1.
Anne Gorter, Edwin Klijn, Rutger van Koert, Marielle Scherer,
Ismee Tames
DATECH 2019
rom Tribunal Archive to Digital Research Facility (TRIADO)
Exploring ways to make archives accessible and usable

2.
TRIADO: from laboratory to ‘reality check’
• Partners: National Archives, NIOD Institute for War,
Holocaust and Genocide Studies, Huygens ING/KNAW
Humanities Cluster, Dutch War Collections
• 2017-2019
• Sample of 13,8 metres (160,000 pages) from Central
Archive of Special Jurisdiction (CABR)

3.
About the CABR

4.
About the CABR
Source: photo album ‘Centraal Archievendepot Justitie’; National Archives

5.
About the CABR
Source: photo album‘Centraal Archievendepot Justitie’; National Archives

6.
Research questions
1. Which digital methods are best suited (in terms of quality, efficiency,
etc) to make large corpora of unstructured, imperfect data, based on
analogue collections, usable as a research facility? Focus is on
applying proven technology on a sample of digitized documents to
work towards four access points: who, what, where and when.
(GENERIC part)
2. Is it possible to answer specific, mainly quantitative statistical
research questions on the basis of the digital data created under 1?
(SPECIFIC part)

7.
Steps
• Preparation for digitization
• Scanning
• Automatic transcription
• Data enrichment
• Web presentation

8.
Diverse quality of materials

9.
A little about OCR
• Abbyy does a good job without much
tweaking
• Preprocessing helps tesseract v4, but
training even more
• Tesseract has a lot of false positives,
Abbyy practically none, but can miss
stamps etc
• Overall best: Abbyy
l Retraining tesseract
l 150 pages, 80% training, 20% testing
0
2
4
6
8
10
12
14
16
18
20
1
1000
2000
5000
10000
20000
30000
40000
50000
60000
70000
80000100000
CER
WER
Error

10.
Named Entities; Who? Where?
• After OCR: NER using frog => results not too good
l Broken OCR
l Different language in the period. For example: Capitalization of Job
titles such Kapitein Jansen, Veldwachter Jansen
• Much better: matching lists of names (almost) exactly
l Can be tweaked with relative Levenshtein distance
l relativeDistance(“hallo”, “ha110”) = 0.3
l relativeDistance(“hallo”, “ha8lo”) = 1.0
• Less false positives, but also missing unknown names
• Searching with relative levenshtein increases false positives and is
much slower

11.
When?
If it looks like a date….
11 = possible day or month
Aug = possible month
1945 = possible year
11 + Aug+ 1945 = date
35 + Aug + 1945 != date
80% correct, remaining errors mostly ocr-errors and
some overdetection

12.
Finding the right document
l Rvl-cdip dataset

13.
Finding the right document
l Questionnaire
l Email
l Budget

14.
Architecture of a CNN. — Source: https://www.mathworks.com/videos/introduction-to-deep-learning-what-are-convolutional-neural-networks--1489512765771.html

15.
Frontend (anonimized)
l Image removed for privacy reasons

16.
Some conclusions
• OCR useful. Word Error Rate (WER) of 15% for reports.
• Auto-classification and date extraction promising. Error rate of 20%.
• Training the software (machine-learning) and preprocessing before
OCR improves the results.
• Extracting names of locations, organisations and persons using off the
shelf NER has lots of errors. Using existing lists of names works better.

17.
Next steps
• Project plan to digitize and provide access to 4 kilometres of archives
• More linking with external data resources (names, places,
organizations, etc)
• Extend automatic transcription

18.
www.oorlogsbronnen.nl
@Oorlogsbronnen
info@oorlogsbronnen.nl
www.oorlogsbronnen.nl/final-report-triado-enrichment-phase