Performance Evaluation of Sentiment Analysis on Text and Emoji Data Using End-to-End, Transfer Learning, Distributed and Explainable AI Models

AI models show significant performance drops when encountering unfamiliar data: The sentiment analysis models achieved nearly 98% accuracy on tweets containing emojis seen during training, but accuracy plummeted to 70% when tested on unseen emojis, demonstrating that these models may not generalize well to new variations in real-world deployment.

Emoji interpretation presents unique challenges for AI systems: The research highlights that emojis—which convey complex emotions and cultural meanings—are particularly difficult for AI models to classify consistently, suggesting that sentiment analysis systems relying on emoji data require careful validation before use in marketing or content moderation applications.

Explainability tools are essential for identifying AI bias: The researchers used SHAP (an explainable AI algorithm) to examine model behavior and detect potential biases in their feature sets, underscoring the importance of transparency mechanisms in understanding what patterns AI systems actually learn and whether they reflect problematic assumptions.

Computational efficiency can be improved without sacrificing accuracy: By using distributed training approaches instead of traditional single-threaded models, the researchers reduced runtime by approximately 15% while maintaining the same accuracy levels, suggesting that responsible AI deployment should consider both performance and computational resource efficiency.