Emotion Recognition in Code-Mixed Social Media Data Using an Optimized Transformer Architecture
Seshendranath Balla Venkata, Abdolwadwd Alzubaidy, N. Srinivasan, M. Sathish Kumar, Bhargavi Vemala
Abstract
In these days, emotion recognition from social media data has highly increased for applications in public opinion monitoring and user behavior understanding. This task becomes more complex in multilingual and code-mixed environments, such as English-Roman Urdu tweets, where irregular language switching, slang, and spelling variations present additional limitations. However, traditional transformer-based methods including Multilingual Bidirectional and Auto-Regressive Transformers (MBART), struggle to maintain accuracy under these noisy situations. Hence, this research proposed an enhanced Multilingual Bidirectional and Auto-Regressive Transformers with Code-Mixed (MBART-CM) framework. Initially, a curated dataset from the 2019 Pakistan general election is preprocessed by the advanced data augmentation techniques such as slang normalization, spelling variation injection, code-mix simulation, and synonym replacement to improve the robustness against informal language patterns. Subsequently, the MBART architecture is fine-tuned on this augmented dataset with optimized hyperparameters, including adaptive learning rate scheduling, gradient accumulation, and early stopping, to confirm stable convergence and avoid overfitting. Furthermore, Latent Dirichlet Allocation (LDA) is employed to detect thematic overlaps which allows a shift to binary classification for better accuracy. Experimental results show that proposed MBART-CM attains 80% accuracy by outperforming the existing MBART in recognizing emotions from noisy code-mixed tweets.