Use of Statistical Language Recognition in Computational Humor

1
Use of Statistical Language Recognition in
Computational Humor

• Julia Taylor (Julia.Taylor_at_ca.com)
• Applied Artificial Intelligence Laboratory
• ECECS Department
• University of Cincinnati
• Cincinnati, Ohio 45221-0039

2
What Is a Joke

• Hetzron A joke is a short humorous piece of oral
  literature in which the funniness culminates in
  the final sentence.
• Punch line -- the final sentence
• Setup -- the text before punch line

3
The structure of a Joke

• Koestler
• a joke calls for the presence of two conflicting
  logics within one story line and the punch line
  is the point of collision between these two
  trains of thought.
• Suls two-stage model
• punch line creates incongruity
• cognitive rule enables the content of the punch
  line to follow naturally from the information
  established in the setup

4
What creates a joke

• Some of the mechanisms for a joke creation
• surprise
• ambiguity
• contradiction
• inconsistency
• disparagement

5
Structure of punch line

• straight-line jokes
• A woman goes to the rabbi Rabbi, what shall I
  do so that I wouldnt become pregnant again?
  The rabbi says Drink a glass of cold water!
  Before of after? The rabbi replies Instead!
• dual jokes
• The Parisian Little Moritz is asked in school
  How many deciliters are there in a liter of
  milk? He replies One deciliter of milk and
  nine deciliters of water In France, this is a
  good joke in Hungary, this is good milk.
• rhythmic jokes
• A newspaper reporter goes around the world with
  his investigation. He stops people on the street
  and ask them Excuse me Sir, what is your
  opinion of the meat shortage? The American asks
  What is shortage? The Russian asks What is
  an opinion? The Pole asks What is meat?
  The New York taxi-driver asks Whats excuse
  me?

6
Suls algorithm

• As text is read, make prediction
• While no conflict with prediction, keep going
• If input conflicts with prediction
• if not ending PUZZLEMENT
• if it is the ending, try to resolve
• no rule found PUZZLEMENT
• cognitive rule found HUMOR

7
Statistical recognition of text

• Analysis of a text linked decisions
• Decisions are based on what we know
• Is it raining
• It is raining outside.
• It is raining in the fall.
• It is raining in the desert.
• It is raining in the house.

8
N-gram model

• Uses n-1 previous words to predict the next one
• each string is assigned the probability in
  relation to all other strings of the same length.

9
N-gram model

• Takes into account where a sentence ends

• A newspaper reporter goes around the world with
  his investigation. He stops people on the street
  and ask them
•  

• A newspaper reporter goes around the world with
  his investigation end-of-sentence he stops people
  on the street and ask them
•  

10
N-gram model training corpus
P(rabbi the) ?

• A woman goes to the rabbi Rabbi, what shall I
  do so that I wouldnt become pregnant again?
  The rabbi says Drink a glass of cold water!
  Before of after? The rabbi replies Instead!

• A newspaper reporter goes around the world with
  his investigation. He stops people on the street
  and ask them Excuse me Sir, what is your
  opinion of the meat shortage? The American asks
  What is shortage? The Russian asks What is
  an opinion? The Pole asks What is meat?
  The New York taxi-driver asks Whats excuse
  me?

11
N-gram value of n

• P(world the) lt P( world around the)
• A newspaper reporter goes around the world with
  his investigation. He stops people on the street
  and ask them Excuse me Sir, what is your
  opinion of the meat shortage? The American asks
  What is shortage? The Russian asks What is
  an opinion? The Pole asks What is meat?
  The New York taxi-driver asks Whats excuse
  me?

12
N-grams and Jokes

• As joke is read, make prediction of the next word
  using N-gram model
• Once punch line is reached (last sentence of
  text), make prediction using regular corpus. At
  some point, probabilities should differ
  significantly, detecting conflict.

13
What has to be addressed

• What n is most accurate?
• Training corpus
• What probability is high enough to match
  prediction?
• What probability is low enough for a conflict?