Modeling Semantic Incongruity for Computational Humor

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 292 papers119 conceptsfrozen 2026-10-04

Humor is often described as the "final frontier" of artificial intelligence. While modern Large Language Models (LLMs) can translate languages, solve math problems, and write code, they frequently fail at the simple task of being funny. As Kristian Kersting and Sophie Jentzsch argue in ChatGPT is fun, but it is not funny!, these models often act as "high-fidelity parrots," reciting a "greatest hits" reel of memorized puns rather than generating original wit.

To move beyond memorization, researchers are focusing on the core mechanism of humor: semantic incongruity.

The Mechanics of the Twist

At the heart of most computational humor research is the Semantic-Script Theory of Humor (SSTH). This theory suggests that a joke works by leading the listener down one logical path (a "script") and then suddenly pivoting to a second, incompatible script. The "funny" part is the cognitive resolution of this conflict.

Igor Labutov and Hod Lipson operationalized this in Humor as Circuits in Semantic Networks by treating jokes as "circuits" in a knowledge graph. A joke setup establishes a path in the network, and the punchline forces the reader to jump to a distant node that nonetheless shares a surprising connection. Similarly, Hao Yang and colleagues introduced Sentence Discrepancy Prediction (SDP) in Unified Humor Detection Based on Sentence-pair Augmentation and Transfer Learning to explicitly measure the "semantic distance" or "twist" between a setup and its conclusion.

The Hierarchy of Difficulty

Not all humor is created equal. Research reveals a clear hierarchy of difficulty for AI:

  1. Phonological Puns: These are the "low-hanging fruit." Systems like HAHAcronym, developed by Oliviero Stock and Carlo Strapparava in HAHAcronym: A Computational Humor System, succeed by manipulating acronyms and wordplay. Because puns rely on phonetic similarity (homophones) or spelling (homographs), they can be modeled using formal linguistic rules and dictionaries, as shown by Kim Binsted and Graeme Ritchie in their Implemented Model of Punning Riddles.
  2. Situational and Topical Humor: This requires "esoteric" real-world knowledge. Tyler Loakman and others found in Comparing Apples to Oranges that while LLMs handle wordplay well, they struggle with jokes that require social reasoning or current events.
  3. Satire and Surrealism: This is the "AI-complete" level. Benchmarks like The New Yorker Caption Contest, introduced by Jack Hessel and Lillian Lee, show that even GPT-4 struggles to connect a visual anomaly (e.g., a giant toaster in a desert) to a witty remark. This requires "reading the room"—a feat of multimodal social commonsense.

From Detection to Explanation

Early research treated humor as a binary classification task: "Is this funny or not?" However, Dongyu Zhang and colleagues argued in Telling the Whole Story that we must ask why it is funny.

This shift has led to "explainable humor" benchmarks. For instance, the Chumor dataset, sourced from the intellectually challenging Ruo Zhi Ba (Weak Brain Forum) community by Ruiqi He and others in Chumor 2.0, tests whether models can identify the specific logical subversion in a joke. Interestingly, Maxime Peyrard discovered in Laughing Heads that Transformer models sometimes spontaneously develop a "laughing head"—a single attention head that learns to pinpoint the exact word that triggers the humor, even without explicit training.

The Hallucination Trap

A major hurdle in this field is the "hallucination trap." In PHUNNY, Alessio Cocchieri and colleagues found that when models are presented with "fake" puns—sentences that look like jokes but make no sense—they often invent elaborate, illogical justifications for why they are funny. This suggests that models are often "over-thinking" humor, applying statistical patterns of "explanation-like" text rather than truly grasping the underlying semantic conflict.

To combat this, researchers like Yuyan Chen are developing frameworks like TalkFunny, which uses "Chain-of-Humor" interpretations and "Humor Mind Maps" to guide models through the logical steps of a joke, forcing them to navigate the gap between conventional expectations and witty subversions.

Papers in this version

Showing 60 of 92 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 58 papers
  2. Version 1 · 2026-10-04 · 92 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_computational_humor_complexity_831c77b7,
  title = {computational humor complexity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/computational-humor-complexity/rdr_831c77b77f514ce4aa0c9e28eedb0c11},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_modeling_semantic_incongruity_for_comput_831c77b7_v1,
  title = {Modeling Semantic Incongruity for Computational Humor},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/modeling-semantic-incongruity-for-computational-humor/rdr_831c77b77f514ce4aa0c9e28eedb0c11/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}