Neural Text Sanitization with Explicit Measures of Privacy Risk

Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 217–229 November 20–23, 2022. ©2022 Association for Computational Linguistics 217 Neural Text Sanitization with Explicit Measures of Privacy Risk Anthi Papadopoulou Language Technology Group, University of Oslo anthip@ifi.uio.no Yunhao Yu École Polytechnique yunhao.yu@polytechnique.edu Pierre Lison Norwegian Computing Center plison@nr.no Lilja Øvrelid Language Technology Group, University of Oslo liljao@ifi.uio.no Abstract We present a novel approach for text sanitiza- tion, which is the task of editing a document to mask all (direct and indirect) personal iden- tifiers and thereby conceal the identity of the individuals(s) mentioned in the text. In con- trast to previous work, the approach relies on explicit measures of privacy risk, making it pos- sible to explicitly control the trade-off between privacy protection and data utility. The approach proceeds in three steps. A neural, privacy-enhanced entity recognizer is first em- ployed to detect and classify potential personal identifiers. We then determine which entities, or combination of entities, are likely to pose a re-identification risk through a range of pri- vacy risk assessment measures. We present three such measures of privacy risk, respec- tively based on (1) span probabilities derived from a BERT language model, (2) web search queries and (3) a classifier trained on labelled data. Finally, a linear optimization solver de- cides which entities to mask to minimize the semantic loss while simultaneously ensuring that the estimated privacy risk remains under a given threshold. We evaluate the approach both in the absence and presence of manually anno- tated data. Our results highlight the potential of the approach, as well as issues specific types of personal data can introduce to the process. 1 Introduction Personal data, also known as Personally Identifi- able Information (PII), often abound in text docu- ments, from emails to patient records, court judg- ments, interview transcripts or customer service chats. Protecting the privacy of the individuals mentioned in those documents is an important task, particularly for sensitive texts which might disclose confidential information such as health status, reli- gious beliefs, ethnicity or sex life. It is, however, possible to apply privacy- enhancing techniques such as text sanitization to conceal the identity of those individuals from the texts, and thereby make it easier to share data to third parties, in particular for the purpose of scien- tific research or statistical analysis. The goal of text sanitization is to transform a document through edit operations such as hiding particular text spans or replacing them by more general values. Although complete anonymization compliant with data pri- vacy frameworks such as the General Data Protec- tion Regulation (GDPR, 2016) has been shown to be very difficult to achieve in practice (Weitzen- boeck et al., 2022), text sanitization can substan- tially enhance the level of privacy protection while simultaneously retaining most of the semantic con- tent expressed in the documents. Existing work on text sanitization has primarily focused on masking …

Content truncated to 500 words