Ensemble Learning for Tweet Classification of Hate Speech and Offensive Language - Winter/Spring Project 2018

These programs are part of a project that will use an ensemble learning model to detect offensive language and hate speech in tweets. It is composed of:

• A Voting classifier
• An LSTM network
• A Bayesian model
• A Proximity model

Link to full project: https://github.com/quinnbp/WT2018

---

This repository contains:

Voting classifier for hate-speech and offensive language detection in tweets:

• Uses a hard-voting (majority voting) classifier that evaluates the outputs of:

  • An SGD Classifier with log loss
  • A LinearSVM Classifier with L1 feature selection and L2 classification
  • A Perceptron

• Features:

  • TFIDF matrix
  • POS-Tags matrix
  • Sentiment analysis
  • Prescence-of-lexicon-terms score
  • Word embeddings (random and GloVe)

TODO:

• Try improving word embeddings using a neural network based on [2]

Weighting system for ensemble learning

• Has 3 different options for applying weighted voting:
  • Precision score of the classifiers' confusion matrices
  • CEN score
  • Precision + CEN score
  • Equal voting

Confusion matrix class

• Creates a confusion matrix given the output predictions of a classifer and the set of true labels
• Contains operations like getting precision score, storing it as a pdf, getting number of false positives, getting the CEN score of the matrix, etc.

---

All written by Daniel Firebanks

Inspired by the research of:

• [1]Davidson et al. (https://github.com/t-davidson/hate-speech-and-offensive-language)
• [2]Badjatiya et al. (https://github.com/pinkeshbadjatiya/twitter-hatespeech)