TL;DR
This paper provides the first systematic analysis of semantic drift in emoji—how their meanings change over time and across different platforms. By analyzing years of Twitter data, the researchers demonstrate that an emoji's meaning is inextricably linked to its visual design; when a platform changes an icon (like turning a realistic pistol into a water gun), its linguistic "neighborhood" shifts fundamentally.
What Problem Does This Paper Address?
While we often think of emoji as static pictures, they function more like evolving linguistic units. However, emoji are unique: unlike words, they are governed by a "four-element" system: a contextual meaning, a strict Unicode "spelling," a formal description (e.g., "face with tears of joy"), and a flexible visual standard decided by platform designers (Apple, Google, etc.).
The authors address a major gap in computational linguistics: how do these visual variations and temporal changes affect how we actually use emoji? Understanding this "semantic drift" isn't just academic—it has real-world implications for sentiment analysis, sarcasm detection, and even legal proceedings where an emoji might be interpreted as a threat or evidence of intent.
Methods
The researchers employed a distributional semantics approach to track how emoji meanings move relative to other words and icons:
- Embedding Models: They trained 108 separate skip-gram (Word2Vec) models—one for each month across three major platforms (iOS, Android, and Web) over a three-year period.
- Quantifying Drift: To measure change, they used a "pair-wise distance" method. Instead of looking at an emoji in isolation, they looked at the distances between all pairs of words/emoji in a vocabulary. If the distance between "pistol" and "death" increases significantly while the distance between "pistol" and "games" decreases, a drift is identified.
- Directional Shift: They developed a statistical test ($\phi$) to determine which specific token in a pair was responsible for the movement, allowing them to isolate whether the emoji changed or the words around it changed.
Datasets
The study used a massive corpus of English tweets collected via the Twitter streaming API:
- Timeframe: May 2016 to April 2019.
- Volume: Over 209 million tweets from 37 million unique users.
- Platform Breakdown: The data was segmented into iOS (113M), Android (56M), and Web (33M) to account for the fact that the same Unicode character looks different to users on different devices.
Key Insights and Ideas
- Visual Change Drives Semantic Change: The most striking example is the Pistol Emoji (🔫). When Apple (and later others) changed the icon from a realistic revolver to a bright green water pistol, its semantic neighbors shifted from "death," "kill," and "2a" (Second Amendment) to "games" and "online trends."
- The "2a" Recovery: Interestingly, after the pistol's visual change, its similarity to "2a" (gun rights) dropped sharply but then showed a "slow but stable recovery." This suggests that while the visual change forced a shift, the community eventually "re-mapped" the new icon back to some of its original political contexts.
- Forced vs. Grass-roots Drift: The paper distinguishes between drift caused by external "regulatory" changes (like a designer changing an icon) and "grass-roots" drift (like the Eggplant 🍆 or Peach 🍑 acquiring sexual connotations purely through community use).
- Instability as Ambiguity: Emoji that "jitter" back and forth in meaning without a clear direction tend to have less cohesive semantic neighborhoods. If an emoji doesn't have a "tight" group of related words, it is more prone to unstable drift.
Key Results
The authors successfully identified several high-profile semantic shifts between 2016 and 2019:
- Pistol (🔫): Shifted from anger/death/gun rights to games/online trends.
- Syringe (💉): Shifted from drug abuse/weapons to medicine (pre-dating the massive COVID-19 shift).
- Cookie (🍪): Shifted from sweets/desserts to baking/food.
- Non-Emoji Drift: The models also captured known word drifts, such as "boomer" shifting from terror/gaming contexts to condescension/millennials (the "OK Boomer" phenomenon).
- Seasonality: The models correctly identified cyclic drifts for seasonal icons like Falling Leaves (🍂) and Skier (⛷️), which move in and out of relevance based on the time of year.
Limitations and Future Work
- Fixed Embeddings: The current method uses fixed embeddings, which struggle to capture emoji with multiple simultaneous senses (polysemy). An emoji might drift in one sense while remaining stable in another.
- Quantifying "Visual Distance": While the paper notes that visual changes cause drift, it does not yet have a mathematical way to quantify how much a visual change (e.g., color vs. shape) contributes to the magnitude of the semantic shift.
- Contextual Nuance: Future work could benefit from using transformer-based models (like BERT) to better understand how emoji meaning changes within specific sentence structures rather than just global corpus statistics.
Summary updated