This is a Common Lisp code walkthrough for generating original jokes. You seed the generator with the knowledge about different objects, and it uses that vocabulary to generate unique jokes. All of the jokes are of the form: “What do you get when you cross X with Y?” This code was originally written for my CS288: An AI Approach to Natural Language Processing class at UC Berkeley in 2004.

The Setup

The joke generator is written in Common Lisp. To compile the code and generate your own jokes, you will need to get a compiler. Download and install the trial version of Allegro CL. Next download the complete jokes.lisp code file.

Running the Joke Generator

(generate) will cycle through the vocabulary and try to make a joke for each set of words. If it finds what it thinks is a good joke, it will print it. It is recommended that joke.lisp be compiled as (generate) is computationally intensive.

(answer-joke word1 word2 mod1 mod2) will try to answer the joke “What do you get when you cross <mod1> <word1> with <mod2> <word2>?” mod1 and/or mod2 may be nil. In order for this to work, *vocab* must be initialized with (seed-knowledge). (generate) does this automatically.

Yeah, but how does it make the jokes?

This is the job of (answer-joke).

It uses transformations of the input words to come up with an answer. Each word in the vocabulary contains information about relations with other words, homophones, and part of speech. The transformations it uses are: substitution of a related word, concatenation to form known compound words or phrases, substitution of a homophone, substring manipulation, and adding common suffixes. Each joke may use one or more of these transformations.

When a joke is found, it is given a rating between 1 and 10, depending on the transformations that were used to create the joke. Jokes can loose points if a similar joke has already been printed, or if the same words occur in the question and the answer. If the final rating is greater than the joke threshold, the joke is printed.

What about the Vocabulary?

The vocabulary is created in (seed-knowledge).

The vocabulary only contains about 200 words. Due to the information needed about each word, the vocabulary had to be created by hand, instead of using a dictionary. The *vocab* hashtable stores word-prop structs keyed by the vocabulary word. The word-prop structs contain part of speech, a relations list, a homophones list, and information about article usage and animation.

Part of speech values can be:
‘n (noun)
‘m (modifier)
‘b (both noun and modifier)
‘x (none)

The none tag is used for words that aren’t really nouns or modifiers, but can potentially make up an answer. For example, “mew” (the noise a cat makes) has POS ‘x.

Relations are simply the declaration that one word is related to another somehow. Two words in a relation could be synonyms (rabbit and hare), “is a” relations (aunt and relative), or important characteristics (cheetah and fast). No information is stored about what kind of relation it is. This is an area for improvement.

The homophones list stores word-prop structs that sound like the word. For example, word-props for “hair” and “hare.”

Article usage (art) values can be either ‘t (true) or ‘f (false).

Each word can also be animate, inanimate, or both. Animate nouns are people, animals, etc. Animate modifiers are modifiers that can be applied to animate nouns. This distinction prevents the use of phrases like “excited rain” while allowing “excited dog.”

To create the vocabulary, I first started by adding words, relations, and homophones that create known jokes. After testing that the generator could produce jokes if given this information, I added additional words that are related to known words, homophones of known words, and compound words containing known words.

To create your own original jokes, replace the vocabulary in (seed-knowledge) with your own relations, homophones, and words.

Improvements

There are a few areas that could be improved.

1. Expand the scope of relationships between words with semantic knowledge. As mentioned earlier, it would be useful to know that “sour rain” is not a good phrase, but “sour lemon” is. It would also be useful to make distinctions between the different types of relations that can occur between words, and to be able to express the idea that the combination of two words can result in another word or phrase based on semantic information.

2. More complete phonetic information. Some transformations rely on substitutions made by partial homophones. For example “mew” (the noise a cat makes) is phonetically similar to the “mu” in music, but it different from the “mu” in murderer. There is currently no way to express this difference.

3. More complete vocabulary. The vocabulary is somewhat contrived. A more telling test of the joke generator would be to import the vocabulary from an unbiased source. A more complex part of speech identifier could also be implemented.

So what jokes did it make?

Here is a list of a few of the jokes that the generator came up with.

WHAT DO YOU GET WHEN YOU CROSS A PORT WITH FROSTED FLAKES?
I don’t know, but it’s serial

WHAT DO YOU GET WHEN YOU CROSS A DANCE WITH A LEMON?
sour balls

WHAT DO YOU GET WHEN YOU CROSS A DANCE WITH A CHEETAH?
fastball

WHAT DO YOU GET WHEN YOU CROSS A POPPY WITH ELECTRICITY?
flower power

WHAT DO YOU GET WHEN YOU CROSS AN ALIEN WITH A CHICKEN?
eggs-traterrestrial

WHAT DO YOU GET WHEN YOU CROSS A SERIOUS THIEF WITH A WET JACKET?
grave robber with raincoat

Tags: jokes, lisp

Related Sweet Hacks - Vol I Jess Johnson’s Programming Project Portfolio About The Top 9½ In a Hacker’s Bookshelf

Like this? Subscribe! Subscribe to the feed Or subscribe via email: