ToxiShield: Promoting Inclusive Developer Communication through Real-Time Toxicity Filtering

A research team has introduced ToxiShield, a novel browser extension designed to combat toxic communication in software development by filtering GitHub pull request comments in real-time. The tool addresses a critical gap in developer tools by providing immediate intervention during code reviews, where harsh feedback can undermine team productivity. Built with three specialized AI modules, ToxiShield first identifies toxicity using a BERT-based model trained on 38,761 code review samples, achieving impressive 98% accuracy and a 97% F1-score. For context, the Communication Coach module uses a prompt-tuned Claude 3.5 Sonnet model to categorize toxicity types with detailed reasoning, while the Reframer employs a fine-tuned Llama 3.2 model to rewrite toxic comments constructively.

The Reframer module, trained on 10,120 code review comments, demonstrates strong performance with 95.27% style transfer accuracy and 97.03% fluency, effectively transforming hostile remarks into professional feedback. The system represents a significant advancement over previous research that focused only on detection, now offering real-time remediation. In human evaluations with 10 participants using the Technology Acceptance Model, ToxiShield showed high perceived usefulness and ease of adoption. By integrating directly into developers' existing GitHub workflow, this tool sets a new benchmark for promoting inclusive communication in open-source communities and professional software engineering teams alike.