Project Text 2: Sentiment Classification

This repository contains the code for the second project of EPFL's Machine Learning course.

Team Members

• Adrien LAYDU
• Manon MICHEL
• François MICHEL

Goal

The goal of the project is to predict as correctly as possible on a large dataset if tweet messages used to contain a positive :) or negative :( smiley, by considering only their remaining text.

Package requirements

To run our code you need to install the following packages:

• Numpy

  pip install numpy
  

• Scikit-Learn

  pip install sklearn
  

• NLTK (The Natural Language Processing Toolkit)

  pip install nltk
  

• Wordninja

  pip install wordninja
  

• Pandas

  pip install pandas
  

• PyTorch

  pip install torch
  

• TQDM

  pip install tqdm
  

• The Transformers repository by Huggingface

  pip install git+https://github.com/huggingface/transformers.git
  

• Regularized logistic regression using GD or SGD.

  pip install numpy
  

Reproducing this Project

To run our Machine Learning algorithm with the best parameters we found:

• Clone this project
• Download the file twitter-datasets.zip on AIcrowd containing the datasets
• Unzip it
• Put it in a folder called /data inside the repository
• Download all the packages required
• Run the run.py file with the following command:

  python3 run.py
  

Files description

Each file can be found in the folder scripts/:

• run.py contains the code to get a prediction of the sentiments on the test dataset.
• helper.py contains functions to import the datasets.
• preprocess.py contains functions to preprocess tweets.
• bert.py contains functions to train the BERT model and evaluate the accuracy of the model.
• tweetToVec contains the function to convert the tweets into vectors using GloVe.
• data_handler.py contains the function to split the datas into a training and a validation set.
• data_processing.py contains functions to tokenize the tweets using the BERT technique.
• bert_collab.ipynb is a notebook used to train the BERT model using Google Colab.
• Simple_models.ipynb is a notebook containing all others the models (excluding BERT) we used to perform the sentiments analysis.