This notebook permit to make a selection model for text classification (Binary and multiclass) using Tensorflow 2.x and Keras. The goal is to compare with different metrics machine learning and deep learning algorithms. It is configure to make classification with French and English texts.

This work has been presented in Model Selection in Text Classification published in Towards Data Science

The models implemented in the notebook for model selection are : Multinomial Naive Bayes, logistic Regression, SVM, k-NN, Stochastic Gradient Descent, Gradient Boosting, XGBoost (Early stopping are implemented to stop the training to avoid overfitting) Adaboost, Catboost, LightGBM and ExtraTrees Classifier.

The models implemented are: Shallow Network, Deep Neaural Network, RNN, LSTM, CNN, GRU, CNN-LSTM, CNN-GRU, Bidirectional RNN, Bidirectional LSTM, Bidirectional GRU, RCNN and Transformers (Early stopping are implemented to stop the training to avoid overfitting).

• Module importation
• Functions for metrics
• Parameters
  • Here you'll choose the column name of the text to be classified and the name of the label column
• List of Models
  • This variables are all boolean and permit to configure the type of models you want to test in the model selection
  • save_results is for the saving the finl dataframe containing the values of all metrics
  • lang is the parameter to detect the language of the data (API Google) if False, Engish is the default
  • sample is the parameter to choose a sample of the data (Default 5000 raws)
  • pre_trained is the parameter to use pretrained fastText model in the deep learning models
• List of Metrics for the Model Selection
  • Contains the metrics considered for the model selection
  • They will be converted with make_scorer (sklearn) for the cross_validate function (sklearn)
• Sand Box to Load Data
  • Here you will load your data and make manipulations on them to prepare them for the model selection
• Start pipeline
  • If lang is True this part will detect the language of the text and select the most present in number of raws
• Prepare data for ML Classic
  • Select a random sample of data (default 5000 raws) if sample is True
  • Select stopwords file in function of the language
  • Create a new column for text without stopwords
• Class Weights
  • Estimate the weight of each class present in the data and determine if the data is balanced or imbalanced
  • Work in progress, if the dataset is imbalanced create generic data with Smothe or Adasyn
• Machine learning
  • Save labels
  • Create empty dataframe to store the results of each metric for each model on each fold
  • Compute One-hot encoding
  • Compute TF-IDF
  • Compute TF-IDF n-grams (2, 3)
  • Compute TF-IDF n-grams characters (2, 3)
  • Load pretrained model fastText
  • Pad sentences in integers word vectors
• All machine learning models
  • report () function based on cross_validate function to compute the metrics
• All deep learning models
  • cross_validate_NN() custom function for cross-validation (Stratified k-fold) and computed metrics
• Save the results if save_results if True

Next steps:

• Use compressed layer with TensorNet like this post
• Distributed Neural Networks
• GridSearch for Hyperparameters tuning
• Transform the notebook in script with dictionnary of models to test

Your contributions are always welcome!

If you want to contribute to this list (please do), send me a pull request or contact me @chris or chris