Sentiment Analysis and Emotion Detection Using Transformer Models in Multilingual Social Media Data

Abstract

The rapid expansion of multilingual social media platforms has resulted in a surge of user-generated content, introducing challenges in sentiment analysis and emotion detection due to code-switching, informal text, and linguistic diversity. Traditional rule-based and machine learning models struggle to process multilingual complexities effectively, necessitating advanced deep-learning approaches. This study develops a transformer-based sentiment analysis and emotion detection system capable of handling multilingual and code-mixed social media text. The proposed fine-tuned Cross-lingual Language Model – Robust (XLM-R) model is compared against state-of-the-art transformer models (mBERT, T5) and traditional classifiers (support vector machine (SVM), Random Forest) to assess its cross-lingual sentiment classification performance. A multilingual dataset was compiled from Twitter, YouTube, Facebook, and Amazon Reviews, covering English, Spanish, French, Hindi, Arabic, Tamil, and Portuguese. Data preprocessing included tokenization, stopword removal, emoji normalization, and code-switching handling. Transformer models were fine-tuned using cross-lingual embeddings and transfer learning, with accuracy, F1-score, and confusion matrices for performance evaluation. Results show that XLM-R outperformed all baselines, achieving an F1-score of 90.3%, while multilingual Bidirectional Encoder Representations from Transformers (mBERT) and T5 scored 84.5% and 87.2%, respectively. Preprocessing improved performance by 7%, particularly in code-mixed datasets. Handling code-switching increased accuracy by 8.9%, confirming the model’s robustness in multilingual sentiment analysis. The findings demonstrate that XLM-R effectively classifies sentiments and emotions in multilingual social media data, surpassing existing approaches. This study supports integrating transformer-based models for cross-lingual natural language processing (NLP) tasks, paving the way for real-time multilingual sentiment analysis applications.