Automating Content Moderation in Windy.app Using Large Language Models

Sammanfattning

The proliferation of user-generated content on platforms designed for social interaction, including Windy.app, poses significant challenges for manual content moderation, impacting scalability, consistency, and timely intervention. This thesis addresses the problem of enhancing content moderation efficiency and effectiveness by leveraging Large Language Models (LLMs). The research investigates the development of a structured moderation taxonomy, the selection of optimal LLMs based on performance and production readiness, and the design of an LLM-powered moderation pipeline incorporating essential human oversight. The study involved creating a bespoke six-category moderation taxonomy for Windy.app, operationalized through prompt engineering. A comprehensive empirical evaluation of fifteen LLMs was conducted, identifying Google's Gemini 2.0 Flash Lite as the most suitable due to its balance of accuracy, coverage, and operational viability. Based on these findings, a proof-of-concept (PoC) system called "WindyGuard" was designed and implemented. This system features a microservice architecture, asynchronous communication, and a Human-on-the-Loop workflow, enabling automated handling of straightforward content and human review for complex cases or appeals. The results demonstrate the feasibility of an LLM-based approach to significantly augment content moderation. The PoC system successfully implements the designed taxonomy and pipeline, showcasing a practical method for balancing automation with human judgment. The selected LLM performed effectively on a standardized benchmark, indicating its potential for real-world application. This thesis provides a framework and a functional prototype for Windy.app to develop a more scalable, consistent, and insightful content moderation system. The findings and the implemented PoC are useful for platforms seeking to integrate LLMs into their moderation processes while maintaining ethical considerations and human control over nuanced decisions, contributing to safer and more respectful online communities.