Using LLMs to automate moderation tasks:

  • Unbundling multiple ideas
    • Extracting distinct ideas from a single user post, and categorizing them into solutions, metrics, and barriers
    • Few shot learning did well
  • Implementing efficient argument mining
    • Identifying and extracting all distinct arguments related to ideas/answers discussed
    • Few shot learning did well, but did as well as zero-shot
  • Argument classification
    • Classifying pro or con
    • Weakest task