TL;DR
Researchers analyzed over 1.2 billion messages to determine why certain emojis go viral while others languish in obscurity. By mapping emojis and words into a shared semantic space, they discovered that popularity is driven by semantic clarity, proximity to popular slang, and an emoji's ability to either replace a word or complement a sentimental context.
What Problem Does This Paper Address?
Despite being a "universal language," emoji usage is highly skewed; a handful of icons like 😂 dominate, while hundreds of others are rarely used. Previous research often focused on social diffusion (who sends what to whom), but this paper investigates the intrinsic semantic factors of the emojis themselves.
The authors tackle three main questions:
- Does the "clarity" of an emoji’s meaning affect its popularity?
- How do emojis interact with the words around them—do they replace words or add new layers of meaning?
- How does the emotional "temperature" of a message influence emoji usage?
Methods
The researchers employed a state-of-the-art network embedding tool called LINE (Large-scale Information Network Embedding) to project emojis and words into a high-dimensional vector space.
They utilized two specific types of proximity:
- LINE-1st (First-order proximity): Captures how often a word and emoji co-occur (e.g., "Cheers" and 🍷).
- LINE-2nd (Second-order proximity): Captures how similar their contexts are, indicating "replaceability" (e.g., "wine" and 🍷).
From these embeddings, they constructed k-nearest-neighbor (kNN) graphs and analyzed the "egonet" (the immediate neighborhood) of each emoji. They measured popularity against several metrics:
- Local Clustering Coefficient (LCC): A measure of how "dense" and unambiguous an emoji's semantic neighborhood is.
- Pointwise Mutual Information (PMI): To measure the association between an emoji and its nearest word.
- Semantic Shifting: A novel approach where they calculated the "shift" in a message's mean vector when an emoji was removed, using a Gaussian Mixture Model (GMM) to distinguish between complementary and supplementary roles.
Datasets
The study used a massive, anonymized dataset from the Kika Emoji Keyboard:
- Scope: 1.22 billion messages from 1.03 million US-based English speakers.
- Timeframe: September 2015.
- Vocabulary: 5.9 million unique tokens (words, emoticons, and emojis).
- Emoji Set: 1,281 supported emojis.
Key Insights and Ideas
- Semantic Clarity is King: Popular emojis tend to have very "dense" semantic neighborhoods (high LCC). If an emoji sits in a "thinner" area of the semantic space with few close neighbors, it is likely ambiguous and less popular.
- The "Company You Keep": An emoji is more likely to be popular if its nearest semantic neighbors are themselves popular words or internet slang (like "lol" or "lmao").
- Two Distinct Roles: Emojis function in two ways:
- Supplementary: Entity-related emojis (like 🍷) often act as direct substitutes for words.
- Complementary: Sentiment-related emojis (like ❤️) usually don't change the core meaning of a sentence but provide emotional nuance.
- Sentiment Sensitivity: Emojis are significantly more prevalent in sentimental contexts. Interestingly, while popular emojis are often used to replace words (supplementary), they are simultaneously used to complement the overall sentiment of the message.
Key Results
- Power-Law Distribution: Emoji popularity follows a classic power-law; a tiny fraction of emojis accounts for the vast majority of usage.
- Regression Analysis: The authors' model achieved an $R^2$ of 0.906, indicating that structural properties in the semantic space (LCC, distance to nearest word, neighbor frequency) are incredibly strong predictors of emoji popularity.
- Semantic Shift Weights: Using GMM, they found that for the popular 😂 emoji, the "supplementary" component (carrying the message's core meaning) was much smaller (0.01) compared to the 🐶 emoji (0.12), confirming that 😂 is primarily used to complement existing text rather than replace it.
Limitations and Future Work
- Correlation vs. Causality: The study identifies strong correlations but cannot definitively prove that semantic clarity causes popularity.
- Temporal Constraints: The data spans only one month. It doesn't account for how the "meaning" of an emoji might drift over years or how new emojis (like the "middle finger" or "taco" which were new at the time) are adopted.
- Platform Aggregation: The data aggregates all apps (SMS, WhatsApp, Facebook). Usage patterns likely differ between a professional email and a casual Snapchat message.
- Future Directions: The authors suggest moving toward causal inference and exploring how cultural differences influence these semantic embeddings.
Summary updated