Sudheer17/sentiment_lstm

🧠 Sentiment Analysis Using Deep Learning (LSTM)

🚀 Deep Learning Based Twitter Sentiment Analysis using LSTM Neural Networks

Predict whether a tweet expresses Positive, Negative, or Neutral sentiment with a trained LSTM model.

---

📌 Overview

This project implements an End-to-End Sentiment Analysis System using Long Short-Term Memory (LSTM) networks.

The model learns contextual information from tweets after performing extensive Natural Language Processing (NLP) preprocessing.

Unlike traditional Machine Learning models, LSTM captures sequential dependencies between words, making it highly effective for sentiment classification.

The trained model predicts one of the following sentiments:

• 😊 Positive
• 😐 Neutral
• 😞 Negative

---

✨ Features

✅ English Tweet Filtering

✅ Advanced Text Cleaning

✅ HTML Entity Decoding

✅ URL Removal

✅ Stopword Removal

✅ Negation Word Preservation

✅ Lemmatization

✅ Tokenization

✅ Sequence Padding

✅ Deep LSTM Network

✅ Automatic Label Generation using TextBlob

✅ Confidence Score Prediction

✅ Save & Reload Model

✅ Ready for Web Deployment

---

🧠 Deep Learning Architecture

Input Tweets
      │
      ▼
Text Cleaning
      │
      ▼
Tokenization
      │
      ▼
Padding Sequences
      │
      ▼
Embedding Layer
      │
      ▼
LSTM (128)
      │
      ▼
Dropout
      │
      ▼
LSTM (64)
      │
      ▼
Dropout
      │
      ▼
LSTM (32)
      │
      ▼
Global Max Pooling
      │
      ▼
Dense (64)
      │
      ▼
Dropout
      │
      ▼
Dense (Softmax)
      │
      ▼
Prediction

---

📂 Dataset

The project uses a Twitter sentiment dataset.

Dataset includes:

• User ID
• Tweet
• Language

Only English Tweets are used for training.

---

⚙️ NLP Preprocessing Pipeline

The preprocessing pipeline performs:

• Convert text to lowercase
• Decode HTML entities
• Remove URLs
• Remove HTML tags
• Remove punctuation
• Remove numbers
• Remove extra spaces
• Preserve negation words
• Remove stopwords
• Lemmatization
• Clean sentence generation

Example

Original Tweet

I absolutely LOVE this phone!! 😍😍
https://abc.com

Clean Tweet

absolutely love phone

---

🏷 Label Generation

Instead of manually labeling data, the project automatically generates sentiment labels using TextBlob Polarity.

Polarity	Sentiment
< 0	Negative
= 0	Neutral
> 0	Positive

---

🔤 Tokenization

Tokenizer Configuration

• Vocabulary Size : 10,000
• OOV Token : <OOV>

Each cleaned tweet is converted into integer sequences.

Example

I love AI

↓

[15, 41, 280]

---

📏 Sequence Padding

Every tweet is padded to a fixed length.

Maximum Length = 30

Padding Type

Post Padding

Truncation

Post Truncation

---

🧠 Model Summary

Embedding Layer

↓

LSTM (128)

↓

Dropout (0.2)

↓

LSTM (64)

↓

Dropout (0.2)

↓

LSTM (32)

↓

Dropout (0.2)

↓

GlobalMaxPooling1D

↓

Dense (64)

↓

Dropout (0.2)

↓

Dense (3)

↓

Softmax

---

⚙️ Training Configuration

Parameter	Value
Optimizer	Adam
Loss	Sparse Categorical Crossentropy
Epochs	20
Batch Size	32
Validation Split	20%

Callbacks

• EarlyStopping
• ModelCheckpoint
• ReduceLROnPlateau

---

📊 Model Evaluation

Evaluation Metrics

• Accuracy
• Loss
• Classification Report
• Precision
• Recall
• F1 Score
• Confusion Matrix

---

🔮 Sample Predictions

Tweet	Prediction
I absolutely love this phone!	😊 Positive
This is the worst service ever.	😞 Negative
The meeting starts at 10 AM tomorrow.	😐 Neutral
India won the cricket match!	😊 Positive
Russia attacked Ukraine today.	😞 Negative

---

💾 Saved Files

best_model.keras

Best model saved during training.

sentiment_lstm.keras

Final trained model.

tokenizer.pkl

Tokenizer object.

label_encoder.pkl

Label encoder.

---

📁 Repository Structure

sentiment_lstm/

│── best_model.keras
│── sentiment_lstm.keras
│── tokenizer.pkl
│── label_encoder.pkl
│── sentiment.csv
│── README.md

---

🚀 Future Improvements

• Streamlit Web Application
• Hugging Face Deployment
• REST API using FastAPI
• Docker Containerization
• ONNX Model Conversion
• Attention-based Bi-LSTM
• Transformer Comparison (BERT)
• Live Twitter/X Sentiment Prediction
• Batch Prediction Support
• Explainable AI (LIME/SHAP)

---

🛠 Tech Stack

• Python
• TensorFlow
• Keras
• NumPy
• Pandas
• NLTK
• TextBlob
• Scikit-Learn
• Matplotlib
• Seaborn

---

📦 Installation

git clone https://github.com/YOUR_USERNAME/sentiment_lstm.git

cd sentiment_lstm

pip install -r requirements.txt

---

▶️ Run

python app.py

or

jupyter notebook

---

🎯 Applications

• Social Media Monitoring
• Customer Feedback Analysis
• Product Reviews
• Political Opinion Mining
• Brand Reputation Analysis
• Market Research
• Public Opinion Analysis

---

👨‍💻 Author

Sudheer Muthyala

🎓 B.Tech - Electronics & Communication Engineering

💻 Python Developer | Machine Learning Enthusiast | Data Science Aspirant

GitHub: https://github.com/M-Sudheer18

---

⭐ Support

If you found this project useful, consider giving it a ⭐ on GitHub.

It helps others discover the project and motivates future improvements.

---

⭐ Star • 🍴 Fork • 🚀 Contribute

Built with ❤️ using TensorFlow, Keras & NLP