How to read a million books?
- 1. How to read a million books Clemens Neudecker Staatsbibliothek zu Berlin 17 May 2017
- 2. How to read a million books? Clemens Neudecker Staatsbibliothek zu Berlin 17 May 2017
- 3. How to read a million books newspapers? Clemens Neudecker Staatsbibliothek zu Berlin 17 May 2017
- 4. How to read* a million newspapers? (* hint: do try this at home) Clemens Neudecker Staatsbibliothek zu Berlin 17 May 2017
- 5. https://xkcd.com/1838 Agenda • Digital Collections • Data, Tools, Formats • Europeana Newspapers • NLP challenges • Experiments & use cases
- 6. About me • Research Coordinator @ Berlin State Library • M.A. Philosophy, Computer Science, Political Science • Mostly curious about – Optical Character Recognition, Document Analysis – Natural Language Processing – Digital Humanities • More: @cneudecker, cneud.net
- 7. Staatsbibliothek zu Berlin • Established 1661 as the Library of the King of Prussia • Today largest research library in Germany, with approx. 11.5m volumes (23m objects) • Part of the „Stiftung Preußischer Kulturbesitz“, a unique union of museums, archives, libraries and research institutes from Berlin • http://staatsbibliothek-berlin.de/
- 10. Europeana • http://www.europeana.eu/portal/ • Europe‘s Digital Library • > 53m objects incl. art, sound, fashion,... • API: http://labs.europeana.eu/api
- 11. DPLA • https://dp.la/ • Digital Public Library of America • > 16m objects • API: https://dp.la/info/developers/codex/
- 12. Hathi Trust • https://www.hathitrust.org/ • Public copy of Google Books • > 15m volumes • API: https://www.hathitrust.org/data
- 13. DDB • http://ddb.de/ • Germany‘s federal Digital Library • > 9m objects • API: https://api.deutsche-digitale-bibliothek.de/
- 14. Trove • http://trove.nla.gov.au/ • Digital Library of Australia • > 540m objects • API: http://help.nla.gov.au/trove/building-with-trove/api
- 15. Formats & Standards • What data is available? • Typically, a digital object is composed of: – Scanned Images in TIFF, JP2 or JPEG – Descriptive metadata in DublinCore – Structural metadata in METS – Text content in ALTO or TEI – Europeana in EDM – Linked Data in RDF or JSON-LD
- 16. Tools (1/3) • OAI-PMH – https://pypi.python.org/pypi/Sickle – https://pypi.python.org/pypi/pyoai – https://pypi.python.org/pypi/oaiharvest • METS – https://pypi.python.org/pypi/metsrw – https://pypi.python.org/pypi/pymets
- 17. Tools (2/3) • DublinCore – https://pypi.python.org/pypi/pydc – https://pypi.python.org/pypi/dcxml • Europeana – https://pypi.python.org/pypi/europeana-search – https://pypi.python.org/pypi/django-europeana
- 18. Tools (3/3) • IIIF – https://pypi.python.org/pypi/iiif/ – https://pypi.python.org/pypi/Flask-IIIF/ • KB NL – https://pypi.python.org/pypi/kb – https://github.com/KBNLresearch/intro-kb-apis – http://lab.kb.nl/
- 19. Europeana Newspapers • EU-project to make Europe‘s historical newspapers searchable & accessible • http://www.europeana-newspapers.eu/
- 20. Europeana Newspapers Collection • 12 million historic newspaper pages text (> 10.000.000.000 tokens) • 40 languages, 4 alphabets • 400 years (1618 – 2016) • http://www.theeuropeanlibrary.org/tel4/newspapers
- 21. OCR / OLR (U.lag nul «chestttetrung- ■geeinoel II, Setch«it,zen I—Ig Ufr sterntpeechee g» U II. für ftrene-geingelpilche: 13 01191 nnd 13 03 11 io"gl f l««lt-beOeu; OetHn *1, blnftraße IS IZeinsptechee; H I Sanemeinummet gurfilrft 8MB); ««de«: gdn.o(tio||e III (ZemlpreAei 284.3»). Iie.gonlen nur nnier heimnnn » Erden bei der veutlchen Bonl »n« Vtdennld-Getelltchoil gttloto bumduig. Commerz- nndprinoldonlN voINchrSomI bomduig u 189 ß>, .»ontbeegee Kochelchlen- eitchelne» 12 mal wSchenNIch. täglich zweimal — morgen« nnd ndendn —, Sonntage nnr morgen». Toonlnge nur abend» Zn den Kochdorerlen wird die Ndend-Nuegode noch am üben! Dieser Entwurf ist. wie Bürgermeister Roß mitteilte, den Fraktionen zur Stellungnahme vorgelegt worden. Zum Donnerstag war eine zweite Sitzung der Fraktionsfübrer vom Vertreter des Senats angeordnet worden, zu der ober zwei Fraktionen, die Teutschnationalen und die Nationalsozialisten, nicht erschienen waren. Von den Nationalsozialisten ist kurz vor Beginn der Sitzung eine telephonische Erklärung abgegeben worden, etwa des Inhalts, daß die Fraktion sich den sachlichen Verhandlungen entziehen müsse, solange nicht gewisse Vorbedingungen erfüllt sein. http://www.theeuropeanlibrary.org/tel4/newspapers/issue/Hamburger_Nachrichten/1932/12/31
- 24. Performance 82.4% 85.3% 80.9% 75.9% 67.5% 83.4% 84.1% 68.1% 93.1% 57.6% 87.0% 68.3% 76.1% 82.6% 54.1% 32.7% 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% SuccessRate Language Setting Bag of Words OCR Evaluation Per Language 79.1% 62.2% 55.9% 58.8% 94.7% 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% Keyword search Phrase search Access via content structure Print/ebook on demand Content based image retrieval SuccessRate(harmonic,areabased) Evaluation Profile Layout Analysis Performance Per evaluation profile
- 25. Experimental(!) Downloads • http://data.theeuropeanlibrary.org/download/ newspapers-by-country/README.html • http://research.europeana.eu/itemtype/ newspapers • http://test-solr-mongo.eanadev.org/ europeana-research-newspapers-dump/ sample-2017-04-26/Staatsbibliothek_zu_Berlin_ Preu%253Fischer_Kulturbesitz/titles.html
- 26. OCR • EU: IMPACT project (2008-2012) • US: eMOP project (2013-2015) • DE: OCR-D project (2016-2018) • Google: –Tesseract –ocropy (fka OCRopus) –Aksara
- 27. Named Entity Recognition • 3 Categories: – PERSON; LOCATION; ORGANIZATION • 3 Languages: – Dutch; French; German • Powered by Stanford CoreNLP - CRF-NER
- 28. Annotations Language # tokens # PER # LOC # ORG French 207,000 5,672 5,614 2,574 Dutch 182,483 4,492 4,448 1,160 German 96,735 7,914 6,143 2,784 Language # tokens # PER # LOC # ORG French 100% 2,75% 2,71% 1,24% Dutch 100% 2,46% 2,44% 0,64% German 100% 8,18% 6,35% 2,88% Language Word-Error-Rate (Bag of Words) Reading Order Success Rate French 16,6% 19,9% Dutch 17,6% 23,2% German 15,9% / 21,9% 13,6%
- 30. Challenges • https://github.com/EuropeanaNewspapers/ ner-corpora/wiki/Corpus-cleanup
- 31. Lack of metadata Issue There is no associated metadata for the annotated text (newspaper title, date, etc.) Solution Automatically match lines with newspaper pages through keyword search
- 32. OCR errors vs. historical spelling Issue Text contains OCR errors but also valid(!) historical spelling variants Solution Document language profiling to distinguish OCR errors and spelling variants theylteil eyeitht ,
- 33. Sentence splits Issue During data pre-processing, (parts) of sentences have been erroneously cut Solution Reconstruct sentences through keyword search and matching procedure
- 34. Hyphenation Issue Text contains hyphenation to be removed but hyphens do also occur in regular text Solution Use a tokenizer to determine hyphens to be removed
- 35. Missing tags Issue Human operators forgot to tag some entities or tagged them with the wrong category Solution ???
- 36. Punctuation Issue According to CONLL, punctuation should be in a separate line from the token - but abbreviations… Solution ???
- 49. EUROPEANA TRANSCRIBATHON CAMPUS BERLIN 2017 22-23 June 2017 Berlin State Library http://pro.europeana.eu/ event/europeana- transcribathon-campus- 2017
- 50. Thank you for your attention! Clemens Neudecker Staatsbibliothek zu Berlin 17 May 2017