Sentiment Analysis of Multilingual Social Media Posts Using Deep Learning Models: A Case Study on Marathi-English Code-Switching

With the widespread adoption of social mass medium platforms, the bulk of user-generated content has expanded dramatically. This digital discourse, especially in multilingual regions such as India, oftentimes features codification-switch—the practice of understudy between two or more words within an exclusive sentence or conversation. This stage unique challenge in sentiment analysis, particularly when analyzing low-imagination language pairs like Marathi and English. This study proposes a comprehensive, intercrossed deep learning glide path to in effect sort sentiments from such computer code-interchange data point. A curated dataset of 15, 000 Marathi-English social sensitive posts is gather and manually annotated for sentiment mutual opposition. Versatile preprocessing stair including Romanized school text normalization and crossbreed tokenization are employed to develop the dataset. A hybrid model integrating Convolutional Neural Networks (CNN), Long Short-Term Memory meshwork (LSTM), and Grey Wolf Optimization (GWO) is propose and benchmarked against traditional modeling such as Naive Bayes, Support Vector Machine (SVM), BiLSTM, mBERT, and XLM-RoBERTa. The results certify superior truth, recall, and F1-account for the purpose CNN-LSTM-GWO poser. Additionally, a detailed cause study examines sentiment patterns in real-world cultural and political effect. This subject field contributes to the progress of sentiment analysis in computer code-switched, low-resourceful voice communication, offering insight and practical method acting applicable to multilingual NLP and social spiritualist monitoring systems.