Design of an Iterative Model Incorporating mBERT, CSLM, and ADAN for Multilingual and Code-Switched Sentiment Analysis

The global spread of code-switched and multilingual content on Twitter highlights the importance of effective sentiment analysis. However, existing algorithms often overlook regional and multilingual tweets, resulting in reduced performance. To address this, our study introduces a model that integrates mBERT, CSLM, and ADAN to enhance sentiment analysis. mBERT, trained on a large bilingual dataset, captures intricate contextual embeddings from both English and Marathi tweets. CSLM is designed to handle the complexities of code-switching in mixed-language content, while ADAN uses adversarial training to reduce biases and improve domain invariance across political, social, and commercial tweets. The combined mBERT-CSLM-ADAN model achieves an accuracy of 89%, outperforming the baseline mBERT and CSLM models. This approach offers a more generalizable, robust, and domain-flexible solution for social media sentiment classification, particularly in multilingual environments.