Multitask Learning for Hate Speech Detection

In this project, we seek to improve classification performance for abusive language detection by leveraging multitask learning on a set of tasks that appear related to abuse detection

TODO

• Modeling

  • Code MTL model

    • Multilabel prediction

  • Try setting aux task weights based on how hard tasks are on the fly

  • Try with BPE (as en encoder)

  • Try with LIWC for every task

• Datasets

  • Sarcasm (https://nlds.soe.ucsc.edu/sarcasm2)

  • Factual or Feeling based argument (https://nlds.soe.ucsc.edu/factfeel)

  • Offensive Language: Davidson et al.

  • Hate speech (Waseem & Hovy, Waseem)

  • Toxicity (https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge/data)

  • Sentiment (Hovy Sentiment gender)

  • Moral Foundations Prediction (MFTC: https://psyarxiv.com/w4f72)

  • Twitter datasets?

  • Demographic dataset: Preotiuc and Ungar.

    Race key:

    • 1: African-American

    • 2: Latinx/Hispanic

    • 3: Asian

    • 4: White

    • 5: Multiracial

    • NULL: Didn't answer/Other race (usually Native American)

• Models

  • LSTM