Classification of inappropriate language

Introduction

The goal of our study is to compare different ML and DL models on classifying text as appropriate or inappropriate language. The ultimate goal is to classify in-game chats, but our study will mainly focus on social media texts.

What is Inappropriate Language ?

We define, for the following work, an Inappropriate Language as a word or a sentence that contains any of the following:

• Profanity - This includes any language that is considered vulgar, offensive, or obscene. This can include swear words, sexual language, and derogatory terms.
• Hate speech - Hate speech is language that is intended to demean, discriminate against, or incite violence or hatred towards a particular group of people based on their race, ethnicity, gender, religion, sexual orientation, or other characteristic.
• Insults - This includes any language that is intended to insult or belittle someone else. This can include name-calling, personal attacks, or derogatory comments about someone's appearance, abilities, or personality.
• Threats Threats are language that is intended to intimidate or harm another person. This can include physical threats, verbal abuse, or intimidation.

Models tested

• Decision Tree
  • Source: sklearn
  • Parameters: None
• Random Forest
  • Source: sklearn
  • Parameters: None
• SVC (SVM)
  • Source: sklearn
  • Parameters: None
• LSTM
  • Source: Tensorflow
  • LSTM -> Dense layer (In -> Out (2))
• Transformer (mobile BERT / distil BERT)
  • Source: Huggingface
  • Zero Shot Classification

Our Dataset

Our dataset is a merge of multiple social media Hate Speech datasets.

Columns:

• Text: text to analyze
• Class:
  • 0 - Appropriate
  • 1 - Inappropriate

Data Sources:

• Hate Speech and Offensive Language (https://github.com/t-davidson/hate-speech-and-offensive-language/raw/master/data/labeled_data.csv) [tweets]
• Measuring Hate Speech (https://huggingface.co/datasets/ucberkeley-dlab/measuring-hate-speech) [social media Reddit Twitter and other]
• https://www.kaggle.com/competitions/detecting-insults-in-social-commentary/
• Jibes and Delights https://github.com/ravsodhi/jibes-and-delights (Insults and Compliments targeting INDIVIDUALS) https://aclanthology.org/2021.woah-1.14.pdf
• https://www.kaggle.com/competitions/jigsaw-toxic-comment-classification-challenge/

Details:

• English only
• From social media comments