Request Correction

Fill-in-the-Blanks: Automatic Generation and Evaluation of Language Models' Pseudonyms for English and Swedish Texts

While considerable effort has gone into developing solutions for detecting Personally Identifiable Information (PII) in linguistic data, less research has gone into automating the generation of appropriate pseudonyms and developing evaluation methods, both relevant for the creation of privacy-friendly language resources. We conduct pilot experiments using Masked and Generative Large Language Models to generate predictions for redacted PII-spans in a cloze-like fashion for English legal texts and parallel news articles in Swedish and English. Furthermore, we explore metrics for automatic evaluation of the generated pseudonyms in the legal data, and investigate the effect of part-of-speech constraints on performance. For the parallel, multilingual data, we contribute our manual PII-annotation and conduct a fine-grained error analysis across two of our pseudonym generation methods and a baseline. Our results illustrate the complexity of pseudonym evaluation and the particular challenge of automatic, at-scale evaluation as well as the models’ tendency to predict prototypical and even stereotypical answers.