Hierarchical Attention based Deep Neural Networks for Toxic Comments Classification

Straight Neural Network methods are effective for text classification tasks but better representation can be achieved by including knowledge of document structure in the model architecture. This can be understood as:

  • Not all parts of the document is relevant for understanding the content
  • Finding relevant sections in a document involves modeling the interactions of the words and not just the presence in isolation

The hierarchical attention that can be used here is:

  1. Words from sentences
  2. Sentences from comments.

The link to the Github Repository can be found here.