An Enriched Dataset of Radio Netherlands Worldwide Broadcast Documentation (1950–1966)

Repository location

The dataset is deposited in Zenodo at: https://doi.org/10.5281/zenodo.20798117

Context

Radio Netherlands Worldwide (RNW) was the Dutch international broadcasting service that operated from 1947 to 2012. This dataset covers the period 1950–1966, during which RNW significantly expanded its reach during the decolonization era, transmitting in Dutch, English, Indonesian, and Spanish to audiences in Southeast Asia, South America, and other regions (de Roo, 2014). The source materials originate from digitized administrative and production documents held at the KITLV (Royal Netherlands Institute of Southeast Asian and Caribbean Studies) archive.¹ These documents include broadcast scripts (draaiboeken), daily transmission reports (dagrapporten), music programs (muziekprogramma’s), and news bulletins. The corpus enables research on: (1) the evolution of international broadcasting strategies during the Cold War and decolonization era (Kuitenbrouwer, 2022); (2) multilingual programming and language policy in postcolonial contexts (Mathe & Motsaathebe, 2025); (3) music selection and cultural diplomacy (Fosler-Lussier, 2015); and (4) the development of radio broadcast infrastructure and practices in the 1950s–1960s (Badenoch, 2018; Potter et al., 2022).

Steps

The dataset was produced through a multi-stage enrichment pipeline:

1. OCR Processing: Source PDFs were converted to markdown format using Infinity-Parser-7B, a vision-language model based on Qwen2.5-VL-7B-Instruct (Wang et al., 2025).² This produced 226,359 markdown files organized by year and month. The model is designed to preserve document structure, including tables and headings, though table row alignment errors occur regularly.

2. Document Classification: Pages were classified into six categories using rule-based pattern matching with fuzzy RegEx (allowing up to 2–3 Levenshtein distance depending on pattern length):

   • nieuws (news bulletins): 124,277 pages (54.9%)

   • draaiboek (broadcast script): 65,795 pages (29.1%)

   • rubriek (program section): 17,631 pages (7.8%)

   • dagrapport (daily transmission report): 11,025 pages (4.9%)

   • centraal_draaiboek (central broadcast script): 7,121 pages (3.1%)

   • muziek_programma (music program): 510 pages (0.2%)

3. Date Extraction: Broadcast dates were assigned using a three-step pipeline: (a) fuzzy RegEx from dagrapport page headers; (b) monotonicity verification with manual correction of 173 out-of-order dates; (c) bidirectional filling to propagate verified dates to surrounding pages. This yields day-precise dates for 218,968 pages (96.7%); the remaining 7,391 pages carry month-level dates derived from the PDF-file names.

4. Language Identification: Language detection was performed using py3langid on full page content and on filtered content rows separately. The dominant languages identified are Indonesian (53.5%), English (23.1%), Malay (13.1%), and Dutch (7.5%), with Spanish, Arabic, Javanese, and others accounting for the remainder. Line-level weighted language distributions were also computed, providing fine-grained multilingual analysis.

5. Music Reference Extraction: Music references were extracted from muziek programma and centraal draaiboek pages using an LLM-based structured extraction pipeline (Qwen2.5-VL-32B-Instruct, served via SURF Willma). Fields extracted include title, composers, artists, label, and catalog number. This produced 72,212 music records. Records were matched against the Discogs API³ via catalog number matching and fuzzy artist–title matching, and those for which a match could be established were enriched accordingly.

The complete pipeline is implemented in Python and documented in the accompanying GitHub repository.

Quality control

OCR quality was evaluated against a gold-standard manual transcription of a stratified random sample of 50 pages spanning 1950-1965. The overall Character Error Rate (CER) is 3.66% and the Word Error Rate (WER) is 11.5%; because this aggregate is estimated over all 50 pages, it is the figure we report as representative of corpus OCR quality. Per-year error rates are provided with the evaluation results in the repository, but they should be read as indicative only: each year is represented by just one to six pages. Classification achieves full coverage of all 226,359 pages across the six types; no formal accuracy evaluation of the rule-based classifier has been conducted. For the two-stage ML classifier for the categories that could not be determined using RegEx, an evaluation on 400 pages achieves 96.2% accuracy. For music references, extraction quality was evaluated on a stratified sample of 44 annotated pages (215 true positives): precision is 98.5% and recall is 99.7%, with field-level error rates of 0.9% for titles, 10.7% for artists, and 2.8% for labels. Discogs matching was assessed on 150 pages (747 extracted records): the record-level match rate is 55.3% (413/747 records matched), with a page-level mean of 51%.

Repository name

Zenodo

Object name

RNW-broadcast-archives-enriched-1950-1966

Format names and versions

The dataset consists of two primary tables:

• Page corpus (rnw_pages_enriched): Parquet/CSV.gz, 226,359 records. Fields include: filename, category, broadcast_date, language (all), language (subset), and weighted line-level language distributions (absolute counts and fractions for both all-language and constrained-subset detection).

• Music references (rnw_music_references): Parquet/JSONL, 72,212 records. Fields include: title, composers, artists, label, catalog_number, date, source_ file, and Discogs enrichment fields (discogs_genres, discogs_styles, discogs_country, discogs_year, match_method, match_score).

The repository also includes the OCR markdown files (226,359 .md files), OCR evaluation results, and processing scripts.

Creation dates

OCR processing and enrichment: 2026-01–2026-05. Source material dates: 1950-01–1966-12 (plus November-December 1969).

Dataset creators

Melvin Wevers (University of Amsterdam), Annabel Simons (University of Amsterdam), Jelmer Datema (University of Amsterdam), Thomas Smits (University of Amsterdam). See author roles statement below.

Language

Dataset content is multilingual (Dutch, English, Indonesian, Spanish). Metadata fields and code are in English.

License

CC0 1.0 Universal (Public Domain Dedication)

Publication date

2026-06-22

Limitations

Researchers should be aware of several constraints. First, the overall Character Error Rate is 3.66% with a Word Error Rate of 11.5%. This is a mean calculated on a relatively small sample, so the quality of individual pages can vary, which should be taken into account when working with this data. Second, table structures are regularly misaligned. Third, while music extraction precision and recall are high (98.5% and 99.7%), artist field accuracy is lower (with a 10.7% error rate), and Discogs matching covers only 55% of extracted records, leaving nearly half without enrichment metadata. Fourth, source materials reflect colonial-era perspectives and institutional language, which researchers should account for in their interpretive work (Stoler, 2009).