Multilingual Emotion Detection in Social Media Using an Attention-Enhanced and Explainable Machine Learning Framework

Abstract

Emotions play a central role in human experience, influencing decisions, behaviors, and social interactions. Despite substantial advances in natural language processing, detecting emotions in multilingual and code-mixed texts remains challenging due to the linguistic diversity of social media users, particularly in low-resource languages. While prior work has focused mainly on monolingual approaches, with limited exploration of multilingual scenarios. This study addresses multilingual emotion detection by creating a machine learning framework enhanced with a custom attention mechanism and fine-tuning strategy. The proposed custom attention mechanism is devised to capture emotionally salient information and improve classification accuracy. Two complementary approaches are explored: 1) a joint multilingual model based on a universal emotion classifier for multiple languages; and 2) a translation-based model that standardizes all inputs into a single target language. Additionally, explainable artificial intelligence (XAI) techniques are implemented to interpret the model’s decisions and enhance transparency. To obtain a proof-of-concept, two existing datasets are transformed into three new multilingual resources with four additional emotion categories. Experiments are conducted using both conventional and advanced machine learning models—integrated with the custom attention mechanism—to evaluate the system’s effectiveness for multilingual emotion recognition tasks. The proposed model achieved superior performance, with accuracy-scores of $93.1\%,79.6\%$, and $93.3\%$ for augmented English, augmented Urdu, and multilingual datasets, respectively.