Towards Privacy Preservation in AI Summarization: Balancing Privacy and Completeness

ACL ARR 2024 December Submission1282 Authors

Abstract: With the rapid integration of AI in virtual meeting platforms, automatic summarization has become essential for productivity across sectors. While text summarization has seen significant progress, dialogue-based summarization remains underexplored, with efforts largely focusing on improving quality and addressing domain adaptation. Privacy concerns, however, are often neglected, exposing sensitive information, particularly in critical settings like healthcare, finance, and legal interactions. This paper introduces a privacy-sensitive taxonomy addressing diverse scenarios and explores strategies to safeguard privacy in AI-generated summaries. Our hybrid approach combines rule-based and learning-based techniques to address direct and indirect privacy threats while maintaining content accuracy. Using a specialized dataset curated around our taxonomy, we fine-tuned large language models and evaluated them with human and automated metrics, including Privacy and Completeness Scores. The results demonstrate the effectiveness of these models in mitigating privacy risks, offering a strong foundation for advancing privacy-preserving AI technologies while balancing privacy and completeness.

Paper Type: Long

Research Area: Summarization

Research Area Keywords: Summarization, Resources and Evaluation, Ethics, Bias, and Fairness, Dialogue and Interactive Systems, Generation, Human-Centered NLP, Machine Learning for NLP, NLP Applications

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Data resources

Languages Studied: English

Submission Number: 1282