Modeling Subjectivity in Humor and Offense Detection

31 concepts18 papers1 proposal

Traditional natural language processing tasks often operate under a comfortable assumption: there is a single, objective "ground truth" label for every piece of text. If you are classifying whether an email is spam or identifying whether a sentence contains a medical symptom, human annotators can usually reach a consensus.

However, when language models confront humor, irony, and offense, this assumption breaks down entirely. What makes one person laugh might make another log off in disgust. Humor and offense are not stable properties of text; they are dynamic, highly subjective experiences that exist at the intersection of language, culture, and individual identity.

To build systems that can truly navigate human social spaces, researchers are shifting away from rigid binary classifications. Instead, they are developing frameworks that embrace subjectivity, quantify controversy, and model language through multiple perspectives.

The Illusion of a Single Ground Truth

When automated systems evaluate subjective language, they encounter what researchers call a "moving target." As Zhengyi Guan et al. observe in their work on Tsia at SemEval-2021 Task 7: Detecting and Rating Humor and Offense, establishing an objective ground truth for training models is fundamentally difficult because individual interpretations vary wildly based on personal backgrounds and cultural contexts.

If a model is forced to learn from data where human disagreement has been averaged out or erased, it fails to capture the actual mechanics of communication. Yang Bai points out in hub at SemEval-2021 Task 7: Fusion of ALBERT and Word Frequency Information Detecting and Rating Humor and Offense that this context-dependent variability represents a fundamental limitation for traditional NLP architectures. When a model treats a highly controversial joke as simply "70% funny," it misses the point: the joke might be 100% funny to one demographic and 100% offensive to another.

The Entanglement of Humor and Offense

Humor and offense are not independent concepts; they are deeply entangled. A major catalyst for exploring this relationship was the SemEval-2021 HaHackathon shared task, introduced by J. A. Meaney et al. in SemEval 2021 Task 7: HaHackathon, Detecting and Rating Humor and Offense. This task forced models to jointly predict both the presence of humor and the intensity of offense, revealing that "humor controversy" is a distinct linguistic phenomenon.

To capture these nuances, models must move beyond simple keyword matching. For instance, Yingjia Zhao et al. developed an ALBERT-based architecture in ZYJ at SemEval-2021 Task 7: HaHackathon: Detecting and Rating Humor and Offense with ALBERT-Based Model specifically designed to quantify these subjective qualities along a continuous scale rather than a binary switch.

Interestingly, the scale and type of model matter depending on the task's subtlety. In RedwoodNLP at SemEval-2021 Task 7: Ensembled Pretrained and Lightweight Models for Humor Detection, Nathan Chi et al. discovered that while massive pre-trained language models excel at identifying intended humor, lighter, more targeted models are often more effective at predicting nuance like humor controversy and offensiveness. This suggests that detecting a joke requires broad semantic knowledge, but predicting whether that joke will cause offense requires a tighter grasp on specific social friction points.

Why Context is Paramount

Why do standard language models struggle so acutely here? Because humor and offense rely heavily on pragmatics—the study of how context contributes to meaning.

Traditional NLP:  [Text] ───────────────────────────► [Single Label]

Perspectivist NLP: [Text] + [Cultural/Social Context] ◄──► [Distribution of Interpretations]

This challenge is magnified in real-world social media text, which is rarely clean or monolingual. Consider the following dimensions of context that researchers are actively trying to model:

From Consensus to Perspectivism

To resolve these challenges, the research community is undergoing a paradigm shift: moving away from forcing a single consensus label and instead explicitly modeling the diversity of human perspectives. This is known as perspectivism.

A prime example of this direction is the work by Silvia Casola et al. on MultiPICo: Multilingual Perspectivist Irony Corpus. Instead of stripping away annotator disagreement to find an average "truth," the authors preserved the rich sociodemographic metadata (such as age, gender, and culture) of the annotators. This allows models to learn how and why irony perception shifts across different social groups.

Similarly, Liesbeth Allein et al. developed the OrigamIM: A Dataset of Ambiguous Sentence Interpretations for Social Grounding and Implicit Language Understanding dataset to capture implicit moral judgments and reader attitudes. They demonstrate that annotator disagreement is not noise to be filtered out; it is valuable data that reflects real-world societal divisions.

Even the way models are trained is adapting to these gray areas. For instance, Han Wang et al. designed a self-training framework in Multi-Agent VLMs Guided Self-Training with PNU Loss for Low-Resource Offensive Content Detection that simulates a negotiation between two AI agents—a strict moderator and a lenient user. By framing offensive content detection as a dialogue between opposing viewpoints rather than a rigid binary classification, the system can better navigate the ambiguous boundaries of human language.

By treating subjectivity as a core feature rather than an error, this research direction is paving the way for AI that doesn't just read words, but understands the social fabric of the people speaking them.

Key Foundational Work

Tsia at SemEval-2021 Task 7: Detecting and Rating Humor and Offense (2021) — contributed 5 concept elements to this direction.
hub at SemEval-2021 Task 7: Fusion of ALBERT and Word Frequency Information Detecting and Rating Humor and Offense (2021) — contributed 4 concept elements to this direction.
Simon @ DravidianLangTech-EACL2021: Meme Classification for Tamil with BERT (2021) — contributed 3 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_subjectivity_of_humor_subjectivity_of_hu_d9d109fe,
  title = {subjectivity of humor (subjectivity of humor and offense)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/subjectivity-of-humor-subjectivity-of-humor-and-offense/rdr_d9d109fedc1746828026a482c6821a68},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_subjectivity_of_humor_subjectivity_of_hu_d9d109fe_v1,
  title = {subjectivity of humor (subjectivity of humor and offense)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/subjectivity-of-humor-subjectivity-of-humor-and-offense/rdr_d9d109fedc1746828026a482c6821a68/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake