Humor, Computer-Generated

There are two main motivations for computational humor. On the one hand, for the sake of humans who interact with them, computers should have active (generating) and passive (detecting and understanding) humor abilities. The underlying assumption is that this will make human-computer interaction more like human-human interaction, which is presumably more enjoyable and thus productive. Here, the main purposes of humor range from improved motivation, increased focus on tasks, and higher user acceptance to improved sales.

For the sake of humor research, on the other hand, putting theories about humor into action by giving their rules and resources to a computer is a very good test of those theories. If the program identifies the same texts as humorous as humans do, or if it generates text that humans identify as humorous, the rules and theories the program is based on are appropriate for humor. This entry discusses those computer programs that have been written for humor generation and how they have evolved from templates with a limited ability to make choices to natural-language processing systems.

It is easier to use computer programs for humor generation than for humor detection, because generation can be controlled more tightly. The rules and resources given to the computer determine the output, possibly in reaction to some additional user input: simple ones like word lists and templates or, more recently, complex ones like dictionaries and knowledge bases. With the help of these resources and sets of rules (algorithms), a program then produces texts intended to be humorous. In contrast to this, the analysis of humor in free-range text necessarily has to involve much larger and complex resources, because it must be able to handle many kinds, or even any kind of humorous text. Because of this, analysis is restricted to simple research questions on preidentified humorous texts or to mere classification of texts as potentially humorous or not, usually using statistical and machine-learning approaches.

The task of humor generation has not used statistical methods in any central function but has worked with knowledge bases, so far mostly simple ones that produced rather clumsy humorous texts that usually have low levels of funniness in any context but are nevertheless applications of artificial intelligence.

The evolution of automatic humor generation over the past 20 years has been determined by the improvement in the quality and quantity of the underlying resources and the artificial intelligence powering the generation. Currently, the third generation of systems, based on text-meaning and aiming for text-humor understanding, is under development.