NLP_Classification

Table of contents

• General info
• Project 1: SMS Spam detection
• Project 2: Fake news detection
• Project 3: Toxic comments detection

General info

In this repository you'll learn how to classify text based on multiple models such as: Naive Baye, LSTM, Transformers (Bert) and One vs All

Project 1 - SMS Spam detection

Using Naive Baye we have 98,56% accuracy ! (This is why companies use this algorithm for spams classification)

Using LSTM we have 97.72% accuracy, just like Naive Baye.

And using BERT we have 97.61% accuracy.

Project 2 - Fake news detection

Using Naive Baye we have 62.92% accuracy and 60.00% of True news are misclassified as Fake news...

Using LSTM we have 67.16% accuracy and 80.90% of True news are misclassified as Fake news...

Using BERT we have 65.20% accuracy, the algorithm doesn't learn much.

Project 3 - Toxic comments detection

For this last project I am using the dataset from kaggle: Jigsaw Unintended Bias in Toxicity Classification

Using Naive Baye we have:

• 86.95% accuracy for normal data
• 33.39% accuracy for severe toxic data
• 19.23% accuracy for obscene data
• 35.06% accuracy for insults

For a total of 72.74% Accuracy.

This shows the limits of Naive Baye algorithm.

Using LSTM we have:

• 91.18% accuracy for normal data
• 34.59% accuracy for severe toxic data
• 48.65% accuracy for obscene data
• 46.97% accuracy for insults

For a total of 76.78% Accuracy.

This is a great improvement from Naive Bayes.

Using BERT we have 74% accuracy in total.