Home - weiyinc11/HateSpeechModerationTwitch GitHub Wiki

Welcome to the HateSpeechModerationTwitch wiki!

Resources

Following the exploration of OCMP43's listed platforms (details can be found in /resources), the following platforms provided the most promise with regards to its ability to support our experimental needs such as Moderation System Customizations and a controlled lab environment:

In-depth summary and Experimental Design can be found here.

  1. Discord
  2. Reddit
  3. Twitch

Project Outline (Updated 06/20)

Phase 1. Exploration of Platforms

  • Report on the tools that these platforms provide
    • Identifying adjacent papers that experiment on platforms : Mindmap of Related Papers
    • API details and mechanics behind what can be customized
    • Utilize the white papers and transparency reports to gain a deeper insight into how these platforms are implementing autoMods
    • Highlight the keywords and policies within community guidelines of each of the platforms that seem impractical as well as most common policies that are demonstrated among these platforms explored.
      • Enforcement of these policies : how do they enforce these "We do not tolerate statements"
      • Other policies or rules around AI-generated content
  • Platforms are not supplying an LLM for access for moderators - perhaps due to Latency
    • Explain that this is the current system that platforms can afford.. Many platforms may utilize GPT and such to implement autoMod
    • Or perhaps they are not able to implement these functionality to a certain confidence

Phase 2. Utilizing NLP / LLM systems

  • Experimental Design
    • Line rate, Automated input of data / content onto these platforms
    • Narrow down into which policies we would like to investigate and focus on
  • NLP datasets and testing against baseline systems
    • Datasets that are tailored towards focused policies
    • Perhaps if these policies are not well developed in the world of NLP, what are these platforms using to inform their LLMs for content moderation

Phase 3. Testing the robustness of the systems

  • Tweaking the data inputs and attempting to identify where the LLMs are most successful or may falter in performance images.
    • Most platforms could be utilizing a streamlined API such as Perspective where the process of breaking these systems would be less essential