SentimentAI is a deep learning-based web application that classifies tweets into four sentiment categories using a hybrid BERT + BiLSTM architecture.
The system leverages transformer-based contextual embeddings along with sequential learning to achieve high accuracy and robust sentiment understanding.
👉 Designed for real-world NLP applications like:
- Social media monitoring
- Brand sentiment analysis
- Opinion mining
| Layer | Details |
|---|---|
| BERT Encoder | bert-base-uncased |
| Dropout | p = 0.10 |
| BiLSTM | 2 layers · hidden size = 128 |
| Linear | 256 → 4 classes |
| Log-Softmax | Output activation |
| Class | Description |
|---|---|
| 🟢 Positive | Happiness, excitement, praise |
| 🔴 Negative | Anger, sadness, criticism |
| 🟡 Neutral | Informational, no emotion |
| ⚪ Irrelevant | Spam or unrelated content |
| Metric | Score |
|---|---|
| Accuracy | 90.0% |
| Precision | 90.0% |
| Recall | 90.0% |
| F1 Score | 90.0% |
- 📊 74,681 tweets
- 🏷️ 4 sentiment classes
- 📌 Real-world Twitter dataset
| Layer | Technology |
|---|---|
| Frontend | Streamlit |
| Model | PyTorch + HuggingFace |
| NLP | Transformers (BERT) |
| Visualization | Plotly |
| Deployment | Streamlit Cloud |
| Model Hosting | HuggingFace Hub |
- 📝 User inputs tweet text
- 🔍 Text is tokenized using BERT tokenizer
- 🧠 BERT extracts contextual embeddings
- 🔄 BiLSTM processes sequence
- 📊 Model predicts sentiment class
- 📱 Result displayed in UI
git clone https://github.com/puneethas26/sentiment-bert-bilstm.git
cd sentiment-bert-bilstm
pip install -r requirements.txt
streamlit run app.py