Tweet Sentiment Analysis project

Authors

Bradley Payne

Nithya Alavala

---

Project Summary

Sentiment analysis is an important task for many domains including for product feedback and how the general public feels about a given company, product or topic. Many people express their ideas, concerns, movie analysis, and other opinions on twitter through tweets. The goal of this project is to take texts from twitter data and classify it as negative, neutral, or positive.

One area that we would like to explore is the use of emojis in sentiment analysis.

---

Datasets

• 1.6 Million Tweet dataset

  • Contains 800,000 positive and 800,000 negative tweets
  • labeled with 0 for negative, 4 for positive tweets
  • Dataset does NOT contain emojis

• GloVe vectors for word representation

  • Pretrained word embedding

---

Models

We used the following methods

• Random Forest
• Decision Tree
• Naive Bayes
• SVM - Classifier
• LSTM
• BERT

SRC files

Below is a brief summary of each of the source code files

• attention.ipynb

  • Contains the LSTM with the added attention layer

• bert-pretrained.ipynb

  • Implements a version of BERT from the huggingface transformers library
  • Trains BERT for 2 epochs
  • Evaluates on the test set

• data_exploration.ipynb

  • Our first notebook to explore the data
  • Fits the random forest as proof of concept

• data_pipeline.py

  • Defines a data preprocessing pipeline that:
    • Loads the data
    • Cleans
    • Tokenizes
    • Splits into testing and training sets

• lstm_models.ipynb

  • Fits an LSTM model on:
    • Raw data
    • Preprocessed data

• models.ipynb

  • Uses data_pipeline.py functionality to get training and testing set
  • Fits sklearn models:
    • Naive Bayes
    • SVM
    • Decision Tree
    • Random Forest