Skip to content

Latest commit

 

History

History
65 lines (46 loc) · 2.5 KB

writeup.md

File metadata and controls

65 lines (46 loc) · 2.5 KB
Raw

Detecting Toxicity with Deep Learning

Ryan Solava

Abstract

In this project, we use deep learning to determine if a social media comment is online harassment or not. The recurrent neural networks with pre-trained embedding perform the best, and outperform the baseline NLP models.

Design

Online harassment is a serious concern. The CDC reported in 2019 that 15.7% of high schoolers experienced online harassment in a 12 month period. Online platforms bare some responsibility for dealing with this issue. While some moderation can be done by hand, a way to detect if text constitutes online harassment automatically would be useful. While natural language processing could be used for this problem, we found that deep learning methods outperformed those models.

Data

The first dataset is a collection of approximately 3,500 YouTube comments. Each comment has been tagged as online harassment or not.

The second dataset is a collection of comments from a variety of platforms. There are approximately 150,000 comments, which have also been tagged as online harassment or not.

Both datasets can be found at the following link. here.

Additionally, the Google News pre-trained embedding was used for transfer learning.

Algorithms

  • The baseline model was a NLP model. TFIDF was used to create the document term matrix, and LDA was used for used for matrix decomposition. These topics were then fed into an XGBoost tree model.
  • The topics were also used in a simple feed forward neural network.
  • A simple RNN with bidirectional LSTM was trained without transfer learning.
  • A variety of RNN's were trained with transfer learning. The best performing model began with the pretrained embedding followed by an LSTM layer with 100 nodes and then two dense layers with 50 and 20 nodes.

Tools

  • Python, Pandas, and Numpy for standard data science tools
  • NLP Tools:
    • sklearn for TFIDF and LDA
    • spaCy for preprocessing
    • scattertext for visualization
  • Keras for deep learning models

Communication

A summary of my project and results can be found on my slides, and were discussed in my presentation. Also, my code and visualizations are posted to my personal GitHub. Specically, you can find a scattertext visualization showing which words are common in each type of comment. Note that this visualization contains profanity and hate speech, since these are common in online harassment.