🐦 Twitter Sentiment Analysis using Deep Learning and Transformer Architectures

📌 Project Overview

This project presents a comprehensive comparative study of deep learning and transformer-based architectures for Twitter sentiment classification.

Four different models were implemented and evaluated on a large-scale Twitter sentiment dataset containing over 1 million tweets:

• 🔹 Simple Recurrent Neural Network (RNN)
• 🔹 Long Short-Term Memory Network (LSTM)
• 🔹 Gated Recurrent Unit (GRU)
• 🔹 Bidirectional Encoder Representations from Transformers (BERT)

The objective is to analyze how modern neural architectures improve sentiment classification performance and identify the most suitable model for large-scale sentiment analysis.

---

📊 Dataset Information

Attribute	Value
Total Tweets	1,048,572
Features	6
Negative Tweets	799,996
Positive Tweets	248,576
Sentiment Classes	Binary Classification

Dataset Class Distribution

Distribution of positive and negative sentiment classes.

---

🏗️ Project Workflow

┌─────────────────────────────────────────────┐
│          Twitter Sentiment Dataset          │
│            1,048,572 Tweets                 │
└─────────────────────────────────────────────┘
                      │
                      ▼
┌─────────────────────────────────────────────┐
│         Exploratory Data Analysis           │
│ • Class Distribution                        │
│ • Tweet Length Analysis                     │
│ • Word Frequency Analysis                   │
│ • Positive/Negative Word Clouds             │
└─────────────────────────────────────────────┘
                      │
                      ▼
┌─────────────────────────────────────────────┐
│            Text Preprocessing               │
│ • Lowercasing                               │
│ • URL & Username Removal                    │
│ • Punctuation Removal                       │
│ • Whitespace Normalization                  │
└─────────────────────────────────────────────┘
                      │
                      ▼
┌─────────────────────────────────────────────┐
│          Train / Validation / Test          │
│        175k / 37.5k / 37.5k Samples         │
└─────────────────────────────────────────────┘
                      │
          ┌───────────┼───────────┐
          │           │           │
          ▼           ▼           ▼
┌─────────────┐ ┌─────────────┐ ┌─────────────┐
│ Vocabulary  │ │ Vocabulary  │ │ Vocabulary  │
│ 20,000      │ │ 20,000      │ │ 20,000      │
│ Max Len=32  │ │ Max Len=32  │ │ Max Len=32  │
└──────┬──────┘ └──────┬──────┘ └──────┬──────┘
       │               │               │
       ▼               ▼               ▼
┌─────────────┐ ┌─────────────┐ ┌─────────────┐
│     RNN     │ │    LSTM     │ │     GRU     │
│ Train from  │ │ Train from  │ │ Train from  │
│ Scratch     │ │ Scratch     │ │ Scratch     │
└──────┬──────┘ └──────┬──────┘ └──────┬──────┘
       │               │               │
       └───────┬───────┴───────┬───────┘
               │               │
               ▼               │
      ┌─────────────────┐      │
      │ Performance     │      │
      │ Evaluation      │      │
      └────────┬────────┘      │
               │               │
               ▼               │
        ┌─────────────┐        │
        │ Accuracy    │        │
        │ Precision   │        │
        │ Recall      │        │
        │ F1 Score    │        │
        │ ROC-AUC     │        │
        └─────────────┘        │
                               │
                               ▼

┌─────────────────────────────────────────────┐
│                  BERT                       │
│ bert-base-uncased                           │
│ Pretrained Transformer                      │
│ Full Fine-Tuning                            │
│ Max Length = 48                             │
└─────────────────────────────────────────────┘
                      │
                      ▼
┌─────────────────────────────────────────────┐
│           Comparative Analysis              │
│ RNN vs LSTM vs GRU vs BERT                  │
│ Accuracy • Precision • Recall               │
│ F1 Score • ROC-AUC                          │
└─────────────────────────────────────────────┘
                      │
                      ▼
┌─────────────────────────────────────────────┐
│              Best Model: BERT               │
│ Accuracy : 87.72%                           │
│ F1 Score : 72.48%                           │
│ ROC-AUC  : 92.21%                           │
└─────────────────────────────────────────────┘

The project follows a complete NLP pipeline:

1. Exploratory Data Analysis
2. Text Preprocessing
3. Dataset Splitting
4. Sequence Preparation
5. Model Development
6. Performance Evaluation
7. Comparative Analysis
8. Best Model Selection

---

🔍 Exploratory Data Analysis

Key findings from the dataset:

• No missing values detected
• Strong class imbalance (76% Negative, 24% Positive)
• Average tweet length ≈ 74 characters
• Median tweet length ≈ 70 characters
• Significant vocabulary overlap between sentiment classes

Word Cloud Analysis

Most frequently occurring words in positive and negative tweets.

---

🧹 Text Preprocessing

The following preprocessing pipeline was applied:

• Lowercasing
• URL Removal
• Username Removal
• Punctuation Removal
• Numeric Filtering
• Whitespace Normalization

Sequence Preparation

RNN / LSTM / GRU

• Vocabulary Size: 20,000
• Maximum Sequence Length: 32

BERT

• Pretrained BERT Tokenizer
• Maximum Sequence Length: 48
• Attention Masks
• Special Tokens

---

🧠 Models Implemented

1️⃣ Simple RNN

Baseline recurrent neural network trained from scratch.

Architecture

• Embedding Layer (128)
• Simple RNN Layer (128 Hidden Units)
• Dropout (0.3)
• Dense Output Layer

RNN Training Dynamics

Training and validation loss during RNN training.

---

2️⃣ LSTM

Long Short-Term Memory architecture designed to capture long-range contextual dependencies.

Architecture

• Embedding Layer (128)
• LSTM Layer (128 Hidden Units)
• Dropout (0.3)
• Dense Output Layer

LSTM Training Dynamics

Training and validation loss during LSTM training.

---

3️⃣ GRU

Gated Recurrent Unit architecture providing efficient sequence modeling with fewer parameters.

Architecture

• Embedding Layer (128)
• GRU Layer (128 Hidden Units)
• Dropout (0.3)
• Dense Output Layer

GRU Training Dynamics

Training and validation loss during GRU training.

---

4️⃣ BERT

Pretrained transformer model fine-tuned end-to-end for sentiment classification.

Architecture

• BERT Base (Uncased)
• Transformer Encoder
• Classification Head

Training Strategy

• Full Fine-Tuning
• AdamW Optimizer
• Learning Rate = 2e-5
• Best Model Checkpointing

BERT Training Dynamics

Training and validation loss during BERT fine-tuning.

---

📈 Model Performance Comparison

Model	Accuracy	Precision	Recall	F1 Score	ROC-AUC
RNN	0.7530	0.4820	0.5637	0.5197	0.7084
LSTM	0.8117	0.5789	0.7541	0.6550	0.8723
GRU	0.8030	0.5613	0.7730	0.6504	0.8697
BERT	0.8772	0.7730	0.6822	0.7248	0.9221

---

📊 Comparative Analysis

Overall Model Comparison

Comparison of Accuracy, Precision, Recall, F1 Score, and ROC-AUC across all architectures.

---

F1 Score Comparison

---

ROC-AUC Comparison

---

🏆 Key Findings

RNN

• Established baseline performance
• Limited contextual understanding
• Lowest overall performance

LSTM

• Significant improvement over vanilla RNN
• Strong sequence modeling capability
• Approximately 26% improvement in F1 Score

GRU

• Comparable performance to LSTM
• Faster and computationally efficient
• Highest Recall score

BERT

• Best overall performance
• Highest Accuracy, Precision, F1 Score, and ROC-AUC
• Benefited from pretrained contextual language representations

---

🎯 Best Model

🥇 BERT

Metric	Score
Accuracy	87.72%
Precision	77.30%
Recall	68.22%
F1 Score	72.48%
ROC-AUC	92.21%

---

📂 Project Structure

sentiment-analysis-comparison/

├── data/
├── notebooks/
│   ├── 01_EDA.ipynb
│   ├── 02_Preprocessing.ipynb
│   ├── 03_RNN.ipynb
│   ├── 04_LSTM.ipynb
│   ├── 05_GRU.ipynb
│   ├── 06_BERT.ipynb
│   └── 07_Comparative_Analysis.ipynb
│
├── models/
├── images/
├── reports/
├── results/
│
├── README.md
└── requirements.txt

---

🛠️ Technologies Used

Programming

• Python

Data Processing

• Pandas
• NumPy

Visualization

• Matplotlib
• Seaborn
• WordCloud

Deep Learning & NLP

• PyTorch
• Hugging Face Transformers
• Scikit-Learn

---

🚀 Future Improvements

• Hyperparameter Optimization
• Threshold Tuning
• RoBERTa Comparison
• DeBERTa Comparison
• Multi-Class Sentiment Classification
• Real-Time Sentiment Monitoring Dashboard
• FastAPI Deployment
• Streamlit Application

---

📜 Conclusion

This project demonstrates the progression of sentiment classification performance from traditional recurrent neural networks to transformer-based language models.

While LSTM and GRU significantly outperformed the vanilla RNN baseline, BERT achieved the strongest overall performance with an Accuracy of 87.72%, an F1 Score of 72.48%, and a ROC-AUC of 92.21%.

The results highlight the effectiveness of transfer learning and contextual language understanding in modern Natural Language Processing applications.