Upgrade to Pro — share decks privately, control downloads, hide ads and more …

ICDE_21_full_presentation

 ICDE_21_full_presentation

Paper Presentation for ICDE 2021

4af679ea7716884dc09bf8b42488bfbb?s=128

_themessier

July 13, 2021
Tweet

Transcript

  1. Hate is the New Infodemic: A Topic-aware Modeling of Hate

    Speech Diffusion on Twitter Contributors: Sarah, Subhabrata, Sakshi, Chhavi, Dr. Vikram, Dr. Amitava, Dr. Tanmoy
  2. • Hate speech detection has become a much needed/explored area

    of research. • How different topics can lead to generation and spread of hate speech in a user network is underexplored. • We aim to determine how a hateful tweet diffuses via retweet and which users are likely to generate hate in the first place. Motivation
  3. Introduction

  4. Approach • Crawl a large-scale Twitter dataset of tweets, retweets,

    user activity history, and follower networks. (spanning 20+ topics) • Manually annotated a total of 17k tweets and trained a Hate Detection model for our dataset. • Additionally crawled online news articles published during the timeframe of Twitter dataset. • We design 2 models one for determining generation of hate and other for diffusion of hate in the follow network.
  5. Observations from Tweets Data

  6. Topic Wise Statistics of Tweets Data

  7. Hate Distribution of Different Users on Different Hashtags

  8. Modeling

  9. Proposed Models: Hate Generation

  10. Results for Hate Generation Model

  11. Proposed Models: Hate Diffusion RETINA Model

  12. Proposed Models: Hate Diffusion RETINA Model • Static retweeter prediction:

    ∆t is ∞ (i.e., all the retweeters irrespective of their retweet time) • Dynamic retweeter prediction: Where we predict on successive time intervals
  13. Observations from Tweets Data

  14. Proposed Models: Hate Diffusion RETINA Model Exogenous attention Dynamic Static

  15. Observations

  16. Results comparing Diffusion Predictions Signify models without exogenous influence

  17. Comparison of Models when Root-tweet is Hateful(dark shade) and Non-Hateful

    (lighter)
  18. Variation in performance of RETINA-S with cascade size

  19. Conclusion: • It is a very first attempt to study

    hate generation and spread on twitter. • We developed multiple supervised models powered by rich feature representation to predict the probability of any given user tweeting something hateful. • We proposed RETINA, a neural framework exploiting extra-Twitter information (in terms of news) with attention mechanism for predicting potential retweeters for any given tweet. • Reference Code: https://github.com/LCS2-IIITD/RETINA • Paper: https://arxiv.org/abs/2010.04377