Sentiment Analysis RNN Project

Project Overview

This project implements a Recurrent Neural Network (RNN) model for sentiment analysis using TensorFlow and Keras. The goal is to classify text data into positive, neutral, or negative sentiments. We use GridSearchCV to fine-tune hyperparameters and improve the model's accuracy.

Table of Contents

• Project Overview
• Installation
• Usage
• Model Architecture
• Hyperparameters
• Results

Installation

To get started, clone this repository and install the necessary packages:

git clone https://github.com/dlynch42/sentiment-analysis.git
cd sentiment-analysis
pip install -r requirements.txt

Requirements

• pandas
• numpy
• tensorflow
• sklearn
• pyprind
• re

Install these packages using pip:

pip install pandas numpy tensorflow scikit-learn pyprind matplotlib seaborn

Usage

To train the model and make predictions, follow these steps:

1. EDA & Preprocessing: Load and preprocess the text data to remove HTML tags, URLs, special characters, and stopwords, then tokenize and stem the text. Create sequence mappings for model.
2. Model Architecutre: Initialize and build the RNN model. Train model on X_train. Tune and adjust hyperparameters using GridSearchCV and Pipe.
3. Test Model: Use the trained model to make predictions on new data.
4. Conclusion: Analyze results.

Model Architecture

The model architecture consists of the following layers:

1. Embedding Layer: Converts input sequences into dense vectors of fixed size.
2. LSTM Layers: Captures dependencies in both forward and backward directions using bidirectional LSTM. Multiple layers can be stacked for deeper representations.
3. Dropout Layers: Helps prevent overfitting.
4. Dense Output Layer: Outputs the final sentiment prediction.
5. Optimizer & Loss Function: Used Adam to optimize and BCE to calculate loss

Hyperparameters

We focused on the following hyperparameters to optimize the model:

• seq_len: Sequence length of the input data.
• lstm_size: Number of units in the LSTM layers.
• num_layers: Number of LSTM layers.
• batch_size: Size of the batches during training.
• learning_rate: Learning rate for the optimizer.

Results

The best model achieved a test accuracy of 73.45%. The results were lower than expected, likely due to over-processing the data. The original text sequences were short, and excessive preprocessing reduced the amount of useful data.