📄 Building the v4 of the Croatian National Corpus

Summary

The Croatian National Corpus (HNK) has not had a new release since v3 in 2013. Its texts, genre mix and annotation are therefore outdated. This work-in-progress paper describes how HNK v4 is being assembled. Unlike earlier versions, v4 will contain only IPR-cleared material and will be released for academic use under a permissive license. Sources include open-access scientific publications, legislation and texts donated by authors, and the team is negotiating with major publishers. The authors acknowledge that this restriction may make the corpus more opportunistic than balanced, but they argue that free access to a large, curated corpus outweighs strict adherence to a planned design. To support the data, they built CorpRepo, a web application that manages metadata, harvesting, text extraction, cleaning and versioning across many git repositories. On the annotation side, v4 will replace the MULTEXT-East tagset with Universal Dependencies (UD) and add layers beyond morphosyntax. Release is planned for late 2026 or early 2027 through the HR-CLARIN concordancer and Sketch Engine.

Key points

• The target size is at least 1 Gw. Under the tentative composition plan, newspapers and magazines make up roughly half, with smaller shares for legislative/public, academic, literary and other texts.
• The Croatian Scientific Corpus subcorpus is estimated at over 600 Mw. It draws on Hrčak (572 journals, more than 322,000 full-text papers) and on theses in the Dabar repositories and the National and University Library. Its metadata carry a domain/field/branch classification, so discipline-specific subcorpora can be built from it.
• Legal texts come from Narodne novine and the Central Catalogue of Official Documents. They extend the earlier Croatian MARCELL Legislative Subcorpus (102 Mw) with material published since then.
• CorpRepo works as follows:
  • It harvests metadata via OAI-PMH.
  • It extracts text with GROBID into XML/TEI.
  • It version-tracks every file-system change in git.
  • It supports semi-automatic and manual cleaning, which matters because PDF extraction, especially from Hrčak, is very noisy.
  • It generates corpora according to chosen parameters.
• HNK v3 was annotated only morphosyntactically, using the MULTEXT-East v4 tagset. v4 adopts UD for readability and cross-linguistic comparability. It will also add precomputed query attributes, such as periphrastic verb forms and traditional lemmatization, to make searching easier.