Content Moderation

Objective

Detection of inappropriate content is a mandatory requirement for most content platforms. Usually we refer to the detection of NOK (non-OK) images or text from OK content.

Below we show how DeepDetect is used to filter both text and images by leaders of content moderation and digital marketing companies.

Image moderation

Typical applications:

  • NSFW detection, setup the existing NSFW model
  • Detection of message passing in images, use the text detection model
  • Detection of eyes, mouth, nose to make sure faces are facing the camera
  • Detection of drinks, violence, …

Examples

Text detection in portrait image.

Train your own model

We recommend training an image classification model with the platform. This is fairly simple:

  • Create two directories, one with the OK images, one with the NOK images. Thanks to the platform [pre-trained models](), you may not require more than a few hundreds images per directory

  • Train your model with the platform by following the image training instructions

  • Test and analyze results with the platform

  • Put into production by using the platform or [deploying with the DeepDetect server]().

Text moderation

Training

We recommend training a text classification model with the platform using a character-based neural network. This is a good choice for dealing with user-generated content since it can work around variations in spelling etc…

Building the model is fairly simple:

  • Create two directories, one with the OK text samples, one with the NOK samples. You may need a few hundreds to few thousands samples per directory, up to hundreds of thousands for best results if your problem is difficult.

  • Train your model with the platform by following the text training instructions

  • Test and analyze results with the platform

  • Put into production by using the platform or [deploying with the DeepDetect server]().

Related