Contextual Analysis of Emojis 😄

Overview

Emojis are one of the most expressive parts of digital language, but their meaning is rarely fixed. The same 😭 can signal grief in one message and laughter in another. For my final project in CSCI 534 (Affective Computing) at USC, my team and I looked at how the surrounding textual context changes the emotional interpretation of an emoji, and whether models can pick up on that shift.

Approach

• Collected a corpus of social media posts containing frequently used emojis, keeping the full text so that context was preserved.
• Annotated and clustered posts by the affective category of the surrounding text (e.g. joy, sadness, sarcasm, affection).
• Trained embedding-based models on emoji-in-context, then compared the resulting representations against the emoji's "dictionary" sentiment.
• Measured how far each emoji's inferred sentiment drifted across contexts to identify the most context-dependent symbols.

What we found

A small set of high-frequency emojis carried the majority of ambiguity: their sentiment flipped polarity depending on the sentence around them, while more literal emojis stayed stable. Models that were given the full sentence consistently outperformed emoji-only baselines at predicting the intended emotion, which supports treating emojis as context-dependent tokens rather than fixed sentiment lexicon entries.

Why it mattered to me

This project sat right at the intersection of the things I care most about — natural language processing, cognitive science, and how people actually communicate. It shaped how I think about ambiguity in language data, and it carried directly into later NLP and social psychology research I worked on.