De-Identification of Emails: Pseudonymizing Privacy-Sensitive Data in a German Email Corpus

Proceedings of Recent Advances in Natural Language Processing, pages 259–269, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_030 259 De-Identification of Emails: Pseudonymizing Privacy-Sensitive Data in a German Email Corpus Elisabeth Eder Ulrike Krieg-Holz Institut f¨ur Germanistik Alpen-Adria-Universit¨at Klagenfurt, Klagenfurt, Austria {elisabeth.eder | ulrike.krieg-holz}@aau.at Udo Hahn Jena University Language & Information Engineering (JULIE) Lab Friedrich-Schiller-Universit¨at Jena, Jena, Germany udo.hahn@uni-jena.de Abstract We deal with the pseudonymization of those stretches of text in emails that might allow to identify real individual persons. This task is decomposed into two steps. First, named entities carrying privacy- sensitive information (e.g., names of per- sons, locations, phone numbers or dates) are identified, and, second, these privacy- bearing entities are replaced by syntheti- cally generated surrogates (e.g., a person originally named ‘John Doe’ is renamed as ‘Bill Powers’). We describe a system architecture for surrogate generation and evaluate our approach on CODEALLTAG, a German email corpus. 1 Introduction With the advent and rapidly increasing adoption of electronic interaction platforms, the communi- cation patterns of modern societies have changed fundamentally. We observe an unprecedented up- surge of digitally transmitted private communi- cation and exploding volumes of so-called user- generated contents (UGC). As a major charac- teristic of these new communication habits, a sender’s individual email, post, comment, tweet is distributed to an often (very) large number of addressees—the recipients of an email, other blog- gers, friends or followers in social media plat- forms, etc.. Hence, hitherto private communica- tion becomes intentionally public. Responding to these changes, digital (social) media communication has become a focus of re- search in NLP. Yet there seems to be a lack of awareness among NLP researchers that the ex- ploitation of natural language data from such elec- tronic communication channels, whether for com- mercial, administrative or academic purposes, has to comply with binding legal regulations (Wilson et al., 2016). Dependent on each country’s law system, different rules for privacy protection in raw text data are enforced (cf., e.g., two recent analyses for the US (Mulligan et al., 2019) and the EU (Hoofnagle et al., 2019)). Even privacy-breach incidents in a legal grey zone can be harmful for the actors involved (including NLP researchers). This is evidenced dramatically in the so-called AOL search data leak.1 In August of 2006, Amer- ican Online (AOL) made a large query log collec- tion freely accessible on the Internet for a limited time. The data were extracted over three months from their search engine to support academic re- search. The collection represented 650k users is- suing 20 million queries without any significant anonymization. The result of this release, among others, was the disclosure of private information for a number of AOL users. The most troubling aspect of the data leak was the ease by which sin- gle unique individuals could be pinpointed in the logs. Even ignoring the existence of social secu- rity, drive license, and credit card numbers, the New York Times demonstrated the ability to deter- mine the identity of a real user.2 The outline of this incident …

Content truncated to 500 words