Toxic-Comment-Detector

This project applies classification models with the aim of automating the detection of toxic comments on social media.

Modeling:

Three main Machine Learning classification models are used:

• Multinomial Naive Bayes
• Logistic Regression
• Support Vector Machine (SVM)

According to the results, Naive Bayes classifier with TfidfVectorizer came with the best f1 score and recall: 0,85 and 0.88, respectively. Because of this reason, it was used as the base model of the web application.

Full results:

Models	precision	recall	f1 score
Multinomial NB	0.82	0.88	0.85
Logistic Regression	0.82	0.80	0.81
SVM	0.82	0.83	0.83

Web application:

To make the web app, I used Streamlit to turn the python script into an app. For the final deployment of an interactive model, I used HuggingFace's Spaces.

You can check it out and try it out here.

Screenshot of the web page:

Data:

Data for this project was obtained from the GitHub repository of SurgeAI, a powerful NLP data labeling platform.

Data Description:

• Dataset consists of 1000 English examples gathered from different popular social media platform.
• Each 500 examples in it are labeled as "toxic" and "non-toxic", respectively.

• Columns:

  • text: the text of the social media comment
  • is_toxic: whether the text is toxic or not

References:

• The script for the web application was inspired by Sidney Kung even though later on it was heavily changed according to the needs of this project. Hence, the link to her personal website: https://www.sidneykung.com/
• SurgeAI Dataset GitHub Repository Link: https://github.com/surge-ai/toxicity