The human resource (HR) domain contains various types of privacy-sensitive textual data, such as e-mail correspondence and performance appraisal. Doing research on these documents brings several challenges, one of them anonymisation. In this paper, we evaluate the current Dutch text de-identification methods for the HR domain in four steps. First, by updating one of these methods with the latest named entity recognition (NER) models. The result is that the NER model based on the CoNLL 2002 corpus in combination with the BERTje transformer give the best combination for suppressing persons (recall 0.94) and locations (recall 0.82). For suppressing gender, DEDUCE is performing best (recall 0.53). Second NER evaluation is based on both strict de-identification of entities (a person must be suppressed as a person) and third evaluation on a loose sense of de-identification (no matter what how a person is suppressed, as long it is suppressed). In the fourth and last step a new kind of NER dataset is tested for recognising job titles in tezts.

The human resource (HR) domain contains various types of privacy-sensitive textual data, such as e-mail correspondence and performance appraisal. Doing research on these documents brings several challenges, one of them anonymisation. In this paper, we evaluate the current Dutch text de-identification methods for the HR domain in three steps. First, by updating one of these methods with the latest named entity recognition (NER) models. The result is that the NER model based on the CoNLL 2002 corpus in combination with the BERTje transformer give the best combination for suppressing persons (recall 0.94) and locations (recall 0.82). For suppressing gender, DEDUCE is performing best (recall 0.53). Second NER evaluation is based on both strict de-identification of entities (a person must be suppressed as a person) and third evaluation on a loose sense of de-identification (no matter what how a person is suppressed, as long it is suppressed).

PurposeThis paper aims that privacy research is divided in distinct communities and rarely considered as a singular field, harming its disciplinary identity. The authors collected 119.810 publications and over 3 million references to perform a bibliometric domain analysis as a quantitative approach to uncover the structures within the privacy research field.Design/methodology/approachThe bibliometric domain analysis consists of a combined directed network and topic model of published privacy research. The network contains 83,159 publications and 462,633 internal references. A Latent Dirichlet allocation (LDA) topic model from the same dataset offers an additional lens on structure by classifying each publication on 36 topics with the network data. The combined outcomes of these methods are used to investigate the structural position and topical make-up of the privacy research communities.FindingsThe authors identified the research communities as well as categorised their structural positioning. Four communities form the core of privacy research: individual privacy and law, cloud computing, location data and privacy-preserving data publishing. The latter is a macro-community of data mining, anonymity metrics and differential privacy. Surrounding the core are applied communities. Further removed are communities with little influence, most notably the medical communities that make up 14.4% of the network. The topic model shows system design as a potentially latent community. Noteworthy is the absence of a centralised body of knowledge on organisational privacy management.Originality/valueThis is the first in-depth, quantitative mapping study of all privacy research.

We explore the use case of question answering (QA) by a contact centre for 130,000 Dutch government employees in the domain of questions about human resources (HR). HR questions can be answered using personnel files or general documentation, with the latter being the focus of the current research. We created a Dutch HR QA dataset with over 300 questions in the format of the Squad 2.0 dataset, which distinguishes between answerable and unanswerable questions. We applied various BERT-based models, either directly or after finetuning on the new dataset. The F1-scores reached 0.47 for unanswerable questions and 1.0 for answerable questions depending on the topic; however, large variations in scores were observed. We conclude more data are needed to further improve the performance of this task.

This paper introduces a novel method to predict when a Google translation is better than other machine translations (MT) in Dutch. Instead of considering fidelity, this approach considers fluency and readability indicators for when Google ranked best. This research explores an alternative approach in the field of quality estimation. The paper contributes by publishing a dataset with sentences from English to Dutch, with human-made classifications on a best-worst scale. Logistic regression shows a correlation between T-Scan output, such as readability measurements like lemma frequencies, and when Google translation was better than Azure and IBM. The last part of the results section shows the prediction possibilities. First by logistic regression and second by a generated automated machine learning model. Respectively, they have an accuracy of 0.59 and 0.61.