Modelling Collaborative Semantics with a Geographic Recommender
Christoph Schlieder
SeCoGIS WorkshopNovember 7, 2007, Auckland
Lehrstuhl für Angewandte Informatik in denKultur-, Geschichts- und Geowissenschaften
Otto-Friedrich-Universität Bamberg
Schlieder: Modelling Collaborative Semantics 08-2
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
A geographic place
Bamberg UNESCO world
heritage site
Schlieder: Modelling Collaborative Semantics 08-3
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
A different conceptualization
Bamberg beer capital of
Bavaria
Schlieder: Modelling Collaborative Semantics 08-4
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Yet another conceptualization
da
ily.e
lsch
.eu
Bamberg ?
Schlieder: Modelling Collaborative Semantics 08-5
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Conceptual modelling
Place concepts „Bamberg“, „Southern
Germany“, „Europe“, … Thematically and spatially
different conceptualizations
Issues Formal semantics of place
concepts Data about different
conceptualizations
Contributions Semantic analysis based on
multi-object (!) tagging User similarity data from a
geographic recommender
Schlieder: Modelling Collaborative Semantics 08-6
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Tripost Recommender
Schlieder: Modelling Collaborative Semantics 08-7
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Part 1Geo-information communities
Part 2Collaborative Semantics
Part 3Geographic Recommender
Schlieder: Modelling Collaborative Semantics 08-8
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Geo-information communities
Information community Gould & Hecht (2001)
A Framework for Geospatial and Statistical InformationOGC white paper
An information community is a group of people who share a common geospatial feature data dictionary (including definitions of feature relationships) and a common metadata schema.
Gould & Hecht (2001)
Schlieder: Modelling Collaborative Semantics 08-9
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Example
Cadastral Communities Data and process models 27 national cadastral
authorities in the EU 1 community designing the
Cadastral Reference Model
Ontological engineering One ontology per
information community Cadastral Reference ModelLemmen et al. (2003)
Schlieder: Modelling Collaborative Semantics 08-10
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Conceptual modelling
High quality 30 experts from cadastral
agencies, GIScience and Knowledge Engineering
Description logic-based modelling (OWL-DL)
High cost 4 years for understanding
and modelling property transaction processes
COST G9Modelling real property
transactions
Schlieder: Modelling Collaborative Semantics 08-11
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Ontological engineering
Hess, Schlieder (2006)
Ontology-based Verification
of Core Model Conformity, CEUS
National cadastral data model+ intended correspondences
Core Cadastral Data Model+ conformity constraints
Conformity checker (OWL-DL)
Schlieder: Modelling Collaborative Semantics 08-12
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Information communities
Traditional view Each information community
defines its ontology Number of communities
or ontologies << 100 Complex conceptualization
uses DL role restrictions
Semantic boundaries Ontologies come with crisp
semantic boundaries The Greek cadastral model
is not the Danish model Semantic Web technologies
are appropriate (OWL-DL)
Schlieder: Modelling Collaborative Semantics 08-13
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Part 1Geo-information communities
Part 2Collaborative Semantics
Part 3Geographic Recommender
Schlieder: Modelling Collaborative Semantics 08-14
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Collaborative geodata acquisition
Social Web Communities of users who
collect geospatial data
Collaborative mapping GPS biking trail libraries
Morris et al. (2004), Matyas (2007)
Public domain street mapswww.openstreetmap.org
dense data for London
sparse data for Brussels
ww
w.o
pen
stre
etm
ap
.org
ww
w.o
pen
stre
etm
ap
.org
Schlieder: Modelling Collaborative Semantics 08-15
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Collaborative metadata acquisition
Social tagging Categorization of geospatial
data by a community Keywords („tags“) describe
spatio-temporal coverage and content type
Folksonomies folk taxonomy
= tag vocabulary www.geograph.org.uk
Schlieder: Modelling Collaborative Semantics 08-16
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Tagging as categorization task
tagged by data producerfarm track
Schlieder: Modelling Collaborative Semantics 08-17
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Tag frequency
Example 422.895 images 2.784 categories (tags)
Power law frequency rank - 36% tags used once only
24% tags used 2-5 times Most frequent tag used
17.360 times
rank tag
1 Church
2 Farmland
3 Farm
…
2782 Windmill stump
2783 Luminous object in space (sun)
2784 Penstock
www.geograph.org.uk
Schlieder: Modelling Collaborative Semantics 08-18
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Spatial coveragew
ww
.pan
ora
mio
.com
/pho
to/2
014
27
Neuschwanstein POI in Google maps
Schlieder: Modelling Collaborative Semantics 08-19
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Folksonomies
Low cost Categorization by voluntary
contributors (non-experts)
Low quality No controlled vocabulary
house vs. housemanson vs. manor house
misclassifications
Misclassificationby a non-expert
Schlieder: Modelling Collaborative Semantics 08-20
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
The semantics of tags
User tagging Not just the contributor but
all users provide tags Conflicting tags (!)
Semantic analysis Ternary semantic relation
for user tagged data
Classical view
tagging(object, tag)
Gruber (2005)
tagging(object, tag, user)
Schlieder: Modelling Collaborative Semantics 08-21
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Multi-object tagging
Collection of objects
place nametag
Semantic analysis
tagging({obj1,…,objN}, tag, user)
Schlieder: Modelling Collaborative Semantics 08-22
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Bug or feature?
Quality problem (bug) Serious for folksonomies Even more serious if user
tagging is permitted Unmanageable for multi-
object user tagging?
Consequence Use folksonomies only as
the poor man‘s ontology
Schlieder: Modelling Collaborative Semantics 08-23
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Bug or feature?
Data source (feature) Multi-object user tagging
informs us about different conceptualizations
Consequence Invert the task of finding a
tag for a multi-object Find n objects from a
collection of m >> n to illustrate a (place) concept
Schlieder: Modelling Collaborative Semantics 08-24
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
The semantics of multi-object tags
Hypothesis Selection is based on two
conflicting criteria Typicality: choose typical
instances of the concept Variablity: show the
variability of the concept
violation of the variability criterion
Schlieder: Modelling Collaborative Semantics 08-25
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Empirical data
Bamberg Patchwork Postcards
0
10
1 3 5 7 9 11 13 15 17 19 21 23
rank
freq
uenc
y
Schlieder: Modelling Collaborative Semantics 08-26
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Conceptual modelling
Issues How can we describe the
semantics of place concepts? How do we obtain data about
different conceptualizations?
Selection task Selection seems based on two
conflictin criteria: typicality and variability
Multi-object tagging User tagging of multi-objects
informs about the place concepts of individual users
tagging({obj1,…,objN}, tag, user)
Schlieder: Modelling Collaborative Semantics 08-27
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Part 1Geo-information communities
Part 2Collaborative Semantics
Part 3Geographic Recommender
Schlieder: Modelling Collaborative Semantics 08-28
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Recommender systems
Item-to-item similarity recommendationswww.amazon.com
Schlieder: Modelling Collaborative Semantics 08-29
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Multi-object recommendation
Use case The user selects images
and captions for a patchwork postcard.
The system generates other patchwork postcards with appropriate captions
www.wiai.uni-bamberg.de/tripost
TriPost Webservice
Schlieder: Modelling Collaborative Semantics 08-30
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Tags of a single user
Bamberg
Cardiff
Dublin
Antwerpen
Anna‘s multi-object tags
Schlieder: Modelling Collaborative Semantics 08-31
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Feature similarity
Anna Bill Clio Don Emma Franz
Bamberg 1-3-6 2-4-5 1-4-6 2-3-4
Cardiff 2-3-5 1-4-6 2-4-6 1-4-5 1-2-3
Southern Germany
1-2-6 2-4-6 3-5-6 1-3-6 4-5-6 1-2-6
Southeast of England
2-4-5 2-5-6 1-3-5 1-2-3 2-5-6
sim(A,B) = |A∩B| / |A∪B|2/3 0.66
Schlieder: Modelling Collaborative Semantics 08-32
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
User-to-user similarity
Anna Bill Clio Don Emma Franz
Anna .66 .22 .66 .33 .66
Bill .66 .22 .33 .44 .77
Clio .22 .22 .55 .66 .42
Don .66 .33 .55 .33 .55
Emma .33 .44 .66 .33 .33
Franz .66 .77 .42 .55 .33
Schlieder: Modelling Collaborative Semantics 08-33
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Spatial similarity
Spatial Partonomy Users visiting a similar
selection of places are considered similar
Example: Europe in 7 daysWhich Countries? Which Cities? Which Fotographs?
Printed patchwork postcard
Schlieder: Modelling Collaborative Semantics 08-34
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Computing similarity
Measures Feature similarity
e.g. Tversky measure User-user similarity
e.g. averaged feature similarity
Central idea User-to-user similarity in the
selection task is interpreted as a measure for shared conceptualization
Information community The community of a user u
consists of the k users most similar to u.
Schlieder: Modelling Collaborative Semantics 08-35
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Tag Communities
A
D F
B
C
E
C
AB
A
D F
B
C
E
E
FD
F 3-neighbors(C) C 3-neighbors(F)
Fuzzy semantic boundary2-, 3-, 4-community?
Schlieder: Modelling Collaborative Semantics 08-36
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Conclusions
Issues Formal semantics of place concepts Data about different conceptualizations
Contributions Semantic analysis based on multi-object (!) tagging User similarity data from a geographic recommender
Consequences Tagging communities are different from information
communities
Schlieder: Modelling Collaborative Semantics 08-37
Chair of Computing in the Cultural Sciences
Laboratory for Semantic Information Processing
Conclusions
Folksonomies modeling of semantics before the emergence of information
communities before crisp semantic boundaries have been established
Semantic Web ontologies modeling of semantics after that phase they assume crisp semantic boundaries