Hate Speech Classifier

This project was created as part of the Metis data science bootcamp.

The goal of this project was to construct a classifier to separate hate speech from other offensive twitter comments.

Repository Description

• The data .csv can be found at this repository. Place it within the data folder, or redirect the call to pd.read_csv in main.py to the appropriate location.
• Run main.py to generate non-deep learning models, according to the configuration settings defined at the top of the script.
• Run keras-cnn.py to train a convolutional neural network with 2 layers of convolution and pooling.

Repository Structure

• bin
  • processing: Contains methods to process and normalize text, including part-of-speech classifier
  • modeling: Contains methods to generate non-deep learning models
  • main.py: Entry point for non-deep learning methods. Configurable settings include:
    • bool to perform PCA or not before analysis, and int number of components
    • bool to convert 3-classes to binary classification problem
    • bool to force performing text normalization
    • bool to determine what type of model to generate
    • str description of run for metrics dataframe identification
    • float between 0 and 1 to define test size
    • int to serve as random seed for train-test splits for reproducibility
  • keras-cnn.py: Convolutional neural network to perform classification
• data: Dump of data, including processed dataframes, word dictionary to serve as reference, etc.
  • models: Generated serialized models
    • v1-nonstratified: Nonstratified sampling
    • v2-stratified: Stratified sampling
    • v3-binary_stratified-class02: Stratified sampling with class 1 and class 0 combined
    • v4-pca_stratified-50components: Stratified sampling with PCA to 50 components
    • v5-pca_stratified-250components: Stratified sampling with PCA to 250 components
  • normalize-checkpoints: Will be generated during text normalization steps, containing serialized dataframes before and after most time-consuming steps
• docs: Documentation associated with projects
• img: Generated images of model results

Dependencies

• sklearn
• pandas
• numpy
• nltk
• keras
• h5py

If I Could Do It Over Again

• Save the TfidfVectorizer() objects fitted to my data
• Design with class imbalance in mind from the beginning
• Much better repository structure and organization