The Multilingual Anonymisation Toolkit for Public Administrations (MAPA) Project

The Multilingual Anonymisation Toolkit for Public Administrations (MAPA) Project ¯E. Ajausks†, V. Arranz‡, L. Bi´e*, A. Cerd`a-i-Cuc´o*, K. Choukri‡, M. Cuadros§, H. Degroote*, A. Estela*, T. Etchegoyhen§, M. Garc´ıa-Mart´ınez*, A. Garc´ıa-Pablos§, M. Herranz*, A. Kohan*, M. Melero**, M. Rosner¶, R. Rozis†, P. Paroubek⋄, A. Vasil¸evskis†, P. Zweigenbaum⋄ ∗ †Tilde {eriks.ajausks,roberts.rozis,arturs.vasilevskis}@tilde.lv ‡ELDA/ELRA {arranz,choukri}@elda.org *Pangeanic - PangeaMT {l.bie,a.cerda,h.degroote,a.estela,m.garcia,m.herranz,a.kohan}@pangeanic.com **Barcelona Supercomputing Center maite.melero@bsc.es ⋄ Universit´e Paris-Saclay, CNRS, LIMSI {pap,pz}@limsi.fr ¶University of Malta mike.rosner@um.edu.mt §Vicomtech {agarciap,mcuadros,tetchegoyhen}@vicomtech.com Abstract We describe the MAPA project, funded un- der the Connecting Europe Facility pro- gramme, whose goal is the development of an open-source de-identification toolkit for all official European Union languages. It will be developed since January 2020 until December 2021. 1 Introduction De-identification may provide the means to share language data while also protecting private or sen- sitive data by spotting then deleting, obfuscating, pseudoymising or encrypting personally identifi- able information. De-identification is typically performed for the purpose of protecting an individ- ual’s private activities while maintaining the use- fulness of the gathered data for research and de- velopment purposes. The Multilingual Anonymisation toolkit for Public Administrations (MAPA) project aims to leverage natural language processing tools to de- velop an open-source toolkit for effective and re- liable text de-identification, focusing on the med- ical and legal domains. The project is funded by the Connecting Europe Facility (CEF) programme, under grant No A2019/1927065, and will run from January 2020 until December 2021. The toolkit developed by the MAPA partners (Pangeanic1, Tilde2, CNRS3, ELDA4, Univer- ∗All authors have contributed equally to this work. ∗c⃝ 2020 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. 1https://pangeamt.com/ 2https://www.tilde.com/ 3http://www.cnrs.fr/ 4http://www.elra.info/en/ sity of Malta5, Vicomtech6 and SEAD7) will ad- dress all official EU languages, including under- resourced ones such as Latvian, Lithuanian, Esto- nian, Slovenian and Croatian, and severely under- resourced ones such as Irish and Maltese. As a part of the project, a connection to eTrans- lation,8 an online machine translation service pro- vided by the European Commission, will be es- tablished to foster the provision of multilingual datasets by public administrations that may in turn improve the coverage and quality of machine translation systems. 2 Approach At its core, the MAPA anonymisation toolkit will rely on Named Entity Recognition and Classifica- tion (NERC) techniques using neural networks and deep learning techniques. The latest deep learn- ing architectures and the availability of pre-trained multilingual language models, such as BERT (De- vlin et al., 2019) have pushed the state of the art in NERC to new levels of performance. In addition, thanks to the transfer learning capa- bilities shown by this type of deep learning models, new systems can be trained using smaller datasets of manually labelled data, and the knowledge ac- quired for a given domain or language can be re- used in a cross-domain or cross-language setting (Garc´ıa-Pablos et al., 2020). MAPA will leverage the most innovative technology to provide robust models for the 24 official European languages, trained to detect …

Content truncated to 500 words