In all of its many flavors, crowdsourcing works. It works for cultural heritage organizations too. During this presentation we look at various aspects of crowdsourced OCR text correction, commenting, and tagging for digitized historical newspapers at the National Library of Australia’s Trove, the California Digital Newspaper Collection (CDNC), and at the Cambridge Public Library in Cambridge Massachusetts as well as the astounding number of historical birth, death, marriage, census, and other records transcribed by “crowd” volunteers at Family Search. Some aspects include: demographics, experiences, motivation, quality, preferred data, economics and marketing. You will see that crowd sourcing is not only feasible but also practical and desirable. You will wonder why your own cultural heritage organization hasn't begun its own crowdsourcing project!
20140628 crowdsourcing, family history, and long tails for libraries [ala annual las vegas]
1. Crowdsourcing, Family History,
and Long Tails for Libraries
!
http://slidesha.re/1qzB8vv
Frederick Zarndt
frederick@frederickzarndt.com
Secretary, IFLA Newspapers Section
Photo held by John Oxley Library, State Library of Queensland. Original from
Courier-mail, Brisbane, Queensland, Australia.
2. Crowdsourcing is the practice of obtaining
needed services, ideas, or content by
soliciting contributions from a large group of
people, and especially from an online
community, rather than from traditional
employees or suppliers. ... [It] is different
from ordinary outsourcing since it is a task or
problem that is outsourced to an undefined
public rather than a specific, named group.
Wikipedia contributors, "Crowdsourcing," Wikipedia, The Free Encyclopedia,
http://en.wikipedia.org/wiki/Crowdsourcing (accessed March 17, 2013)
5. • On the date of publication of Jeff Howe’s Wired
magazine article, 1-Jun-2007, Wikipedia did not have
an entry (list) of crowdsourcing projects*.
• On 25-Jan-2010 Wikipedia’s list of crowdsourcing
projects had 35 entries*.
• On 17-Mar -2013 Wikipedia’s list of crowdsourcing
projects had 158 entries+.
* From Internet Archives’ Wayback Machine.
+ Wikipedia contributors, "List of crowdsourcing projects," Wikipedia, The Free Encyclopedia,
https://en.wikipedia.org/wiki/List_of_crowdsourcing_projects (accessed March 17, 2013).
6. Amazon Mechanical Turk was launched Nov 2005
Alexa global / country rank of Amazon Mechanical Turk (June 2014): 6,465 / 2,046
crowdsourcing
8. Galaxy Zoo was 1st launched July 2007
Alexa global / country traffic rank of Galaxy Zoo (June 2014): 606,971 / 100,298
citizen science
9. Kickstarter was 1st launched in 2009
Alexa global / country traffic rank of Kickstarter (June 2014): 782 / 326
60,000+ projects successfully funded with more than USD $1,000,000,000
crowd funding
11. Family Search Indexing was 1st launched (beta) 2004
Alexa global / country traffic rank of FamilySearch (June 2014): 4,385 / 1,321
12. Project Gutenberg was 1st launched Dec 1971
Alexa global / country traffic rank of Project Gutenberg (June 2014): 6,615 / 4,066
13.
14. Alexa global / country traffic rank of National Library of Finland
2,535,854 (31-Oct-2012) / 199 (2-Apr-2012)
15. so what? why should
a library care about
crowdsourcing?
Time Life Pictures
Getty Images
16. “user engagement refers to the quality of
the user experience that emphasizes the
positive aspects of the interaction with a web
application, and in particular the phenomena
associated with wanting to use that
web application longer and frequently”
Elad Yom-Tov, Mounia Lalmas, Georges Dupret, Ricardo Baeza-Yates, Pinard Donmez, and
Janette Lehmann. 2012. The effect of links on networked user engagement. In Proceedings of the
21st international conference companion on World Wide Web (WWW '12 Companion). ACM, New
York, NY, USA, 641-642. DOI=10.1145/2187980.2188167 http://doi.acm.org/
10.1145/2187980.2188167
17. “in addition to increasing search accuracy or
lowering the costs of document transcription,
crowdsourcing is the single greatest advancement
in getting people using and interacting with library
collections”
Paraphrased from Trevor Owen’s blog http://www.trevorowens.org/2012/03/
crowdsourcing-cultural-heritage-the-objectives-are-upside-down/ (accessed June 2013).
18. “While [the National Library of Australia’s]
Trove offers a range of user engagement
features, and use of each of these features
continues to grow, it is Trove’s newspaper
text correction features that have attracted
the highest level of user engagement.”
Marie-Louise Ayres. 2013. ‘Singing for their supper’: Trove, Australian newspapers, and the
crowd. Paper presented at IFLA WLIC 2013, Singapore. Accessed June 2014 IFLA Library http://
library.ifla.org/id/eprint/245.
19. Alexa global / country rank of National Library of Australia (June 2014): 10,964 / 249
Trove gets ~78% of all National Library web traffic.
20. National Library of Australia
• Online since 2008
• More than 13,000,000 / 127,437,967 newspaper
pages / articles (May 2014)
• Top text corrector 2,625,205+ lines (May 2014)
• 2,682,119 lines corrected each month (average for
1st 5 months 2014)
• 129,046,297 lines corrected as of May 2014, up from
66,527,535 lines corrected May 2012
• 129,300 / 8,218 registered / active users (May 2014)
24. California Digital
Newspaper Collection
• CDNC began digitizing newspapers in 2005 as
part of NDNP
• Newspapers digitized to article-level as well as
to page-level as required by NDNP
• Hosted on Veridian beginning 2009
• Collection size 61,412 issues, 545,955 pages,
6,364,529 articles (May 2014)
25. OCR text correction
• OCR text correction added Aug 2011
• Corrections are done line by line
• 2246 registered / 1,266 active users (Jun 2014)
• 2,656,497+ lines of text corrected (Jun 2014)
• ~2% of the collection corrected, 98% to go!
• Top corrector 717,855 lines > 2x 2nd corrector
26.
27. Cambridge Public Library
Historic Newspaper Collection
• Cambridge Historic Newspapers online since Jan 2012.
• Cambridge Massachusetts Public Library digitized local
newspapers (http://cambridge.dlconsulting.com/)
• Newspapers digitized to article-level
• Collection size 6,346 issues, 59,070 pages, 669,406
articles (May 2014)
• Collection includes 13,099 obituary cards
30. Deaths. lln»rieff, Esq. of <c .. Qn.
Sunday, the till. greatly Drandrellt, of
Orms4irJi.- ~ ; ;✓ ' • * On ijfr r inn
l j j j i l F i i j ' 1 1 f H a v o d i v y d ,
Carnarvonshire, S ; **" *- ' « ' March
Oxford, F. Tfovmeud, Uerald. » • V .
•On Tncsdav last, Mr. Charles.
IWilinson, this 8 ; had vf thesis#,, a
week ago, which tcrminate<i'iu his
death. . / ' ■ O'i Sunday, dJst nit. at.
AsbtCnvHall, mar Lancaster,
Mr.,Geo. Worn ick, many years
house'steward hit late Once The
Hamilton and Brandon. He locked
himself h»oWn'r«wte<: soon. twelve
o'clock" that dny, and fii»-d a loaded
pistol "through Ins bead, 1 which
instantaneously killed him. Coronet's
Verdict, shot himself in a temporary fit of
Friday week,
raw OCR text
Excerpt from The British Newspaper Archive, Chester Courant, Tuesday 6-Apr-1819, page 3.
newspaper image
33. uncorrected OCR accuracy by
newspaper title
Title
OCR character
accuracy
~OCR word
accuracy
PRP Pacific Rural Press 1871 - 1922 92.6% 68.1%
SFC San Francisco Call 1890 - 1913 92.6% 68.1%
LAH Los Angeles Herald 1873 - 1910 88.7% 54.9%
LH Livermore Herald 1877 - 1899 88.6% 54.6%
DAC Daily Alta California 1841 - 1891 88.2% 53.4%
CFJ California Farmer and Journal
of Useful Sciences 1855 - 1880
86.5% 48.4%
SN Sausalito News 1885 - 1922 70.4% 17.3%
*Word accuracy assumes average word length is 5 characters
34. corrected OCR accuracy by
newspaper title
Title
OCR character
accuracy
Corrected
accuracy
PRP Pacific Rural Press 1871 - 1922 92.6% 99.3%
SFC San Francisco Call 1890 - 1913 92.6% 99.6%
LAH Los Angeles Herald 1873 - 1910 88.7% 99.1%
LH Livermore Herald 1877 - 1899 88.6% 99.9%
DAC Daily Alta California 1841 - 1891 88.2% 99.9%
CFJ California Farmer and Journal
of Useful Sciences 1855 - 1880
86.5% 99.8%
SN Sausalito News 1885 - 1922 70.4% 100.0%
35. Title
OCR character
accuracy
~OCR word
accuracy
Corrected
accuracy
~Corrected
word accuracy
PRP 1871 - 1922 92.6% 68.1% 99.3% 96.5%
SFC 1890 - 1913 92.6% 68.1% 99.6% 98.0%
LAH 1873 - 1910 88.7% 54.9% 99.1% 95.6%
LH 1877 - 1899 88.6% 54.6% 99.9% 99.5%
DAC 1841 - 1891 88.2% 53.4% 99.9% 99.5%
CF 1855 - 1880 86.5% 48.4% 98.3% 91.8%
SN 1885 - 1922 70.4% 17.3% 100.0% 100.0%
*Word accuracy assumes average word length is 5 characters
corrected OCR accuracy by
newspaper title
36. correction accuracy
by user
User
Average OCR
accuracy
Correction
accuracy
A 70.4% 100.0%
B 87.1% 99.5%
C 95.4% 99.5%
D 86.5% 98.3%
E 95.3% 100.0%
F 91.0% 100.0%
G 91.0% 99.8%
H 90.5% 99.0%
I 96.6% 99.8%
J 94.8% 100.0%
K 86.8% 99.3%
38. Graphic from Kaufmann et al. “More than fun and money. Worker
Motivation in Crowdsourcing – A Study on Mechanical Turk.”
Motivation
39. Motivation
Genealogists and family
historians
• National Library of Australia’s 2012 Trove
status report showed that ~50% of Trove users
are family historians
• National Library of New Zealand survey found
that ~50% of PapersPast users are genealogists
PAPERSPAST
40. • 72% visit UDN for genealogical research
• 20% visit for various other types of historical research
• 87% find obituaries useful
• Over 60% find the other genealogical article types (birth and
wedding announcements) useful
• Only 7% do not find genealogical articles useful
• Many are writing family histories and consequently also look
for general background information
• Older content is much more highly valued than more recent
content (see more detailed explanation that follows)
• 44% find smaller, rural papers more useful, while only 15%
find larger, metropolitan papers more useful
Motivation
2012 user survey
John Herbert and Randy Olsen. Small town papers: still delivering the news.
WLIC 2012, Helsinki Finland. http://conference.ifla.org/past-wlic/2012/119-
herbert-en.pdf
41. • CDNC and Cambridge Public Library
published a user survey in Mar 2013
• 604 / 32 responses
• Surveys are (mostly) identical except
for organization name
Motivation
2013 user survey
46. • “I enjoy the correction - it’s a great way to learn more
about past history and things of interest whilst doing a
‘service to the community’ by correcting text for the benefit
of others.”
• “I have recently retired from IT and thought that I could be
of some assistance to the project. It benefits me and other
people. It helps with family research.”
Rose Holley. March 2009. Many Hands Make Light Work. National Library of Australia.
Accessed June 2014 http://www.nla.gov.au/ndp/project_details/documents/
ANDP_ManyHands.pdf.
Motivation
Trove users’ report
47. “The ‘typical’ Trove user is a very well educated,
highly paid, English speaking employed woman
aged fifty or over, with a significant or primary
interest in family or local history, who visits the
Trove website very frequently. Users of Trove
newspapers are older than the average Trove
user; only 13% of newspaper users are under 40
years or age.”
Marie-Louise Ayres. ‘Singing for their supper’: Trove, Australian newspapers, and the
crowd. WLIC 2013,Singapore. http://library.ifla.org/245/1/153-ayres-en.pdf.
Motivation
Engaged users: Who are they?
48. “Many of Trove’s user engagement features are
very popular. More than 100,000 users have
registered to date, and more than 2 million tags
and nearly 60,000 comments had been added…
[Trove] text correction, however, stands head and
shoulders above any other user engagement
features.”
Motivation
Engaged users: What do they do?
Marie-Louise Ayres. ‘Singing for their supper’: Trove, Australian newspapers, and the
crowd. WLIC 2013,Singapore. http://library.ifla.org/245/1/153-ayres-en.pdf.
49. “when someone transcribes a document, they are
actually better fulfilling the mission of a cultural
heritage organization than someone who simply stops
by to flip through the pages”
Paraphrased from Trevor Owen’s blog http://www.trevorowens.org/2012/03/
crowdsourcing-cultural-heritage-the-objectives-are-upside-down/ (accessed June 2013).
Motivation
Engaged users
50. “I am interested in all kinds of history. I have pursued genealogy
as a hobby for many years. I correct text at CDNC because I see
it as a constructive way to contribute to a worthwhile project.
Because I am interested in history, I enjoy it.”
Wesley, California
Personal communications with CDNC text correctors.
Motivation
CDNC users’ report
51. !
“I only correct the text on articles of local interest - nothing at
state, national or international level, no advertisements, etc.
The objective is to be able to help researchers to locate local
people, places, organizations and events using the on-line
search at CDNC. I correct local news & gossip, personal items,
real estate transactions, superior court proceedings, county and
local board of supervisors meetings, obituaries, birth notices,
marriages, yachting news, etc.”
Ann, California
Personal communications with CDNC text correctors.
Motivation
CDNC users’ report
52. “I am correcting text for the Coronado Tent City Program for
1903. It is important to correct any problems with personal
names and other information so that researchers will be able
to search by keyword and be assured of retrieving desired
results. ... type fonts cause a great deal of difficulty in
digitizing the text and can cause problems for searchers. Also,
many of the guests' names at Tent City and Hotel Del
Coronado were taken from the registration books and reported
in the Program. This led to many problems in spelling of last
names and the editors were not careful to be consistent in the
spellings. This Program is an important resource since it
provides an excellent picture of daily life in Tent City and
captures much of the history of Coronado itself.”
Gene, California
Personal communications with CDNC text correctors.
Motivation
CDNC users’ report
53. “I have always been interested in history, especially the
development of the American West, and nothing brings it alive
better than newspapers of the time. I believe them to be an
invaluable source of knowledge for us and future generations.”
David, United Kingdom
Personal communications with CDNC text correctors.
Motivation
CDNC users’ report
54. CDNC is an excellent source of information matching my
personal interest in such topics as sea history, development
of shipbuilding, clippers and other ships etc. ...
Unfortunately, the quality of text ... is rather poor I’m
afraid. This is why I started to do all corrections necessary
for myself ... and to leave the corrected text for use of
others. .... I am not doing this very regularly as this is just
my hobby and pleasure.
Jerzey, Poland
Personal communications with CDNC text correctors.
Motivation
CDNC users’ report
55. As an amateur historical researcher my time for research is very
limited. Making time to travel to archives, libraries, and historical
societies does not happen as often as I would like. The Cambridge
Public Library’s online newspaper collection has been an invaluable
resource and it is fun. I am very grateful for all the help I have received
over the years from so many research organizations. Correcting text
has several benefits. It makes it much more likely that I will find a
story if I decide to search for it in the future. It is a way of saying
‘thank you’ to the Cambridge Library for having such a great resource
available and maybe I can make the next person’s research a little
easier. It is my own little historical preservation project.
Cambridge Historical Newspapers Text Corrector
Personal communications with CDNC text correctors.
Motivation
Cambridge users’ report
56. so old, boring, easily
entertained people correct
text. convince me there are
real benefits.
58. $
Economics
Financial value of outsourced OCR text correction
for newspapers?
The Assumptions
• 25 to 50 characters per line in a newspaper column:
Assume 40 characters per line (CDNC sample average)
• Outsourced text transcription or correction costs USD
$0.35 to $1.20 per 1000 characters: Assume $0.50
per 1000 characters
59. $$ 2,656,497 lines x 40 characters per line x
1/1000 x $0.50 = $53,130
$ 129,046,297 lines x 40 characters per line x
1/1000 x $0.50 = $2,580,926
Economics
60. $Financial value of in-house OCR text
correction?
The Assumptions
• Correction takes 15 seconds per line
• Cost is hourly wage plus benefits of lowest level
employee, $10 for CDNC, $41.88* for Australia
AUD $40.38 = USD $41.88 is the actual labor value assumed by the National Library of Australia
to calculate avoided costs due to crowdsourced OCR text correction in its 2012 Trove Status
Report.
Economics
61. $$ 2,656,497 lines x 15 seconds per line x 1/3600
hrs per second x $10.00 per hr = $110,687
$ 129,046,297 lines x 15 seconds per line x
1/3600 hrs per second x $41.88 per hr =
$22,518,579
Economics
62. Accuracy
“His Accuracy Depends on Ours!"
Office for Emergency Management. Office of
War Information. Domestic Operations
Branch. Bureau of Special Services. [Photo
held at US National Archives and Records
Administration]
63. Accuracy
• Edwin Kiljin (Koninklijke Bibliotheek the Netherlands)
reports raw OCR character accuracies of 68% for early 20th
century newspapers
• Rose Holley (National Library of Australia) reports raw
OCR character accuracy varied from 71% to 98% on a
sample Trove digitized newspapers
Rose Holley. How good can it get? Analysing and improving OCR accuracy in large scale historic
newspaper digitisation programs. D-Lib Magazine. Mar/Apr 2009. Accessed June 2014 http://
www.dlib.org/dlib/march09/holley/03holley.html.
Edwin Kiljin. The current state-of-art in newspaper digitization. D-Lib Magazine. Jan/Feb 2008. Accessed
June 2014 http://www.dlib.org/dlib/january08/klijn/01klijn.html.
Public domain graphic courtesy of Wikimedia Commons.
64. Accuracy
MAPPING TEXTS* assesses digitization quality of digital
newspapers by comparing the number of words recognized
to the total number of words scanned
* Mapping texts is a collaboration between the University of North Texas and Stanford University aimed at experimenting
with new methods for finding and analyzing meaningful patterns embedded in massive collections of digital newspapers.
65. How does low text accuracy affect search recall?
The Facts
• Average uncorrected OCR character accuracy of the
CDNC sample data is ~89%
• Average length of an English word is 5 characters
• Average word accuracy is 89% x 89% x 89% x 89% x 89%
= 55.8% - round up to 60% or 6 out of 10 words correct
Accuracy
67. Accuracy
The Facts
• Average corrected character accuracy of the CDNC
sample data is ~99.4%
• Average word accuracy of CDNC corrected text is 99.4%
x 99.4% x 99.4% x 99.4% x 99.4% = 97.0%
69. A search for “Arndt” at Chronicling America gives
10,267 results*
• If Chronicling America text accuracy is 55.8% (same as
uncorrected CDNC sample), then 8,133 instances of
“Arndt” were not found
• If text accuracy is 97.0%, then 317 instances of “Arndt”
were not found
Accuracy
* Search performed 31 Oct 2012
70. Accuracy
Suppose the word/name is longer than 5
characters?
The Facts
• Assume that average uncorrected / corrected OCR
character accuracy is ~89% / ~99% same as CDNC.
Name Name length Raw text accuracy Corrected text accuracy
Eklund 6 49.7% 94.2%
Kennedy 7 44.2% 93.25
Espinosa 8 39.4% 92.3%
Bonaparte 9 35% 91.4%
Chatterjee 10 31.2% 90.4%
71. Accuracy
Name
Number of
search results
Missing results with
raw text accuracy
Missing results with
corrected text accuracy
Eklund 2,951 2,987 182
Kennedy 360,723 455,392 26,111
Espinosa 1,918 2,950 160
Bonaparte 44,664 82,947 4,203
Chatterjee 19 42 2
Chronicling America searches done 19-Mar-2013
(6,025,474 pages from 1836 to 1922).
72. but you left
out long
tails…
Public domain illustration
from "On The Genesis of
Species" by St. George Mivart
73.
74. the long tail* of crowdsourced
OCR text correction
a probability distribution has a long tail if a larger
share of population rests within its tail than it would
under a normal distribution
!
the most productive users represent a small fraction
of the total user population and ~50% of total
production, or, said a different way, the largest
fraction but individually not quite so productive
users are as important as the most productive users
The phrase “long tail” was popularized by Chris Anderson in the October 2004 Wired
magazine article The Long Tail and by Clay Shirky’s February 2003 essay “Power laws,
web logs, and inequality”.
76. OCR text correction long tails
0
75000
150000
225000
300000
CDNC lines corrected by text corrector
0
750,000
1,500,000
2,250,000
3,000,000
NLA lines corrected by text corrector
top corrector 242,965 top corrector 1,456,906
50%
50%
50%
50%
77. Future considerations
• How to market / advertise
crowdsourcing?
• How to motivate
crowdsourcers?
• Is authentication / identity of
crowdsourcers an issue?
• How to administer
crowdsourced data?
Photo of Aleister Crowley [Public domain] from Wikimedia
Commons
78. Conclusions
Conclusion of the Sonata for piano #32, opus 111 by
Ludwig van Beethoven
• Lots of crowdsourcing in cultural heritage
organizations and elsewhere
• Benefits are multi-faceted: Economic, data
accuracy, user engagement, increased web traffic
80. Resources
Public domain photo “A useful instruction for young sailors from the Royal
Hospital School, Greenwich” from the National Maritime Museum.
81. Correct California newspapers at http://cdnc.ucr.edu
Correct Cambridge MA newspapers http://bit.ly/cambridgepublic
Correct Australian newspapers http://trove.nla.gov.au
Correct Virginia newspapers http://virginiachronicle.com
Try crowdsourcing!
82. Other resources
Mapping Texts at http://mappingtexts.stanford.edu/
Wragge Labs at http://wraggelabs.com/
Wikipedia list of crowdsourcing projects
https://en.wikipedia.org/wiki/
List_of_crowdsourcing_projects
Wikipedia list of digitized newspapers
http://en.wikipedia.org/wiki/
List_of_online_newspaper_archives
83. ?
Photo held by John Oxley Library, State Library of Queensland. Original from
Courier-mail, Brisbane, Queensland, Australia.
Frederick Zarndt
frederick@frederickzarndt.com
Secretary, IFLA Newspapers Section