"Blissfully Happy" or "Ready to Fight": Varying Interpretations of Emoji [pdf] (cdn.aaai.org)

Researchers at ICWSM 2016 (Miller et al.) investigated how the same emoji can be interpreted differently depending on their graphic rendering and the platform that displays them. Using an online survey that showed popular emoji as rendered on multiple platforms (e.g., Apple, Google, Android vendors), participants labeled meanings and sentiment for each image. The study measured variance in both semantics and sentiment, produced quantitative metrics of disagreement, and identified which emoji are most and least likely to be misread — demonstrating that visual differences across vendor renderings can flip an emoji’s perceived emotion (hence examples like “blissfully happy” vs. “ready to fight”). This matters for anyone building or evaluating AI/ML systems that rely on emoji as signals: sentiment analysis, social media mining, annotation efforts, and user-interface design. Platform-specific rendering noise can introduce label noise, bias classifiers, and produce cross-platform miscommunication in downstream analyses. The paper’s methodology—crowdsourced interpretation across multiple vendor images and quantified disagreement—provides a blueprint for auditing emoji-driven datasets and suggests practical caution (or normalization) when using emoji as ground truth for training or interpreting models.