Modeling spread of curse words on a social network

Abstract Social media sites provide an online discussion forum for exchange of ideas and opinions. Unfortunately, on- line harassment presents a common problem in which verbal attacks (e.g., via the use of curse words) are propagated on a social networking platform. Modeling the spread of such curse words with an online, social network is a challenging issue. This paper describes a statistical and deep learning approach to this prob- lem. First we present a modeling of curse word ut- terance at any node in the social network as a Hawkes point process (which unlike the Poisson point process is self-excitatory) and using a LSTMs (Long Short Term Memory to capture self-excitation using recurrent con- nections). Second, we present an extension of these models to capture the social network structure (using Graph-Hawkes and Graph-LSTM), wherein utterances by near by nodes in the social network influence the node’s utterances. A detailed experimental evaluation is conducted on a Twitter dataset and with simula- tion data (using parameters learnt from the Twitter dataset). The results show that the proposed Graph- based approaches can improve the performance over the Node-centric counterparts by 30%, demonstrating that immediate neighbor ties influence dissemination in the network.
Authors
  • Mudhakar Srivatsa (IBM US)
  • Diane Felmlee (PSU)
  • Roger Whitaker (Cardiff)
  • Supriyo Chakraborty (IBM US)
  • Ceryl Giammanco (ARL)
Date Sep-2020
Venue 4th Annual Fall Meeting of the DAIS ITA, 2020