Comprehensive Sentiment Analysis Using Different Methods

Introduction

This project performs comprehensive sentiment analysis, focusing on Twitter data. The goal is to predict if a tweet message used to contain a positive :) or negative :( smiley, by considering only the remaining text.

Project Focus

The project compares various sentiment analysis techniques to identify the most effective ones, particularly for Twitter data. It involves:

• Analyzing performance in different scenarios.
• Evaluating accuracy and efficiency.
• Exploring text representation methods like Bag-of-Words, TF-IDF, Word2Vec, and GloVe.

Data Collection and Preprocessing

Data includes tweets, processed via steps like:

• Cleaning and normalizing text.
• Handling missing values and outliers.
• Utilizing advanced text representation methods.

Methods Evaluated

The project evaluates:

• Traditional models: Decision Tree, Random Forest, Logistic Regression.
• Advanced models: LSTM, BERT, RoBERTa.

Training and Evaluation

• Data split into training, validation, and testing sets.
• Emphasis on transformer-based models for their effectiveness.
• Evaluation metrics include accuracy, precision, recall, F1-score, and model robustness.

Structure of this repository

--Data              To store the original dataset
--Manipulated       To store the trained models
--Src               Providing the code we used in this project
run.py              To get the submission file using the best model we find

Dependencies

Dependencies include Python 3.8, Pandas, Numpy, Sklearn, NLTK, TensorFlow, PyTorch.

Models Used

Models used are Decision Tree, Random Forest, Logistic Regression, LSTM, BERT, RoBERTa.

Environment & Execution Instructions

• Run in a Python environment, preferably a virtual environment.
• Main execution file is run.py.
  • The run.py is simply a prediction function of the best model we find (Roberta with 5 fold majority voting). On GPU, the prediction should be given within minutes. To train the model, please go to src folder and check different methods. But training is very time-consuming.