# Which anonymization technique is best for which NLP task? -- It depends. A Systematic Study on Clinical Text Processing

arXiv: https://arxiv.org/abs/2209.00262 · HTML: https://mlsanity.com/paper/2209.00262/

## Key facts

- Cited by 4 papers in the map (data through 14 Oct 2026).
- Builds on 4 of its 34 references that are in the map.
- First posted on arXiv 1 Sep 2022.

## Builds on (most cited first, 4 of 4)

- [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://mlsanity.com/paper/1810.04805/) (2018)
- [Publicly Available Clinical BERT Embeddings](https://mlsanity.com/paper/1904.03323/) (2019)
- [Phenotyping of Clinical Notes with Improved Document Classification Models Using Contextualized Neural Language Models](https://mlsanity.com/paper/1910.13664/) (2019)
- [Closing the Gap: Joint De-Identification and Concept Extraction in the Clinical Domain](https://mlsanity.com/paper/2005.09397/) (2020)

## Built on by (most cited first, 4 of 4)

- [DeID-GPT: Zero-shot Medical Text De-Identification by GPT-4](https://mlsanity.com/paper/2303.11032/) (2023)
- [CAPID: Context-Aware PII Detection for Question-Answering Systems](https://mlsanity.com/paper/2602.10074/) (2026)
- [Subject-level Inference for Realistic Text Anonymization Evaluation](https://mlsanity.com/paper/2604.21211/) (2026)
- [Silencing the Risk, Not the Whistle: A Semi-automated Text Sanitization Tool for Mitigating the Risk of Whistleblower Re-Identification](https://mlsanity.com/paper/2405.01097/) (2024)

## Abstract (arXiv)

Clinical text processing has gained more and more attention in recent years. The access to sensitive patient data, on the other hand, is still a big challenge, as text cannot be shared without legal hurdles and without removing personal information. There are many techniques to modify or remove patient related information, each with different strengths. This paper investigates the influence of different anonymization techniques on the performance of ML models using multiple datasets corresponding to five different NLP tasks. Several learnings and recommendations are presented. This work confirms that particularly stronger anonymization techniques lead to a significant drop of performance. In addition to that, most of the presented techniques are not secure against a re-identification attack based on similarity search.