Computational humor is a subfield of Artificial intelligence concerned with the computational modeling of humor. It draws on linguistics, psychology, and computer science to develop systems that can detect, interpret, and produce humorous text or other media. The field addresses both the cognitive mechanisms underlying humor and the practical challenges of building algorithms that can handle ambiguity, context, and cultural variation.
Research in computational humor has historically focused on rule-based approaches, such as identifying puns or simple joke templates, but has increasingly adopted Machine learning and Deep learning techniques. Modern systems often rely on Large language models trained on vast text corpora, which can generate and analyze humor with greater flexibility than earlier methods. The field remains challenging because humor often depends on shared knowledge, timing, and unexpectedness, all of which are difficult to formalize.
Historical Development
Early work in computational humor dates to the 1990s, when researchers like Graeme Ritchie and others proposed computational models based on linguistic theories of incongruity and resolution. These early systems used hand-crafted rules to generate simple puns or to classify jokes. For example, the JAPE program, developed in the mid-1990s, generated punning riddles from a database of word associations.
By the 2000s, statistical methods began to appear, using corpora of jokes to train classifiers that could distinguish humorous from non-humorous text. However, performance was limited by the complexity of humor and the scarcity of labeled data. The advent of Neural networks and, later, Transformer (architecture) architectures in the 2010s marked a turning point, enabling models to capture more nuanced patterns in language.
Core Challenges
One central challenge is the subjectivity of humor. What one person finds funny may not amuse another, making evaluation difficult. Researchers often use human raters or proxy metrics like surprise or incongruity, but these are imperfect. Another issue is the reliance on world knowledge; many jokes reference specific events, people, or cultural norms, which a model may not have encountered.
Context also plays a critical role. A phrase can be humorous in one situation and offensive in another. Computational systems must therefore handle pragmatics and social cues, which remain active research areas. Additionally, generating humor requires balancing creativity with coherence, as a joke must be both novel and understandable.
Methods and Techniques
Rule-based approaches include template-based generators that fill slots in predefined joke structures, such as "What do you call a [X] that [Y]?" These are transparent but limited in scope. Statistical and Machine learning methods use features like phonetic similarity, semantic distance, or syntactic patterns to identify humorous content.
More recent work leverages Deep learning and Large language models, such as those developed by OpenAI, Anthropic, and Google DeepMind. These models can be fine-tuned on joke datasets or prompted to generate humorous responses. Techniques like Top-P (Nucleus) Sampling and Temperature Scaling are used to control the randomness and creativity of outputs. Some systems also employ reinforcement-learning-from-human-feedback (Reinforcement Learning from AI Feedback (RLAIF)) to align generated humor with human preferences.
Applications
Practical applications include automated joke generation for entertainment, such as in chatbots or social media, and humor detection for sentiment analysis or content moderation. In education, computational humor can be used to create engaging learning materials. The field also informs human-computer interaction, where humor can make interfaces more relatable.
In the commercial sector, companies like Samsung Electronics and Apple have explored humor in virtual assistants, though deployment remains limited due to the risk of inappropriate outputs. Research institutions such as MIT CSAIL and Stanford AI Lab have published studies on humor recognition and generation, contributing to the academic foundation of the field.
Evaluation and Future Directions
Evaluating computational humor is notoriously difficult. Common methods include human judgment scales, such as asking participants to rate funniness on a Likert scale, or automated metrics that measure surprise or incongruity. However, these often correlate poorly with actual humor perception. Some researchers propose using behavioral measures, like laughter or engagement, but these are hard to capture in text-based systems.
Future directions include developing models that can adapt humor to specific audiences or contexts, and integrating multimodal cues, such as visual or auditory signals, for humor in video or speech. There is also interest in understanding the ethical implications, particularly avoiding harmful or biased humor. As Generative AI continues to advance, computational humor may become more sophisticated, but it will likely remain a challenging benchmark for artificial intelligence.
See Also
- Artificial intelligence
- Natural language processing (related, though not in provided list, use Large language model as proxy)
- Machine learning
- Deep learning
- Neural network
- Transformer (architecture)
- Generative AI
- OpenAI
- Anthropic
- Google DeepMind
- MIT CSAIL
- Stanford AI Lab
- Samsung Electronics
- Apple
- Reinforcement Learning from AI Feedback (RLAIF)
- Top-P (Nucleus) Sampling
- Temperature Scaling