Redact or Keep? A Fully Local AI Cascade for Educational Dialogue De-Identification

Short summary

Researchers propose a fully local AI cascade system for de-identifying educational dialogue that preserves curriculum content while removing personally identifiable information. Their approach achieves 95.8% accuracy on a single laptop, outperforming LLM-based baselines (76.7%) and commercial APIs (70.6%). This suggests that specialized problem formulation matters more than raw model scale for privacy-critical tasks.

• •Local cascade framework for educational data de-identification preserves curriculum terms while removing PII
• •Achieves 95.8% F1 on a single laptop, outperforming cloud-based and commercial approaches
• •Demonstrates that specialized problem formulation beats larger models for privacy-critical tasks

Generated with AI, which can make mistakes.