The Divergence of Intended and Perceived Sarcasm

20 concepts11 papers1 proposal

Sarcasm is a linguistic flip-flop: the literal words say one thing, but the intended meaning screams the opposite. For years, natural language processing treated sarcasm detection as a standard text classification problem. To build datasets, researchers relied on two primary sources of "ground truth": crowdsourced third-party annotators or automated distant supervision via hashtags like #sarcasm.

However, this approach rests on a flawed assumption: that sarcasm is an objective property of text visible to any observer. As Aniruddha Ghosh et al. point out in Magnets for Sarcasm: Making Sarcasm Detection Timely, Contextual and Very Personal, external observers lack insight into a speaker's true psychological state and immediate mood. Without this internal context, third-party labeling introduces significant noise. This creates a fundamental disconnect: models trained on crowdsourced data learn to identify what looks sarcastic to a stranger, rather than capturing the actual communicative intent of the creator (Debanjan Ghosh et al. in Sarcasm Analysis Using Conversation Context).

Quantifying the Perception Gap

What happens when we stop guessing and ask authors themselves what they meant? To answer this, Silviu Oprea et al. created iSarcasm: A Dataset of Intended Sarcasm, a dataset where labels were provided directly by the text's creators.

The results exposed a massive "perception gap." State-of-the-art models that performed exceptionally well on traditional datasets failed dramatically when tasked with identifying true intended sarcasm. The authors demonstrated that external observers frequently miss subtle, genuine irony or incorrectly attribute sarcasm to entirely literal statements.

As Ibrahim Abu Farha et al. later showed in Sarcasm Detection is Way Too Easy! An Empirical Comparison of Human and Machine Sarcasm Detection, traditional datasets make the task look deceptively simple because they rely on explicit, superficial markers. When ground truth is anchored to the author's actual intent, the problem becomes significantly harder. This reveals that traditional training data degrades downstream performance by optimizing for the wrong signal (Ibrahim Abu Farha et al. in SemEval-2022 Task 6: iSarcasmEval, Intended Sarcasm Detection in English and Arabic).

Two Distinct Linguistic Phenomena

The divergence between intent and perception is not just a matter of label noise; they are two entirely different linguistic phenomena that require different modeling approaches.

In Exploring Author Context for Detecting Intended vs Perceived Sarcasm, Silviu Oprea et al. discovered that an author's historical social media activity is a powerful predictor of intended sarcasm. Because an individual's ironic style is grounded in their personal history and behavioral patterns, models that incorporate these user-specific features can learn to identify their intent. However, these author-specific features do not translate to general human perception. An external reader lacks access to the author's history, meaning audience perception is driven by entirely different, highly subjective factors.

This structural difference deeply affects predictability. Joan Plepi et al. utilized Graph Attention Networks in Perceived and Intended Sarcasm Detection with Graph Attention Networks to model the relationships between authors and their conversational neighborhoods. They found a distinct gap in predictability: intended sarcasm is significantly easier for a model to identify when given access to the author's historical context, whereas perceived sarcasm remains highly volatile because it depends on the audience's interpretation.

In fact, the two phenomena have completely different linguistic signatures. By listening to online conversations where users explicitly clarify their intent (e.g., "I was being sarcastic"), Boaz Shmueli et al. built the SPIRS dataset in Reactive Supervision: A New Method for Collecting Sarcasm Data. Their work confirmed that what an author intends and how an oblivious audience responds follow distinct textual patterns.

The Limits of Contextual Cues

A natural counterargument is that the perception gap can be bridged by providing more context, such as images or dialogue threads. If the audience can see what the author sees, won't perception align with intent?

The evidence suggests otherwise. In Detecting Sarcasm in Multimodal Social Platforms, Rossano Schifanella et al. fused visual and textual data from platforms like Instagram. Even when full multimodal context was available, a substantial discrepancy persisted between the creator's intent and the audience's interpretation. Visuals help resolve blatant contradictions, but they cannot fully expose a creator's internal motivation.

Because capturing true internal motivation at scale is incredibly difficult, practical systems often have to choose which side of the gap to model. For instance, Trisha Mittal et al. developed the "Intent-o-meter" in Towards determining perceived audience intent for multimodal social media posts using the theory of reasoned action. Instead of claiming to read the creator's mind, the authors used psychological frameworks to predict the perceived audience intent, acknowledging that what the audience extracts is often the only measurable signal available without direct author feedback. When manual labeling is forced to rely on annotator perception rather than true intent, the resulting models will inherently reflect the audience's subjective biases rather than the author's reality (Aparna K Ajayan et al. in Amrita_CEN at SemEval-2022 Task 6: A Machine Learning Approach for Detecting Intended Sarcasm using Oversampling).

Designing for the Right Perspective

This leaves the field of sentiment analysis with a critical architectural choice. As posed by Ibrahim Abu Farha et al. in Sarcasm Detection is Way Too Easy! An Empirical Comparison of Human and Machine Sarcasm Detection, should computational systems prioritize identifying the speaker's internal intent or the audience's likely interpretation?

The answer depends entirely on the application:

  • If the goal is personalized AI or user modeling, systems must treat sarcasm as an act of intent, requiring author-grounded datasets and historical behavioral context.
  • If the goal is content moderation or public trend analysis, systems must model perception, mapping how a general audience is likely to interpret a post regardless of what the author originally meant.

By recognizing that intended and perceived sarcasm are separate, diverging tracks, researchers can move away from noisy, monolithic datasets and build models tailored to the specific perspective they need to capture.

Key Foundational Work

Detecting Sarcasm in Multimodal Social Platforms (2016) — contributed 3 concept elements to this direction.
Magnets for Sarcasm: Making Sarcasm Detection Timely, Contextual and Very Personal (2017) — contributed 1 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_intent_perception_divergence_1ca3455c,
  title = {intent-perception divergence},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/intent-perception-divergence/rdr_1ca3455cfacd47458778b755f7f42e77},
  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_intent_perception_divergence_1ca3455c_v1,
  title = {intent-perception divergence},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/intent-perception-divergence/rdr_1ca3455cfacd47458778b755f7f42e77/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