Summarizing archival collections using storytelling techniques Yasmin AlNoamany Michele C. Weigle Michael L. Nelson Old Dominion University Web Science & Digital Libraries Research Group www.cs.odu.edu/~mln/ @phonedude_mln Research Funded by IMLS LG-71-15-0077-15 Dodging the Memory Hole Los Angeles, CA, 2016-10-14

1. Summarizing archival collections
using storytelling techniques
Yasmin AlNoamany
Michele C. Weigle
Michael L. Nelson
Old Dominion University
Web Science & Digital Libraries Research Group
www.cs.odu.edu/~mln/
@phonedude_mln
Research Funded by IMLS LG-71-15-0077-15
Dodging the Memory Hole
Los Angeles, CA, 2016-10-14

2. Archive-It, a subscription-based service,
allows creation of collections
2
> 3,000
collections
> 340
institutions
> 10B archived
pages

3. 3
Collection
title
Collection
categorization
based on the
curator
Seed URI
Metadata
about the
collection
Text search
box
The group that
the resource
belongs to
List of
the seed
URIs
Timespan of the
resource
and the number
of times it has
been captured

4. Collection understanding and collection
summarization are not currently supported
Not easy to answer “what’s in that collection?” or
“how is this collection different from others”?
4

5. There is more than one collection about
“Egyptian Revolution”
5
• “2010-2011 Arab Spring” https://archive-it.org/collections/3101
• “North Africa & the Middle East 2011-2013” https://archive-it.org/collections/2349
• “Egypt Revolution and Politics” https://archive-it.org/collections/2358

6. 6
One of at least seven Human Rights collections…

7. 7

8. 8

9. Our early attempts at collection understanding
tried to include everything…
9
“Visualizing digital collections at Archive-It”, JCDL 2012.
http://ws-dl.blogspot.com/2012/08/2012-08-10-ms-thesis-visualizing.html

10. 1000s of seeds X 1000s of archived pages ==
Conventional Vis Methods Not Applicable
10

11. Idea:
Storytelling
11

12. Stories in literature
Story elements: setting, characters, sequence, exposition, conflict,
climax, resolution
Once upon a time
http://www.learner.org/interactives/story/
12

13. Stories in social media
“It's hard to define a story, but I know it when I see it” (Alexander, 2008)
basically, just arranging web pages in time
13

14. “Storytelling” is becoming a popular
technique in social media
14

15. What are the limitations of
storytelling services?
15

16. The Egyptian Revolution on Storify
16

17. Bookmarking, not preserving!
17

18. Despite these limitations, how do we
combine storytelling & archives?
18

19. Use interface people already know how to use
to summarize collections
Archived collectionsStorytelling services
Archived enriched
stories
19

20. We sample k mementos from N pages of the
collection (k << N) to create a summary story
S
1
S
2
S
3
S
4
S
2
S
1
S
3
Collection Y
S
3
S
2
S
1
Collection Z
Collection X
20

21. Yasmin hand-crafted stories to summarize the
Egyptian Revolution collection for her son, Yousof
https://storify.com/yasmina_anwar/the-egyptian-revolution-
on-archive-it-collection
https://storify.com/yasmina_anwar/the-story-of-the-egyptian-
revolution-from-archive- 21

22. How do we generate this automatically?
22

23. Collections have two dimensions:
{Fixed, Sliding} X {Page, Time}
t1 t3t2 t5t4 tk
…
URI
Time
t6
23
…
…

24. Fixed Page, Fixed Time
A desktop Chrome user-agent
http://www.cnn.com/2014/02/24/world/africa/egypt-
politics/index.html?hpt=wo_c2
Android Chrome user-agent
http://www.cnn.com/2014/02/24/world/africa/egypt-
politics/index.html?hpt=wo_c2
Schneider and McCown, “First Steps in Archiving the Mobile Web: Automated Discovery of Mobile Websites”, JCDL 2013.
Kelly et al. “A Method for Identifying Personalized Representations in Web Archives”, D-Lib Magazine 2013 .
24

25. Feb 1 Feb 1 Feb 2
Feb 4 Feb 5 Feb 7
Feb 9 Feb 11 Feb 11
25
Fixed Page, Sliding Time

26. Feb. 11, 2011
Mubarak resigns
26
Sliding Page, Fixed Time

27. Jan 27 Jan 31
Feb 7Feb 4
Feb 11 Feb 11
Feb 2
Jan 25
Feb 10
27
Sliding Page, Sliding Time

28. The Dark and Stormy Archives (DSA) framework
Establish a
baseline
Reduce the candidate
pool of archived pages
Select good
representative
pages
Characteristics of
human-generated
Stories
Characteristics of
Archive-It
collections
Exclude duplicates
Exclude off-topic pages
Exclude non-English Language
Dynamically slice the collection
Cluster the pages
in each slice
Select high-quality
pages from each
cluster
Order pages
by time
Visualize
28
https://pbs.twimg.com/media/BQcpj7ACMAAHRp4.jpg

29. Establish a baseline of
social media stories
"Characteristics of Social Media Stories”, TPDL 2015, IJDL 2016.
29

30. What is the length of a story
(the number of resources per story)?
This story has
31 resources
1
3
2
30

31. What are the types of resources that
compose a story?
Quotes
Video
31
This story has
• 19 quotes
• 8 images
• 4 videos

32. What are the most frequently used domains?
Twitter.com
Twitter.com
Twitter.com
32
This story has
• 90% twitter.com
• 7% instagram.com
• 3% facebook.com

33. Top 25 domains represents 92%
of all domains
33

34. What differentiates a popular story?
(popular = stories with the top 25% of views)
19,795 views 64 views
34

35. The distributions for the features of the stories
• Based on Kruskal-Wallis test, at the p ≤ 0.05 significance level, the popular and the
unpopular stories are different in terms of most of the features
• Popular stories tend to have:
• more web elements (medians of 28 vs. 21)
• longer timespan (5 hours vs. 2 hours) than the unpopular stories
35

36. Do popular stories have a lower decay rate?
The 75th percentile of decay rate per popular story is 10% of the resources,
while it is 15% in the unpopular stories
36

37. We found that 28 mementos is a good
number for the resources in the stories.
37

38. Establish a baseline of current
Archive-It collections
"Characteristics of Social Media Stories. What makes a good story?", International Journal on Digital Libraries 2016.
38

39. The mean and median number of
URIs in a collection
This collection has 435 seed URIs 39

40. The mean and median number of
mementos per URI
This seed URI has 16 mementos 40

41. The most frequent used domains
abcnews.go.com
blogspot.com
This collection has 30% abcnews.com, 10% blogspot.com, 3% facebook.com
41

42. Archive-It top 25 is fundamentally
different than Storify top 25
42

43. Archive-It top 25 is fundamentally
different than Storify top 25
43
Twitter
is #10
not #1

44. What we archive and
what we share on social media
are different subsets of the web
(seeds != shares)
44
see also: Brunelle, et al., “The impact of JavaScript on archivability”, IJDL 2015

45. Detecting off-topic pages
"Detecting Off-Topic Pages in Web Archives”, TPDL 2015, IJDL 2016.
45

46. Archive-It provides their partners with tools
that allow them to build themed collections
46

47. Archive-It tools are about HTTP events /
mechanics, not “content”
47

48. These tools won’t detect that > 60% of
mementos of hamdeensabahy.com are off-topic
May 13, 2012: The page started as
on-topic.
May 24, 2012: Off-topic due to a
database error.
Mar. 21, 2013: Not working because of
financial problems.
May 21, 2013: On-topic again June 5, 2014: The site has been hacked Oct. 10, 2014: The domain has expired.
http://wayback.archive-it.org/2358/*/http://hamdeensabahy.com
48

49. How do we automatically detect
off-topic pages?
49

50. Textual content
cosine similarity, intersection of the most frequent terms,
Jaccard similarity
Method Similarity
cosine 0.7
TF-Intersection 0.6
Jaccard 0.5
50

51. Textual content
cosine similarity, intersection of the most frequent terms,
Jaccard similarity
Method Similarity
cosine 0.7
TF-Intersection 0.6
Jaccard 0.5
Method Similarity
cosine 0.0
TF-Intersection 0.0
Jaccard 0.0
51

52. Semantics of the text
Web based kernel function using the search engine (SE)
52
Sahami and Heilman, A Web-based Kernel Function for Measuring the Similarity of Short Text Snippets, WWW 2006

53. Semantics of the text
Web based kernel function using the search engine (SE)
Method Similarity
SE-Kernel 0.7
53
Sahami and Heilman, A Web-based Kernel Function for Measuring the Similarity of Short Text Snippets, WWW 2006

54. Structural methods
no. of words, content-length
100 109
Method % change
WordCount 0.09
54

55. Structural methods
no. of words, content-length
100 109
100 5
Method % change
WordCount 0.09
Method % change
WordCount -0.95
55

56. We built a gold standard data set to
evaluate the methods
56

57. We manually labeled 15,760 mementos
Egypt Revolution and Politics
URI-Rs: 136
URI-Ms: 6,886
Off-topic URI-Ms: 384
Occupy Movement
URI-Rs: 255
URI-Ms: 6,570
Off-topic URI-Ms: 458
Columbia Univ. Human Rights collection
URI-Rs: 198
URI-Ms: 2,304
Off-topic URI-Ms: 94 57

58. Evaluated 6 methods + combos at 21 thresholds
Averaged the results at each threshold over the three gold standard
collections
Similarity Measure Threshold FP FN FP+FN ACC F1 AUC
(Cosine,WordCount) (0.10,-0.85) 24 10 34 0.987 0.906 0.968
(Cosine,SEKernel) (0.10,0.00) 6 35 40 0.990 0.901 0.934
Cosine 0.15 31 22 53 0.983 0.881 0.961
(WordCount,SEKernel) (-0.80,0.00) 14 27 42 0.985 0.818 0.885
WordCount -0.85 6 44 50 0.982 0.806 0.870
SEKernel 0.05 64 83 147 0.965 0.683 0.865
Bytes -0.65 28 133 161 0.962 0.584 0.746
Jaccard 0.05 74 86 159 0.962 0.538 0.809
TF-Intersection 0.00 49 104 153 0.967 0.537 0.740
58

59. Average precision of 0.89 on 18 different
Archive-It collections
59
(Cosine,WordCount) with (0.10,-0.85) thresholds

60. How do we dynamically divide the
collections into appropriate slices?
(in other words, how do we pick just 28?)
60

61. We expected most collections to look like this…
The Global Food Crisis collection at Archive-It
61

62. This is what we found
Egypt Revolution and Politics
Human RightsApril 16 Archive Virginia Tech Shooting
Jasmine Revolution 2011 Wikileaks Document Release
62

63. Selecting representative pages for
generating stories
(skipping clustering details, but goal is k=28)
63

64. Quality metrics for selecting mementos
• In the DSA, memento quality Mq is calculated as
following:
Mq = (1 − wm*Dm) + wql*Sql + wqc*Sqc
• Dm is the memento damage (Brunelle, JCDL 2014)
• Sql is the snippet quality based on the URI level
• Sqc is the snippet quality based on URI category
• wm, wql, wqc are the weights of memento damage, level,
and category
64

65. We prefer a higher quality memento (Dm)
http://wayback.archive-it.org/2358/20110201231457/
http://news.blogs.cnn.com/category/world/egypt-world-latest-news/
http://wayback.archive-it.org/2358/20110201231622/
http://www.bbc.co.uk/news/world/middle_east/
65
Brunelle et al. Not All Mementos Are Created Equal: Measuring The Impact Of Missing Resources, JCDL 2014

66. We prefer pages with attractive snippets
https://wayback.archive-it.org/2358/20110207193404/http://news.blogs.cnn.com/2011/02/07/egypt-crisis-country-
to-auction-treasury-bills/
https://wayback.archive-
it.org/2358/20110207194425/http://www.cnn.com/2011/WORLD/africa/02/07/egypt.google.executive/index.html?hpt=T1
66

67. We prefer deep links over
high level domains (Sql)
Feb. 11, 2011: the homepage of BBC on Storify
Feb. 11, 2011: the homepage of BBC Middle East section on Storify
Feb. 11, 2011: the article of BBC on Storify
https://wayback.archive-it.org/2358/20110211191429/http://www.bbc.co.uk/
https://wayback.archive-it.org/2358/20110211192204/http://www.bbc.co.uk/news/world-middle-east-12433045
https://wayback.archive-it.org/2358/20110211191942/http://www.bbc.co.uk/news/world/middle_east/
67

68. Social media pages may not produce
good snippets (Sqc)
http://wayback.archive-it.org/1784/20100131023240/http:/twitter.com/Haitifeed/http://wayback.archive-it.org/2358/20141225080305/https:/www.facebook.com/elshaheeed.co.uk
68

69. Visualizing stories in Storify
69

70. Remember Yasmin’s hand-crafted stories?
70

71. Remember Yasmin’s hand-crafted stories?
71

72. We extract the metadata of the pages
and order them chronologically
{ "elements":[
{
"permalink":"http://wayback.archive-it.org/694/20070523182134/http://www.usatoday.com/news/nation/2007-04-16-
virginia-tech_N.htm", "type":"link",
"source":{"href":"http://www.usatoday.com",
"name":"www.usatoday.com
@ 23, May 2007"}
},
{
"permalink":"http://wayback.archive-
it.org/694/20070530182159/http://www.time.com/time/specials/2007/vatech_victims", "type":"link",
"source":{"href":"http://www.time.com",
"name":"www.time.com
@ 30, May 2007" }
},
{
"permalink":"http://wayback.archive-it.org/694/20070530182206/http://www.collegiatetimes.com/",
"type":"link", "source":{"href":"http://www.collegiatetimes.com",
"name":"www.collegiatetimes.com
@ 30, May 2007" }
},
{
"permalink":"http://wayback.archive-it.org/694/20070606234248/http://hokies416.wordpress.com/",
"type":"link", "source":{ "href":"http://hokies416.wordpress.com",
"name":"hokies416.wordpress.com
@ 06, Jun 2007" }
},
…
{ "permalink":"http://wayback.archive-it.org/694/20070620234329/http://www.hokiesports.com/april16/",
"type":"link", "source":{"href":"http://www.hokiesports.com",
"name":"www.hokiesports.com
@ 20, Jun 2007" } },
],
"description":"This is an automatically generated story from Archive-It collection.", "title":"April
16 Archive ”
}
72
Using the Storify API, we
override the default metadata
to generate more attractive
snippets

73. Example of an automatically generated story
73
Notice the good
metadata: images,
titles with dates,
favicons

74. Evaluating the Dark and Stormy
Archive framework
(how good are the automatically generated stories?)
74

75. Evaluation is tricky!
(two perfectly good stories could have non-overlapping k=28
elements!)
• We use human evaluators (via Amazon's
Mechanical Turk) to compare:
• Human-generated stories
• DSA (automatically) generated stories
• Randomly generated stories
• Successful evaluation means:
• Human and DSA stories are indistinguishable
• Human and DSA stories are better than Random
75

76. Our guidelines for expert archivists at Archive-It
for generating stories from the collections
76

77. We received 23 stories for 10
Archive-It collections
SPST is “Sliding Page, Sliding Time”
SPFT is “Sliding Page, Fixed Time”
FPST is “Fixed Page, Sliding Time” 77

78. https://storify.com/mturk_exp/3649b1s-57218803f5db94d11030f90b 78
• Generated by domain experts
• Sliding Page, Sliding Time
• The Boston Marathon
Bombing collection

79. Automatically generated stories from
archived collections
1. Obtain the seed list and the TimeMap of URIs from the front-end
interface of Archive- It
2. Extract the HTML of the mementos from the WARC files (locally
hosted at ODU) and download the collections that we do not have in
the ODU mirror from Archive-It
3. Extract the text of the page using the Boilerpipe library
4. Eliminate the off-topic pages based on the best-performing method
((Cosine, Word-Count) with the suggested thresholds (0.1, −0.85))
5. Exclude duplicates in each TimeMap
6. Eliminate the non-English language pages
7. Slice the collection dynamically and then cluster the mementos of
each slice using DBSCAN algorithm
8. Apply the quality metrics to select the best representative pages
9. Sort the selected mementos chronologically then put them and their
metadata in a JSON object
79

80. https://storify.com/mturk_exp/3649b0s
80
• Automatically generated story
• Sliding Page, Sliding Time
• The Boston Marathon
Bombing collection

81. Random stories
28 mementos were randomly selected from each
collection before excluding off-topic and duplicate
pages
81

82. https://storify.com/mturk_exp/3649b2s-57227227bb79 048c2d0388dc 82
• Randomly generated story
• Sliding Page, Sliding Time
• The Boston Marathon
Bombing collection

83. https://storify.com/mturk_exp/3649bads 83
if someone prefers this story,
we exclude their results
• Poorly generated story
• The same memento, 28 times
• The Boston Marathon
Bombing collection

84. MT experiment setup
• Three HITs for each story (69 HITs to evaluate 23 stories); two
comparisons per HIT:
• HIT1: human vs. automatic, human vs. poor
• HIT2: human vs. random, human vs. poor
• HIT3: random vs. automatic, automatic vs. poor
• 15 distinct turkers with master qualification (i.e., high acceptance rate)
for each HIT
• We rejected the submissions contained poorly-generated stories and
the HITs that were completed in less than 10 seconds (mean time per
HIT = 7 minutes)
• 989 out of 1,035 (69*15) valid HITs
• We awarded the turker $0.50 per HIT
84https://www.mturk.com/mturk/help?helpPage=worker#what_is_master_worker

85. A sample HIT
85

86. DSA == Human
(Human,DSA) > Random
86

87. Automatic versus Human
87
Sliding Page, Sliding Time Sliding Page, Fixed Time Fixed Page, Sliding Time

88. Human versus Random
Sliding Page, Sliding Time Sliding Page, Fixed Time Fixed Page, Sliding Time
88

89. Automatic versus Random
89
Sliding Page, Sliding Time Sliding Page, Fixed Time Fixed Page, Sliding Time

90. Success!
DSA-generated stories are just as good as stories
generated by human experts
90

91. Use interface people already know how to use
to summarize collections
Archived collectionsStorytelling services
Archived enriched
stories
91
All the code, datasets,
papers, slides, etc.:
http://bit.ly/YasminPhD