SentimentFlow-RNN-and-LSTM-Powered-Tweet-Analysis

People share opinions on Twitter every second. Companies, governments, and researchers want to know what people feel about products, events, or topics in real time. Reading tweets manually is impossible at scale. SentimentFlow is a deep learning based sentiment analysis project for tweets. It trains two models like a basic RNN and an LSTM network. SentimentFlow is a deep learning based sentiment analysis project that uses two architectures

• a Simple RNN model
• an LSTM model

Both models classify tweets as positive or negative, allowing easy comparison of their performance.

---

Table of Contents

1. Features
2. Dataset
3. Preprocessing
4. Model Architectures
5. Training Details
6. Results
7. Confusion Matrices
8. Future Work
9. License

---

Features

• Text sentiment classification for tweets
• Comparison between RNN and LSTM models
• Clean preprocessing pipeline
• Metrics including accuracy, F1 score, and confusion matrix
• Simple and clear notebooks for learning deep NLP
• Works in Jupyter Notebook environment

---

Dataset

This project uses a tweet sentiment dataset (negative vs positive).

Typical columns include:

• text – tweet content
• label – sentiment (0 = negative, 1 = positive)

You can replace this section with your exact dataset link.

---

Preprocessing

Both notebooks include standard NLP preprocessing steps:

• Lowercasing
• Removing URLs, mentions, hashtags and extra symbols
• Tokenization using Keras Tokenizer
• Padding sequences to fixed length
• Splitting into train, validation and test sets

---

Model Architectures

1. RNN Model

A simple Recurrent Neural Network that processes text word by word.

Components:

• Embedding layer
• Simple RNN layer
• Dense layer
• Softmax or sigmoid output layer

Strengths: easy to train, fewer parameters Weaknesses: struggles with long term dependencies

---

2. LSTM Model

A Long Short Term Memory model designed to retain long term information.

Components:

• Embedding layer
• LSTM layer
• Dense layer(s)
• Output layer

Strengths: handles long text patterns better, avoids vanishing gradients Weaknesses: higher computation cost

---

Training Details

• 10 epochs for both models
• Adam optimizer
• Binary cross entropy loss
• Validation performance monitored each epoch

---

Results

LSTM Model

• Train Accuracy: 86.12 percent
• Val Accuracy: 83.29 percent
• Val F1 Score: 0.8325
• Final Loss: 0.3211

RNN Model

• Train Accuracy: 81.06 percent
• Val Accuracy: 79.55 percent to 79.7 percent
• Val F1 Score: 0.7917 to 0.7935
• Final Loss: 0.4248

Summary

The LSTM model performs better in every metric. It learns long term dependencies in tweets more effectively and shows stronger generalization.

---

Confusion Matrices

LSTM Confusion Matrix

True Label	Predicted Negative	Predicted Positive
Negative	100225	19600
Positive	20501	99674

RNN Confusion Matrix

True Label	Predicted Negative	Predicted Positive
Negative	65046	15005
Positive	17727	62222

LSTM shows fewer misclassifications and more consistent predictions.

---

Future Work

• Add GRU or Bidirectional LSTM models
• Use pretrained embeddings like GloVe
• Add attention mechanism
• Extend project to 3 class sentiment (positive, negative, neutral)
• Use transformer models such as BERT for modern NLP performance

---

License

MIT License (Modify if needed)

---