ajaykrishnan23 / Content-Violation-Detection

Models built to detect and tag content that violates social media guidelines.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Content-Violation-Detection

Problem Statement

Cyber bullying involves posting and sharing wrong, private, negative and harmful information about victims. With the world under quarantine, this is only going to become worse.

Prerequisites

  • Python (3.7 Recommended)
  • For installation of PyTorch, refer their official website.

Image Classification

Dataset: NSFW Image Classification Dataset
The classification model has been pretrained using Self Supervised Learning using only unlabeled Imagenet data(Pretrained model and script to be released soon :p). The model is then trained with a linear classifier over the above mentioned data over 5 categories:

  1. pornography - Nudes and pornography images
  2. hentai - Hentai images, but also includes pornographic drawings
  3. sexually_provocative - Sexually explicit images, but not pornography. Think semi-nude photos, playboy, bikini, beach volleyball, etc. Considered acceptable by most public social media platforms.
  4. neutral - Safe for work neutral images of everyday things and people
  5. drawing - Safe for work drawings (including anime)

This achieved an accuracy of 83% using purely unlabaled data and a single linear layer.

Text Toxicity Prediction

The text classification model is built to predict the toxicity of texts to pre-emptively prevent any occurrence of cyberbullying and harassment before they tend to occur. We're using BERT to overcome the current challenges including understanding the context of text so as to detect sarcasm and cultural references, as it uses Stacked Transformer Encoders and Self-Attention Mechanism to understand the relationship between words and sentences, the context from a given sentence. A combination of datasets have been used for this purpose.

Text_Input: I want to drug and rape her 
======================
Toxic: 0.987 
Severe_Toxic: 0.053 
Obscene: 0.100 
Threat 0.745 
Insult: 0.124 
Identity_Hate: 0.019 
======================
Result: Extremely Toxic as classified as Threat, Toxic 
Action: Text has been blocked. 

Referred to lonePatient for training my BERT model

License

MIT © ajaykrishnan23



Stay Fab ❤️

About

Models built to detect and tag content that violates social media guidelines.

License:MIT License


Languages

Language:Jupyter Notebook 86.2%Language:Python 13.8%