The rapid growth of data and its usage by Artificial Intelligence applications leads to heightened concerns about data privacy. Researchers often need to analyze datasets that contain personal information, sometimes paired with sensitive attributes such as medical records or political views. To support such analyses without exposing identifiable content, the Scientific Software Center (SSC) of Heidelberg University developed the mailcom package for pseudonymization. This capability is especially important when employing web-hosted Large Language Models for downstream analysis. As a use case, we applied mailcom to a multilingual email corpus in Spanish, French, and Portuguese contributed by multiple donors as part of a pilot study, in collaboration with the research group of Sybille Große (Department of Romance Studies, Heidelberg University). To protect donor privacy, sensitive information such as names, email addresses, and numbers is extracted and pseudonymized. The package processes text from email subjects and bodies in eml and html formats, as well as from csv rows, making it applicable to a wide range of textual data beyond email. mailcom is built entirely on open-source libraries and is designed for configurability and extensibility. Its core features are: (i) language identification, (ii) named-entity recognition, (iii) extraction of temporal expressions, and (iv) de-identifying sensitive data via pseudonyms. Three aforementioned languages are supported by default, with options to add further languages and change back-end libraries via configuration. We present these features in end-to-end processing pipelines using examples from our use case. The main parts include: (1) General workflow from raw text to pseudonymized output,(2) Default libraries and techniques (e.g. eml-parser, spaCy, langid, langdetect, transformers, and rule-based)(3) Mechanisms for adapting to new languages, transformer pipelines, and spaCy models with minimal effort. Since pseudonymized outputs still require human review to guarantee full anonymization, the package serves as a scalable pre-processing layer that reduces manual work while establishing a principled baseline of privacy protection. This reproducible, privacy-aware tool enables empirical research on digital text under current data-ethics and governance standards.
Le et al. (Tue,) studied this question.