Enhanced Sentiment Classification in Code-Mixed Texts Using Hybrid Embeddings and Synthetic Data Generation

Social media platforms provide a space for users to share thoughts, connect, and discuss various topics. While some communicate in a single language, others use multiple languages, reflecting the linguistic diversity of their nations. Existing research has largely focused on monolingual data, with less attention on code-mixed data. Studies have explored hybrid models, which have been shown to improve sentiment analysis accuracy compared to single models across various datasets. To enhance sentiment classification in this type of data, we employed a hybrid model combining the distilled multilingual embeddings of BERT with BiLSTM to capture sequential dependencies. Further, we address class imbalance through generating synthetic data with ChatGPT, a large language model with human language comprehension. Our approach shows significant performance improvements across various hyperparameter settings over its vanilla counterparts. Best result showed a 4.82% increase in both accuracy and macro F1 score against DistilmBERT alone. We observed that when prompting ChatGPT for generating synthetic data, every prompt should be distinct to avoid data duplication during data generation. With LIME (Local Interpretable Model-agnostic Explanations) technique, we provide interpretations for correctly classified and misclassified instances weighing words according to their contribution to our model's prediction. A potential area of further study adapting the model to a more large, extensive, and diverse datasets, including various social media platforms and languages, could improve its generalizability and robustness.