Local AI Cascade Excels at De-Identifying Educational Dialogue

Educational transcripts are rich resources for research but contain sensitive personally identifiable information (PII) often intertwined with academic content. Current de-identification methods present a dilemma: commercial large language models (LLMs) can handle contextual ambiguity but require transmitting sensitive data to external parties, while local named entity recognition (NER) systems, though privacy-preserving, tend to over-redact relevant curricular terms. To address this, a new, entirely local cascade framework has been proposed. This system redefines de-identification as a constrained privacy triage rather than broad entity recognition. It employs a "recall-first union proposer" that uses lightweight encoders and deterministic rules to initially identify a broad set of potential PII. A subsequent "context-aware reviewer" then makes a binary decision—Redact or Keep—for each candidate, leveraging surrounding dialogue and speaker roles. Evaluations on math tutoring transcripts showed the local framework's superior performance. The strongest configuration achieved a macro F1 score of 0.958, significantly outperforming a same-family LLM-only baseline (0.767) and a commercial API (0.706), all while operating entirely on a single laptop. This robust performance, particularly in handling ambiguous terms like "Riemann" (which could be a student or a mathematical concept), suggests that the problem's formulation and the system's design are more critical than the sheer scale of the underlying model for effective educational de-identification.