1
Workshop on
FOLKSONOMIES
Singapore, February 9, 2009
2
We are very proud to hold a workshop in the „informational city“
of Singapore.
3
Wolfgang G. Stock
Isabella Peters
Researcher, Dept. for Information Science, Heinrich-Heine-University Düsseldorf, Germany
Lectures on Web 2.0 Services and Information Retrieval
Main research area: Folksonomies in Knowledge Representation and Information Retrieval
Professor, Head of Dept. for Information Science, Heinrich-Heine-University Düsseldorf, Germany
Lectures on Information Retrieval, Knowledge Representation, Informetrics and Information Market
Main research areas: Folksonomies, Emotional Information Retrieval and Informetrics of LIS Journals
4
Social Network on our Workshop
• literature, slides, links• Start discussions! Start a forum! Invite more people! Blog!, etc.
http://taggingworkshop.ning.com/
5
Agenda
1. Folksonomies – Indexing without rules2. Folksonomies in information services and library catalogues
Short Break
3. Folksonomies and knowledge representation4. Folksonomies and information retrieval
Short Break
5. Tag gardening for folksonomy enrichment and maintenance6. Find „more like me!“. The social function of a folksonomy
6
Lesson 1
Folksonomies – Indexing without Rules
7
Indexing without Rules
“Anything goes”“Against method”, 1975 (Paul K. Feyerabend, Austro-American philosopher)
Tagging• no rules • no methods – or even against methods• indexing a single document
– synonyms – why not? (New York – NY – Big Apple – … )– homonyms – never heard! (not: Java [Programming Language] – Java
[Island], but Java)– translations – why not? (Singapore – Singapur – …)– typing errors – nobody is perfect (Syngapur)– hierarchical relations (hyponymy) – why not? (Düsseldorf –
North Rhine-Westfalia – Germany)– hierarchical relations (meronymy) – why not? (tree – branch – leaf)
8
Indexing
9
Prosumer
“Prosumer”
• introduced 1980 by Alvin Toffler (American economist) in “The Third Wave”
• prosumerism: characteristic property of the knowledge society
Producer Consumer
Prosumer
10
Tri-partite System
• document (resource)• prosumer (user)• tag
11
Cognitive Indexing Processes
Source: Sinha (2005)
12
Library 2.0
13
15
Network Economy: Positive Feedback Loop
Time
Number of active
(tagging) users
Critical Mass
New userscome along.New users
come along.
Value of the networkincrease.
Value of the networkincrease.
Number of users of the network increase.Number of users of
the network increase.
„success breeds success“
only one standard (in a technological area)
16
Time
red:
Competitor
yellow:
Your Library 2.0 Service
„Combat Area“
„Take off“
Positive Feedback
„Saturation“
„Entry“
Number of active
(tagging) users
Critical Mass
17
How to Become a Standard (Part 1)
• marketing– product
• invite the catalogue’s users to tag• make it easy!• no additional password
– price• “price” = user’s time• Save the time of the user! (Ranganathan): time for
tagging – time for searching
– place• add the folksonomy to a well-known service (e.g., your
library catalogue)
18
How to Become a Standard (Part 2)
• marketing (continued)– promotion
• advertising / public relations• communicate the benefits: “Search with your own tags!” –
“Create your personomy!” – “Share your knowledge!” – “Find more – and better – resources!” – “Find other users similar to your interests!” – …
• awards: “Tagger of the Week” – “Super-Poster” – “Best tagger award” / prizes (e.g., books)
– personnel• especially at the entry phase: your staff has to tag• for promotion• always: software specialists
– processes• process management: knowledge representation tasks
(e.g., tag gardening)• process management: information retrieval tasks (e.g.,
relevance ranking)
19
Lesson 2
Folksonomies in Information Services and Library Catalogues
20
Narrow Folksonomies
• only one tagger (the content creator)
• no multiple tagging
• example: YouTube
Tags
21
Extended Narrow Folksonomies
• more than one tagger
• no multiple tagging
• example: Flickr
Source: Vander Wal (2005)
Tags
Add Tags Option
22
Broad Folksonomies
• more than one tagger
• multiple tagging
• example: Delicious
Source: Vander Wal (2005)
Tags
23
Tagging of OPACs
2 possibilities:
• 1) tagging of resources within the library’s website
• 2) tagging of resources outside the library’s firewall
24
Tagging of OPACS: Within Library’s Website: PennTags
http://tags.library.upenn.edu/
25
Tagging of OPACS: Within Library’s Website: Ann Arbor District Library
http://www.aadl.org/catalog
26
Tagging of OPACS: Within Library’s Website: University Library Hildesheim
http://www.uni-hildesheim.de/mybib/all_tags
27
Tagging of OPACS: Within Library’s Website
• advantages:
– user behaviour can be directly observed and exploited for own applications
– used knowledge organization system (KOS) can profit from user behaviour and user language
– users will be “attracted” to the library
– library will appear “trendy”
28
Tagging of OPACS: Within Library’s Website
• disadvantages:
– development and implementation (costs and manpower) of the tagging service have to be taken over from the library
– if only users may tag: librarians may loose their work motivation or may have a feeling of uselessness
– “lock-in”-effect of users
29
Tagging of Resources Outside the Library‘s Firewall: HEIDI
http://katalog.ub.uni-heidelberg.de/cgi-bin/search.cgi
30
Tagging of Resources Outside the Library‘s Firewall: LibraryThing
http://www.librarything.com/search
31
Tagging of Resources Outside the Library‘s Firewall: BibSonomy
http://www.bibsonomy.org/
32
Tagging of Resources Outside the Library‘s Firewall
• advantages:
– development and implementation (costs and manpower) of the tagging service haven‘t to be taken over from the library
– the library may profit from the “know-how” of the provider of the tagging system
– users may profit from tagging activities of hundreds of other users no lock-in
– library appears “trendy”
33
Tagging of Resources Outside the Library‘s Firewall
• disadvantages
– user behaviour cannot be observed or exploited
– your users support other tagging service
– used KOS cannot profit from user behaviour
34
Social OPAC – Thoughts which have to be made in advance
according to Furner (2007)
during indexing
• the degree of restriction (if any) placed on the number and/or combination of tags that a tagger may assign to a given resource;
• the degree of restriction (if any) placed on the tagger’s choice and form of tags;
• the provision (if any) of context-sensitive suggestions for tags, or for facets that the tagger may wish to consider;
35
Social OPAC – Thoughts which have to be made in advance
according to Furner (2007)
during indexing
• the provision of access (if any) to structured vocabularies of tags;
• the provision of access (if any) to lists or clouds of most frequently- or recently-assigned tags;
• the provision of online access to the full content of resources.
36
Social OPAC – Thoughts which have to be made in advance
according to Furner (2007)
during retrieval
• the degree of restriction (if any) placed on the number and/or combination of tags that a searcher may use in a given query;
• the degree of restriction (if any) placed on the searcher’s choice and form of tags;
• the provision (if any) of context-sensitive suggestions for tags, or for facets that the searcher may wish to consider;
37
Social OPAC – Thoughts which have to be made in advance
according to Furner (2007)
during retrieval
• the provision of access (if any) to structured vocabularies of tags;
• the provision of access (if any) to lists or clouds of most frequently- or recently-searched tags;
• the extent to which tag search is integrated into the existing OPAC search.
38
Social OPAC – Thoughts which have to be made in advance
according to Furner (2007)
for system design and user environment
• to engender a sense of community among library users in separate and remote locations;
• to allow library users to identify other individuals with whom they share interests;
• to engender a sense of empowerment among library users who may not otherwise participate in or contribute to library activities;
39
Social OPAC – Thoughts which have to be made in advance
according to Furner (2007)
for system design and user environment
• to encourage library users to engage with the resources that they tag, and thereby to allow users to come to a deeper understanding of those resources and of the contexts in which they were produced;
• to improve the effectiveness of retrieval of records and discovery of resources;
• to improve the effectiveness of personal rediscovery of resources;
40
Social OPAC – Thoughts which have to be made in advance
according to Furner (2007)
for system design and user environment
• to allow library users to determine which kinds of resources and/or topics are currently popular, newsworthy, or receiving attention;
• to improve the entertainment value of, and thereby the level of user satisfaction with, the search experience;
• to reduce the costs normally incurred in manually cataloging, indexing, or classifying the resources in a collection;
41
Social OPAC – Thoughts which have to be made in advance
according to Arch (2007)
for system design and user environment
• how to handle spam or spagging;
• how to handle linguistic variations, synonyms, homonyms, etc.;
• „ramp up-problem“: who will provide the first content?– subject specialist at your library– forwarding links to users, who are interested in the topic– ...
42
Social OPAC: Benefits
according to Furner (2007)
How do the users benefit?
they participate in the activities of a community of like-minded people;
they identify other individuals with whom they shareinterests;
they contribute to the activities of the library;
43
Social OPAC: Benefits
according to Furner (2007)
How do the users benefit?
they engage with the resources being tagged and/ orwith the records that describe them;
they contribute to improvements in the effectivenessof other users’ searches;
they bookmark resources to which repeated personal access is foreseen;
44
Social OPAC: Benefits
according to Furner (2007)
How do the users benefit?
they determine which kinds of resources and/ortopics are currently receiving attention;
they pass the time in a manner that providesentertainment;
they share their knowledge of the content of resources with others;
45
Social OPAC: Benefits
according to Furner (2007)
How do the users benefit?
they demonstrate the extent of their knowledge of the content of resources;
they instantly recognize the „aboutness“ of the ressource via the tags;
they benefit from the receipt of any concreteincentives supplied by the implementing institution in return for tagging efforts.
46
Short Break
47
Lesson 3
Folksonomies and Knowledge Representation
48
Collective Intelligence
Collective Intelligence• “Wisdom of the Crowds” (Surowiecki)• “Hive Minds” (Kroski) – “Vox populi” (Galton) – “Crowdsourcing”• no discussions, diversity of opinions, decentralisation• users tag a document independently from each other• statistical aggregation of data
Collaborative Intelligence• discussions and consensus• prototype service: Wikipedia (but: 90 + 9 + 1 – rule)
“Madness of the Crowds”• e.g., soccer fans – hooligans• no diversity of opinion – no independence – no decentralisation –
no (statistical) aggregation
49
Power Law Tag Distribution
Source: http:// del.icio.us
Tags zu www.visitlondon.com
0
10
20
30
40
50
60
70
Lond
on
Trav
el
UKEn
gland
Tour
ism
Guide
Cultu
reInf
ormat
ionEn
terta
inmen
Holid
ayLo
ndre
sLo
ndra
f (x)= C / xa
Users
Tags
80/20-Rule
Power Tags
Long Tail
50
Tags zu www.asis.org
0
5
10
15
20
25
30
35
Associa
tions
Librar
yInf
ormati
on
Inform
ation
scien
ce IATe
chno
logy
Profes
siona
lRes
earch
Usabili
tyScie
nce
Librar
ies Web
Inform
ation
archit
ectur
e
ITOrga
nizatio
nsArch
itectu
reOrga
nzatio
nCom
puter
sCon
feren
ce
Inform
ation
_arch
itectur
e
Inform
ation
_scie
nceSoci
etyInverse-logistic Tag Distribution
Source: http:// del.icio.us
Users
Tags
f (x)= e-C‘(x-1)b
Long Trunk
Long Tail
Power Tags
51
Document-specific Tag Distributions
distributions of the top 10 tags in a broad folksonomy (sample: Delicious)
N = 650 bookmarks (minimum of 100 different taging users)
Source: Reher (2008); unpublished
52
Power Tags
• Power Law Distribution • Inverse-logistic Distribution
Power Tags Power Tags
53
Tagging Behaviour
• 1 … 3 tags per document and user• motivations for tagging
– future (own) retrieval– contribution and sharing– attract attention– play and competition– self presentation– opinion expression
• factors which influence tagging– conformity– the role of recommendation
Source: Sen et al. (2006)
Source: Rader & Wash (2008 )
54
Sentiment Tags• negative tags: “awful” – “foolish”, …• positive tags: “amazing” – “useful”, …• applicable for sentiment analysis of documents
Source: Yanbe et al. (2007); service: Hatena Bookmarks
55
Documents which Provoke Emotions
Tagging using scroll-bars
Source:
Schmidt and Stock (2009)
56
Tag Types
57
Discrimination Power of Tags
Tags in Folksonomy
Tags in Concrete Document
frequent rare
frequent“Power tags”
low discrimination
strong discrimination
rare“Long tail”
very low discrimination
low discrimination
58
Benefits of Indexing with Folksonomies
• authentic user language – solution of the “vocabulary problem”• actuality• multiple interpretations – many perspectives – bridging the semantic gap• raise access to information resources• follow “desire lines” of users• cheap indexing method – shared indexing• the more taggers, the more the system becomes better – network effects• capable of indexing mass information on the Web• resources for development of knowledge organization systems• mass quality “control”• searching - browsing – serendipity • neologisms • identify communities and “small worlds”• collaborative recommender system• make people sensitive to information indexing
59
Disadvantages of Indexing with Folksonomies• absence of controlled vocabulary• different basic levels (in the sense of Eleanor Rosch)• different interests – loss of context information• language merging• hidden paradigmatic relations• merging of formal (bibliographical) and aboutness tags• no specific fields• tags make evaluations (“stupid”) • spam-tags• syncategoremata (user-specific tags, “me”)• performative tags (“to do”, “to read”)• other misleading keywords
60
Lesson 4
Folksonomies and Information Retrieval
61
Knowledge Representation and Information Retrieval• two sides of the same coin
• Immanuel Kant (German philosopher): Thoughts without content are empty, intuitions without concepts are blind. ...
Knowledge Representation without Information Retrieval is empty.
Information Retrieval without Knowledge Representation is blind.
FeedbackLoop
62
Information Linguistics• “cleaning tags up”• but: only additionally to raw tags• important basic tasks:
– language identification– word identification (problems:
“informationscience”, “information_science”, …)
– detection and correction of typing errors
– context-specific tags (“me”)– identification of named entities– word form conflation (using, e.g.,
Porter stemmer)– decompounding, phrases– homonymy – synonymy
• “higher” tasks:– semantic relations– translation
63
Relevance Ranking: State of the Art
Interestingness Ranking (Yahoo! / Flickr)• number of tags to the document
• number of users, who tagged the document
• number of users, who retrieved the document
• time (the older the document the less relevant)
• relevance of metadata
Personalized Interestingness Ranking
• user preferences (e.g. favorites)
• user‘s residence
64
Relevance Ranking of Tagged Documents
Source: Peters and Stock (2007)
65
Retrieval Status Value – Factor 1: Tags
• tag frequency or TF*IDF or TF*ITF– index tags (only in broad folksonomies)– search tags
• tag evaluation (feedback of users: „Is this tag useful for finding this document?“)
• more than one search argument: vector space model• time (new tags in platform: higher weight)• Super Poster (term tagged by super poster: higher weight)• Power Tag (higher weight)
tag evaluation
66
Retrieval Status Value – Factor 2: Collaboration
• click rates of a document• number of tagging users• number of comments• linked documents: PageRank
67
Retrieval Status Value – Factor 3: Prosumer
• performative document weight• sentiment weight• rating weight
69
Lesson 5
Tag Gardening for Folksonomy Enrichment and Maintenance
70
The Folksonomy Tag Garden
71
Goal of Tag Gardening: Emergent Semantics
72
• removing “bad tags”: spelling variants (plural vs. singular, conflation of multi-word tags) and spam through “pesticides”
• achieved by type-ahead functionality during indexing, editing functionalities for tags afterwards the application (remove, change, etc.), Natural Language Processing of index tags and search tags, indexing and retrieval tutorials or guidelines for users, authorised users as pesticides
• in order to enhance recall and a consistent indexing vocabulary
simplest form of tag gardening because of neglecting semantics of tags
Weeding
73
• extending the folksonomy with rarely used “baby tags” as high- frequency tags do not sufficiently discriminate resources
• achieved by displaying an inverse tag cloud during indexing or particular “green house” areas where the seedlings may develop and grow, discrete tag suggestions during indexing
• in order to enhance precision and expressiveness of the folksonomy
Seeding
74
• shaping the folksonomy into “flower beds”, distinguishing similar looking “plants”, assigning their “species”, branding each species with labels and giving additional information regarding their application (e.g., cooking, healing, etc.)
• achieved by conflation of multi-language tags, summarization of synonyms, division of homonyms, establishment of semantic relations by comparison with KOS (afterwards indexing)
• in order to enhance precision and expressiveness of the folksonomy by adding semantics, for query expansion during retrieval via semantic relations, for enhanced navigation within the folksonomy, as basis for semantic-oriented displays
Landscape Architecture
75
• combination of folksonomies and KOS during indexing and retrieval
• achieved by semantic-oriented tag suggestions during indexing and retrieval ( tag suggestions not based on tag popularity to avoid self-fulfilling “success breeds success-effect”) or field-based tagging which stimulates semantically richer index tags and search tags
• in order to enhance precision and recall and the expressiveness of the folksonomy by adding semantics, for query expansion during retrieval via semantic relations, for enhanced indexing functionalities, for enhanced navigation within the folksonomy, as basis for semantic-oriented displays
Fertilizing
76
Emergent Semantics
• folksonomies have no explicit structure; there are no visible paradigmatic semantic relations
• document-specific co-occurring tags are linked by syntagmatic relations
• task: to identify paradigmatic relations and to use them in a controlled vocabulary
Synonyms
Is_a
77
(geographical) meronymy
synonymy
hyponymy
Hidden Paradigmatic Relations
Source: http:// www.flickr.com
78
Hidden Paradigmatic Relations. Flickr Landscape Photos
Flickr landscape photos: N=491; analysable tags: 3,618; tags per photo: 7.4Possible document-specific relations (tag-pairs, co-occurrences): 16,098
Document-specific relationsSynonymy 0.56%Abbreviation 0.12%Quasi-Synonymy 0.21%Translation 2.65%
Equivalence (sum) 3.54%Taxonomy 0.23%Simple hyponymy 0.06%
Hyponymy (IS-A relation) (sum) 0.29%Geographical meronymy (administrative) 4.94%Geographical meronymy (not administrative)3.91%Element-collection-relation 0.21%Component-complex-relation 0.84%Segment-time-bond event-relation 0.11%Other meronymy 0.01%
Meronymy (IS-PART-OF relation) (sum) 10.02%Instance 0.23%
Instance 0.23%All relations 14.08%
Source: own research project
79
From Tag Gardening to Collaborative KOS Development
community members als gardeners
• tagging
• evaluation of tags
• field-specific tagging
additional: professional chief-gardener
• KOS development
• new concepts / new words for known concepts
• relations between concepts
80
Maintenance of KOS and Folksonomy
Source: Christiaens (2006)
Folksonomy KOS
Tag Gardening
new terms – new relations
81
From Tag Clouds to Tag Clusters
tag cloud
• alphabetical arrangement
• font size = „importance“ (but mostly no concrete data)
• no relations between tagstag cluster
• tags located in a network
• tuneable granularity (threshold value of similarity)
• relations between tags
• processes:
- calculation of similarity (Jaccard-Sneath, …)
- cluster algorithmsSource: Knautz (2008)
82
Lesson 6
Find „More like me!“ The Social Function of a Folksonomy
83
Users – Tags - Documents
thematically linked
shared users thematically linked
shared documents
84
Shared Documents & Thematically Linked Users
more like this ...
similar documents
detection of documents
more like me ...
similar users
detection of communities
thematically linked
shared documents
85
More like me! Or: More like This User!
• starting point: single user (ego)• processing
– (1) tag-specific similarity• all tags of ego: a(t)• all tags of another user B: b(t)• common tags of ego and another user B: g(t)
– (2) document-specific similarity• all tagged documents of ego: a(d)• all tagged documents of another user B: b(d)• common tagged documents of ego and another user B: g(d)
– calculation of similarity• tag-specific: Jaccard-Sneath: Sim(tag; Ego,B) = g(t) / [a(t) + b(t) – g(t)]• document-specific: Jaccard-Sneath: Sim(doc; Ego,B) = g(d) / [a(d) + b(d) – g(d)]• ranking of Bi by similarity to ego (say, top 10 tag-specific and top 10 document-
specific users)• merging of both lists (exclusion of duplicates)• cluster analysis (k-nearest neighbours, single linkage, complete linkage, group
average linkage)
– result presentation: social network of ego in the centre
86
More like me! Or: More like This User!
single linkage clustering (fictitious example)
Sim(tag) = 0.21
Sim(doc) = 0.25
Sim(tag) = 0.65
Sim(doc) = 0.55
Sim(tag) = 0.33
Sim(doc) = 0.29
Sim(tag) = 0.17
Sim(doc) = 0.23
Sim(tag) = 0.08
Sim(doc) = 0.11
Sim(tag) = 0.15
Sim(doc) = 0.17
Sim(tag) = 0.45
Sim(doc) = 0.36
87
The Social Function of a Folksonomy
objectives:• recommendation of other users with similar interests• hints for forming a virtual community• and – perhaps – for forming a (real) social group
88
Final Discussion
Folksonomies in Library Catalogues – Lessons Learned
89
Lessons Learned
Folksonomies – Indexing without rules
• tagging: anything goes – against methods
• actor: prosumer
• tri-partite system: document – prosumer – tag
• folksonomy behaves like a network good– only one standard– “success breeds success”
• essential: marketing
90
Lessons Learned
Folksonomies in information services and library catalogues
• folksonomy types– narrow folksonomy (only one tagger per document – no
multiple tagging)– extended narrow folksonomy (more than tagger per document – no
multiple tagging)– broad folksonomy (more than one tagger per document – multiple
tags)
• “best” solution for library catalogues– broad folksonomy or– extended narrow folksonomy (only usable if search tags can be
processed)
• platform– own platform (example: PennTags)– third party platform (example: LibraryThing)
91
Lessons Learned
Folksonomies and knowledge representation
• collective intelligence (diversity of options, independence of taggers, decentralisation, statistical aggregation of data)
• document-specific tag distributions– power law– inverse-logistic distribution– Power Tags
• tagging behaviour of the users– 1 … 3 tags per document and user– conformity– recommendation (very problematic)
• sentiment tags (positive – negative)
• documents which provoke emotions
92
Lessons Learned
Folksonomies and information retrieval
• information linguistics (natural language processing)– additional to raw tags: “cleaning” of tags– important tasks: language identification, error detection,
word form conflation
• relevance ranking criteria (calculation of retrieval status values)– tags (TF*IDF, tag evaluation, super posters, time, power
tags)– “collaboration” (click rates, number of tagging users, number
of comments, PageRank)– prosumer (performative tags, sentiment tags, rating)
93
Lessons Learned
Tag gardening for folksonomy enrichment and maintenance
• “weeding”: information linguistics (NLP): core tasks
• “seeding”: baby tags, inverse tag cloud
• “landscape architecture”: combination of folksonomy and KOS (afterwards indexing)
• “fertilizing”: using KOS during indexing and retrieval
• emergent semantics: identification of (hidden) paradigmatic relations (e.g., synonymy and hierarchy)
94
Lessons Learned
Find „more like me!“. The social function of a folksonomy
• a new function: “More like me!”
• recommendation of other users with similar interests
• helpful for community building
95
We would like to thank you very much for attending this Workshop.
Greetings from Düsseldorf, Germany!