Presented at AAAI-94, the JAPE-1 system represents a foundational milestone in computational creativity. Moving beyond simple "canned" humor, the research explores whether a machine can generate original jokes using only formal linguistic rules and a standard, humor-independent dictionary.
The Mechanics of Humor: Humor as Structure
Historically, AI research struggled to model humor due to its perceived subjectivity. Binsted and Ritchie addressed this by focusing on a tractable subset: the punning riddle (e.g., "What do you use to flatten a ghost? A spirit level."). They argued that puns are highly regular and can be treated as structural collisions between two semantic paths.
The JAPE-1 model relies on two primary components:
- Schemata: Logical blueprints that define the relationships between lexemes (words and meanings). A schema might, for instance, look for a word that is a homonym of a part of a common phrase.
- Templates: Surface-level "skins" that convert logical structures into readable question-answer pairs (e.g., "What do you call a [Characteristic] [Noun]?").
A critical part of this process is the creation of a "Fake Meaning"—a temporary semantic bridge that justifies the pun. The system doesn't just find wordplay; it constructs a nonsense concept that links the setup to the punchline's double meaning.
The Generative Process
To ensure the AI wasn't "cheating," the authors used a lexicon created by volunteers unfamiliar with the project. This ensured the data was "humor-independent," proving that the resulting humor derived from the system's mechanics rather than pre-loaded jokes.
The workflow follows a specific sequence:
- Lexical Search: The system scans its database for words and properties (e.g., "is green," "is a clothing item").
- Homonym Identification: It locates words that sound identical but have different meanings.
- Instantiation: The system plugs words into a schema that fits the required relationships.
- Template Selection: Finally, it selects a template to output the text.
As seen in the example above, the system maps the properties of "sheep" and the action "leap" onto the components of the noun phrase "woolly jumper" to create a semantic collision.
Results and Evaluation
JAPE-1 generated 188 jokes, which were evaluated by 14 human judges on a scale of 0 to 5. While the average score was a modest 1.5/5, this was expected for the "groan-worthy" genre of children's riddles. Some generated jokes were indistinguishable from human-written ones, such as:
- "What kind of tree can you wear? A fir coat."
- "What do you call a murderer that has fibre? A cereal killer."
- "What kind of rain brings presents? A bridal shower."
The evaluation also highlighted the extreme subjectivity of humor; jokes that one judge rated 5/5 were often rated 0/5 by another.
Limitations and Future Directions
The authors identified several paths for improvement. The current lexicon lacked the depth needed for more sophisticated setups (e.g., knowing what an object is used for). Additionally, they proposed adding "post-production heuristics" to help the system self-evaluate and discard poor-quality jokes. Future work could also extend the model to handle more complex wordplay, such as spoonerisms.
Summary updated