PIIvot: A Lightweight NLP Anonymization Framework for Question-Anchored Tutoring Dialogues

Matthew Zent, Digory Smith, Simon WoodheadEMNLP 2025

TL;DR

This paper introduces PIIvot, a novel NLP anonymization framework that leverages contextual information to improve PII detection and replacement in educational dialogues. To support research in this area, they also release QATD2k, the largest open-source real-world tutoring dataset of its kind, addressing a critical need for quality educational data.

What Problem Does This Paper Address?

Sharing sensitive, real-world educational data is crucial for advancing AI in education, but it's hampered by privacy concerns, particularly regarding Personally Identifiable Information (PII). Existing PII anonymization techniques often struggle with a trade-off between precision and recall, limiting their practical adoption. Furthermore, the scarcity of high-quality, open-source conversational educational datasets makes it difficult to train and evaluate new models. This paper tackles both issues by proposing a more effective anonymization framework and releasing a valuable new dataset.

Methods

PIIvot is built on two core principles:

  • Recall-First Potential-PII Identification: Instead of aiming for perfect PII identification upfront, PIIvot uses a "recall-first" approach. This means it's designed to identify potential PII with high recall, even if it means a slightly lower precision. This reduces the cognitive load on human annotators and simplifies the initial labeling task. They use a DeBERTa model fine-tuned on a prior dataset for this step, labeling entities like names, locations, URLs, and dates.
  • Context-Aware Surrogate Anonymization (HIPS): Once potential PII is identified, PIIvot employs a "Hidden-In-Plain-Sight" (HIPS) strategy. This involves using an LLM (specifically GPT-4) to generate contextually appropriate replacements for the identified PII. The key here is that the replacements are designed to maintain the coherence and integrity of the dialogue, ensuring that the anonymized data remains useful for downstream tasks. The LLM is prompted to preserve characteristics like gender and ethnic background where relevant, and feedback-based reprompting is used to ensure the replacements are sufficiently different from the original.

Datasets

  • QATD2k (Question-Anchored Tutoring Dialogues): This is a newly introduced, large-scale, open-source dataset of real-world tutoring dialogues collected from a UK-based learning platform. It contains over 1971 dialogues with approximately 46,249 turns, focusing on math tutoring. The dataset was curated through a rigorous filtering and downsampling process to ensure pedagogical relevance and diversity of tutoring strategies.
  • Prior datasets for NER model training: The paper mentions using a prior set of 40k labeled student-tutor utterances to fine-tune the DeBERTa NER model.

Key Insights and Ideas

  • Context is King for Anonymization: The paper strongly advocates for leveraging the context of the dialogue to improve anonymization. By understanding the surrounding text, PIIvot can generate more plausible and less disruptive replacements.
  • Potential-PII as a Stepping Stone: Reframing PII identification as a "potential-PII" labeling task significantly simplifies the annotation process and improves the feasibility of creating large, high-quality datasets.
  • The Value of Real-World Data: The QATD2k dataset highlights the limitations of synthetic tutoring data, which can lead to unrealistically long messages and potentially overfit models. Real-world data, despite its challenges, is essential for building robust educational AI.
  • Balancing Privacy and Usability: PIIvot demonstrates a practical approach to balancing the critical need for privacy with the desire for usable, shareable data, a common tension in educational data science.

Key Results

  • High PIIvot Performance: The PIIvot framework achieved a micro-averaged F1 score of 0.984 for potential-PII detection on dialogues and 0.996 on questions, outperforming more challenging PII detection tasks.
  • Minimal Impact on Text Characteristics: Anonymization with PIIvot had a negligible impact on key text characteristics of the QATD2k dataset, such as words per turn and N-gram entropy, indicating that the anonymized data retains its structural integrity.
  • QATD2k Dataset Creation: The paper successfully curated and released QATD2k, a valuable resource for researchers in educational AI, dialogue modeling, and tutoring systems.

Limitations and Future Work

  • Domain Specificity: The assumption that the content of labeled entities is insignificant might not hold true across all domains. Future work could explore improved prompting strategies or feedback mechanisms for anonymization in different contexts.
  • NER Model Limitations: The framework relies on an effective NER model, and the decision to focus on potential-PII limited the ability to assess relative improvements over existing PII detection tools.
  • Residual PII Risk: Despite extensive measures, there remains a non-zero risk of residual PII in the anonymized dataset. The authors emphasize that releasing the underlying NER models or labels alongside datasets could expose this risk.
  • Talk Move Label Generalizability: The "Getting Students to Relate" talk move label might not fully generalize to all 1:1 tutoring contexts, especially with fictional arguments in DQs.
  • Student Consent and Linkage: The decision to prioritize student privacy by removing links across tutoring sessions, while ethically sound, might impact downstream applications that rely on longitudinal student data.
  • Future Work: The authors suggest exploring LLM-generated feedback for anonymization, self-hosting LLMs for highly sensitive contexts, and preparing more in-situ datasets for benchmarking. They also encourage researchers to validate third-party APIs used in anonymization pipelines.

Summary updated

BibTeX
@inproceedings{zent2025piivot,
  author = {Matthew Zent and Digory Smith and Simon Woodhead},
  title = {{PIIvot: A Lightweight NLP Anonymization Framework for Question-Anchored Tutoring Dialogues}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}

Loading related content...

Figure