Multimodal Sentiment Analysis using Transformers

A deep learning-based sentiment analysis system integrating textual and visual modalities.
Built with transformer architectures (BERT, RoBERTa, GPT) and CNNs for accurate and context-rich opinion mining.

🧠 Key Features

• ✅ Multimodal Fusion: Combines textual (BERT-based) and visual (CNN-based) inputs.
• 🔍 Aspect-Based Sentiment Analysis (ABSA): Fine-grained analysis of sentiment per aspect.
• 🐦 Twitter & Market Sentiment: Analyzes public sentiment from tweets and financial posts.
• ⚙️ Fine-Tuned Transformers: Custom fine-tuning of BERT, RoBERTa, and GPT for domain-specific accuracy.
• 📈 Benchmark Results: Achieved up to 92% accuracy on Stanford Movie Review dataset.
• 🧩 Attention Head Analysis: Interprets model behavior via attention head visualization.
• 💬 Explainable NLP: Visualizes influential tokens and features for each prediction.

🛠 Technologies

• Python, TensorFlow, HuggingFace Transformers (BERT, RoBERTa, GPT)
• Scikit-learn, XGBoost, OpenCV, NumPy, Pandas
• Pretrained datasets: IMDb, Twitter COVID-19, Stanford Sentiment Treebank