riddle models based on
• Schemata define word relationships
• Templates produce surface joke/riddle
• (Binsted Ritchie, 1994)
• (shown) What do you get when you cross ____ and
  ____?

Figure 1
5
Development

• NLTK
• WordNet semantic web
• CMU pronunciation dictionary
• Sets of words produced to fit schema
• e.g, top -gt peak -gt peek -gt peep
• e.g help--gtservice --gt nervous--gtuneasy
• Both homophones and near homophones

6
Development and Experiment

• Where to start for search?
• Problems
• WordNet and CMU Dic.
• Obscure words, British/American variants, proper
  nouns

7
Development

• Palindromes
• Stack to store the words picked forming the front
  of the palindrome
• Segmentation algorithm for the tail
• e.g., fine, position . noitisopenif -gt
  no it is open if
• Recursive backtracking

8
Development

• Acronyms
• Letters of the input make acronym
• Meaning of the input seeds the words to fill
  acronym
• E.g., input dog -gt D____ O____ G____
• All three words related to or associated with
  dog, possibly some filler words

9
Development

• For output to make sense
• Grammar filter
• E.g. for noun phrases
• (determiner) (adjectives) (nouns) head-noun
  OR pronoun
• postmodification
• Implementation (d?anp)(r(d?an)g)?

10
Development

• Anagrams
• Three methods 2nd/3rd use preliminary pruning
  of dictionary -gt n candidate words
• Brute force inefficient can produce all
  combos
• Better brute force less inefficient ()
• Hill Climbing Random restart produces 1 combo
• Solution checker by matrix reduction
• Each state is a solution vector of word presence
  (0 or 1)

26
26
n
n
11
Results

• Punning Riddles
• rabbit-gtconey-gtphony-gtdissimulator.
• no, just no
• rabbit-gthare-gtfair-gthonest
• perhaps
• E.g. What do you call an honest rabbit? A fair
  hare.
• Why do some combinations work and others dont?

12
Results

• Palindromes - X indicates passed the noun phrase
  filter
• X race car
• X level level
• X liar rail
• X rebuttal at tuber
• no ill im million
• X red art trader
• X test set
• bore rob
• X spilt lips
• X battalion oil at tab
• oh whose hall action no it call a hes oh who

13
Results

• Acronyms
• FRUIT ___ ___ ___ ___ ___
• INFINITE Innumerable Non-finite ___ ___
  Numberless ___ ____ ____
• ORDER Orderliness Rules Decree Edict Rescript
• BAD Below Average Decency
• STUPID Stunned ___ Unintelligent Person ___
  Dolt
• BUSH Born Under Sophistication Herbert
• CIA Collecting Independent Activities
• CIA Collecting Intelligence Abroad
• WORD Writings of Restricted Discussion
• Various degrees of success

14
Results

• Anagrams
• Brute force, e.g.
• enter word -gt 'red', 'went', 'or', 'red',
  'or', 'went', 'or', 'went', 'red', 'or',
  'red', 'went', 'went', 'or', 'red', 'went',
  'red', 'or', 'went', 'order', 'order',
  'went'
• Hill climbing examples
• enter word -gt order went
• multi vites -gt it must live
• computer systems -gt curt my mets posse (?)

15
Results

• Anagram algorithm efficiency (for finding one
  solution)

16
Conclusion

• Barrier of sophistication
• Several resource-based problems
• Efficiency
• Possibilities for computer-generated word play