Sentiment Analysis on tweets (using BiLSTM)

📌 Introduction

This project focuses on building a Sentiment Analysis Model that classifies tweets as positive or negative using Deep Learning and Natural Language Processing (NLP) techniques. The model is trained on the Sentiment140 dataset, which contains 1.6 million tweets, making it ideal for large-scale text sentiment prediction tasks.

The project demonstrates essential components of NLP pipelines, including data cleaning, tokenization, sequence modeling using LSTMs, and evaluation using performance metrics and visualizations.

---

✨ Features

• ✔️ Automated text preprocessing (removal of URLs, mentions, hashtags, special characters)
• ✔️ Stemming and stopword removal using NLTK
• ✔️ Deep Learning model using Bidirectional LSTM
• ✔️ Embedding layer for dense word representation
• ✔️ Train/Validation/Test split for robust evaluation
• ✔️ Performance visualization (accuracy, loss plots)
• ✔️ Confusion matrix & classification report
• ✔️ Custom prediction function for classifying new tweets
• ✔️ High accuracy with regularization, dropout, and optimization techniques

---

📂 Dataset

Sentiment140 Dataset (Kaggle)

• Total tweets: 1,600,000

• Labels:

  • 0 → Negative
  • 4 → Positive (converted to 1 during preprocessing)

Dataset Contents

• Original tweet text
• Polarity label
• Metadata (IDs, timestamps, query info — not used)

Preprocessing Steps

• Remove URLs
• Remove mentions (@username)
• Remove hashtags (#happy → happy)
• Remove non-alphabetic characters
• Convert to lowercase
• Tokenization
• Stopword removal
• Stemming using PorterStemmer

The cleaned text is stored in a new column: stemmed_content.

---

Model Architecture

Layer	Description
Embedding (128-dim)	Converts tokens to dense vector representations
SpatialDropout1D (0.3)	Regularization for embedding layer
Bi-LSTM (128 units)	Captures bidirectional sequence context
Bi-LSTM (64 units)	Additional layer for deeper learning
Dense (32, ReLU)	Fully connected hidden layer with L2 regularization
Dropout (0.5)	Prevents overfitting
Dense (1, Sigmoid)	Output layer for binary classification

Hyperparameters

• Tokenizer vocabulary size: 50,000

• Sequence length: 50 tokens

• Optimizer: Adam (lr = 0.0005)

• Loss: Binary Crossentropy

• Batch size: 512

• Epochs: 12

• Callbacks:

  • EarlyStopping
  • ReduceLROnPlateau

---

📊 Results

Evaluation Metrics

• Accuracy: (Add final accuracy observed during training)

• Classification Report:

  • Precision
  • Recall
  • F1-Score

Confusion Matrix

Displays True Positive, True Negative, False Positive, False Negative counts.

Training Curves

Two graphs:

• Accuracy vs Epochs
• Loss vs Epochs

These help visualize learning behavior and overfitting.

Sample Predictions

Tweet: I absolutely love this new phone!
→ Positive 😀 (Confidence: 0.982)

Tweet: This is the worst movie ever.
→ Negative 😡 (Confidence: 0.876)

---

📚 References

• Sentiment140 Dataset — Kaggle
• TensorFlow Documentation
• NLTK Stopword Corpus
• scikit-learn
• Research papers on LSTM and Word Embeddings