1. Keyword Search on Structured Data using
Relevance Models*
Veli Bicer
INFORMATIK
FZI Research Center for Information Technology
Karlsruhe, Germany
FZI FORSCHUNGSZENTRUM
Joint work with Thanh Tran from Semantic Search Group, AIFB
Institute, KIT
* based on the papers @ 20th ACM Conference on Information and Knowledge
Management (CIKM’11) and @ 10th International Semantic Web Conference (ISWC’11)
© FZI Forschungszentrum Informatik 1

2. About the presenter
Veli Bicer
 Research Scientist at FZI Research Center for Information
Technology, Karlsruhe, Germany
 Associated Researcher at Karlsruhe Service Research Institute (KSRI)
 KSRI founded by IBM Germany
Research Interests
 Semantic Data Management/Search
 Relational Learning
 Software Engineering (for Services)
Projects
 German Internet Research Programme THESEUS
 KOIOS Semantic Search in Core Technology Cluster
 TEXO Internet-of-Services Use-case
 Previously, EU ICT Artemis, Satine, Saphire and Ride
10.04.2012 © FZI Forschungszentrum Informatik 2

3. Agenda
Introduction
 Keyword search on structured data
 Relevance models
Approach
 Ranking scheme using relevance models
 Top-k Query processing
Experiments
Application
 Search on environmental data
Conclusion
© FZI Forschungszentrum Informatik 3

4. INFORMATIK
FZI FORSCHUNGSZENTRUM
Introduction
10.04.2012 © FZI Forschungszentrum Informatik 4

5. Keyword Search on Structured Data
Rationale
 4 billion web searches daily
 Data-driven websites have relational database backend
 Predefined search forms constrain retrieval
 SQL difficult to learn
 simplify data retrieval by not using SQL
© FZI Forschungszentrum Informatik 5

6. Keyword Search on Structured Data
Example
 Who is the character played by Audrey Hepburn in Roman Holiday?
Query result Person Character
 A tree of tuples that is reduced id name id name pid mid
with respect to the query. p1 Audrey Hepburn c1 Princess p1 m1
Ann
Which would you rather write? p3 Kate Winslet
c3 Iris p3 m2
… ………
Simpkins
SELECT C.name
… ……..
FROM Person, Character, Movie
WHERE Person.id = Character.pId Movie
AND Character.mid = Movie.id id title plot
AND Person.name = ‘Audrey Hepburn' m1 Roman Holiday Princess Ann is a royal princess
AND Movie.title = ‘Roman Holiday' ; of unknow of an …
m2 The Holiday Iris swaps her cottage for the
 or “Hepburn Holiday” holiday along the next two …
m3 The Aviator Hughes and Hepburn go to a
holiday and fly together ..
… …… …..
© FZI Forschungszentrum Informatik 6

7. Keyword Search on Structured Data
Many approaches are proposed recently
 Performance focus
 Less consideration of ranking
Recent study (Coffman and Weaver, CIKM 2010)
 effectiveness of previous works are below expectations
 problem about ranking strategies, not performance
Two major types of ranking schemes:
 IR-inspired TF-IDF ranking
 (Liu et al, 2006) (SPARK, 2007)
 Proximity based approaches
 (Banks, 2002) (Bidirectional, 2005)
Problem:
 Missing a robust and principled approach!!
© FZI Forschungszentrum Informatik 7

8. Relevance Models
Proposed by Lavrenko and Croft (SIGIR 01) Q D
Assumes that Classical Model
 queries and documents are samples from a
hidden representation space and
 generated from the same generative model
Initial representation of relevance is R
unknown
 Estimated from query
Q D
Language Model
R
Q D
© FZI Forschungszentrum Informatik 8
Relevance Model

9. INFORMATIK
FZI FORSCHUNGSZENTRUM
Approach
10.04.2012 © FZI Forschungszentrum Informatik 9

10. Overview of Approach
1 Query
2 PRF
3 Query RM
4 Res. RM
words p words p words p
hepburn 0.5 hepburn 0.21 5 Res. Score
hepburn 0.12
holiday 0.5 holiday 0.15 holiday 0.18
audrey 0.13 audrey 0.11
katharine 0.09 D(RMQ||RMR) katharine 0.05
princess 0.01 princess 0.00
roman 0.01 roman 0.06
…. … …. …
Title Name
Roman Holiday Audrey
Hepburn
Breakfast at Tiff. Audrey
Hepburn
The Aviator Katharine
Hepbun
The Holiday Kate
Winslet
6 Query Generation 7 Structured Queries 8 Top-k Query Proc.
9 Result Ranking
© FZI Forschungszentrum Informatik 10

11. Data Model
Different kinds of data
 e.g. relational, XML and RDF data
Data Graph of nodes and edges (G=(V,E))
Resource nodes, attribute nodes
 Every resource is typed
 Resources have unique ids, (e.g. primary keys)
10.04.2012 © FZI Forschungszentrum Informatik 11

12. Edge-Specific Relevance Models 1 2 3
A set of feedback resources FR are retrieved from an inverted keyword index:
 E.g. Q={Hepburn, Holiday}, FR = {m1, p1, p4,m2, c2,m3}
Edge-specific relevance model for each unique edge e:
Probability of word at resource
Importance of resource w.r.t. query
Inverted Index FR Edge-specific Relevance Models
princess  m1, c1
breakfast  m3 p1
name birthplace
hepburn
hepburn  m3,p1,p4,c2
Audrey Hepburn Ixelles Belgium
melbourne  p2
iris  c3 m3
title The Holiday
holiday
holiday  m1,m2,m3 plot
breakfast  m3 Iris swaps her
cottage for the
ann  m1,c2 holiday along the
next two …..
………. … ……. © FZI Forschungszentrum Informatik 12

13. Edge Specific Resource Models 4 5
Each resource (a tuple) is also represented as a RM
 …as final results (joint tuples) are obtained by combining resources
Edge-specific resource model:
The score of resource: cross-entropy of edge-specific RM and
ResM:
© FZI Forschungszentrum Informatik 13

14. Smoothing
Well-known technique to address data sparseness and improve
accuracy of RMs (and LMs)
 is the core probability for both query and resource RM
Local smoothing
Neighborhood of attribute a is another attribute a’:
 a and a’ shares the same resources
 resources of a and a’ are of the same type
 resources of a and a’ are connected over a FK
Neighborhood of a
© FZI Forschungszentrum Informatik 14

15. Smoothing
words P name (v | p1 )
r a
Person Character audrey 0.5 0.4 0.37 0.36
type type type hepburn 0.5 0.4 0.39 0.38
pid_fk
p1 c1 ixelles 0.1 0.09 0.08
p4
birthplace belgium
name name 0.1 0.09 0.08
name
Audrey Hepburn Ixelles Belgium Princess Ann katharine 0.02 0.01
Katharine
Hepburn birthplace connecticut 0.02 0.01
Connecticut USA usa 0.02 0.01
princess
0.035
ann
0.035
Smoothing of each type is controlled
by weights:
where γ1 ,γ2 ,γ3 are control parameters
set in experiments
10.04.2012 © FZI Forschungszentrum Informatik 15

16. Ranking JRTs 9
Ranking aggregated JRTs:
 Cross entropy between edge-specific RM (Query Model) and geometric
mean of combined edge-specific ResM:
The proposed score is monotonic w.r.t. individual resource scores
 …a desired property for most of top-k algorithms
© FZI Forschungszentrum Informatik 16

17. Query Translation* 6 7
Mapping of keywords to data elements
Hepburn Hepburn Holiday Holiday
title
name name title
 Result in a set of keyword elements p4 p1
m1
m3
Data Graph exploration type
type
 Search for substructures (query graph) pid_fk
Character
Person mid_fk
connecting keyword elements
bornIn Movie
 Bi-directional exploration of query
Is-a Location
graphs operates on summary of data hasDist
hasLoc
graph only Summary Producer Studio
Top-k computation
Graph worksFor
 Search guided by a scoring function to Person Character Movie
output only the top-k queries type type type
pid_fk mid_fk
Query graphs to be processed name
?p ?c ?m
title
 Free vs. Non-free variables Hepburn Holiday
*[Tran et al. ICDE’09]
© FZI Forschungszentrum Informatik 17

18. Top-k Query Processing 8
Top-k query processing (TQP) is highly common in Web-
accessible databases
 return K highest-ranked answers
 avoid unnecessary accesses to database
TQP assumes
 Scoring function and attribute values to be known a-priori (e.g. RankJoin)
 Combine attribute values by aggregation function
 Sorted access (SA), random access (RA) probes
How to adapt TQP to return top-k relevant results?
 Results are joined set of resources
 Scores are query-dependent
 No indexing is possible
Idea:
 Retrieve resources for non-free variables and rank
 Use SA on those initially retrieved resources
 Use RA to find other resources
© FZI Forschungszentrum Informatik 18

19. Top-k Query Processing
Result candidate c=<(x1,…,xk),score>
 complete when all variables are bound to some resources
 xi =* indicates unbounded
Threshold
Binding operator 0.50
 c’=(c,xiri)
Threshold determines upper bound for unseen resources
 Scheduling between SA and RA
 Tight bound is desired
Priority Queue
<(p1,*,*),0.50>
Person Character Movie <(*,*,m2),0.50>
type type type
pid_fk mid_fk
?p ?c ?m title
name
Hepburn Holiday
Person Character 0.11 Movie
id name S(r) id name S(r) id title S(r)
p1 Audrey Hepburn 0.20 c1 Princess Ann m2 The Holiday 0.19 Output K=1
p3 Katharine Hepburn 0.18 c2 Katharine Hepburn m1 Roman Holiday 0.18
p5 Philip Hepburn 0.13 c3 Iris Simpkins m3 Holiday Blues 0.09
p6 Anna Hepburn 0.12 c4 Louise m4 Family Holiday 0.08
© FZI Forschungszentrum Informatik 19

20. Top-k Query Processing
Result candidate c=<(x1,…,xk),score>
 complete when all variables are bound to some resources
 xi =* indicates unbounded
Threshold
Binding operator 0.48
 c’=(c,xiri)
Threshold determines upper bound for unseen resources
 Scheduling between SA and RA
 Tight bound is desired
Priority Queue
<(p1,*,*),0.50>
Person Character Movie <(*,*,m2),0.50>
type type type
pid_fk mid_fk <(p3,*,*),0.48>
?p ?c ?m title
name
Hepburn Holiday
Person Character 0.11 Movie
id name S(r) id name S(r) id title S(r)
p1 Audrey Hepburn 0.20 c1 Princess Ann m2 The Holiday 0.19 Output K=1
p3 Katharine Hepburn 0.18 c2 Katharine Hepburn m1 Roman Holiday 0.18
p5 Philip Hepburn 0.13 c3 Iris Simpkins m3 Holiday Blues 0.09
p6 Anna Hepburn 0.12 c4 Louise m4 Family Holiday 0.08
© FZI Forschungszentrum Informatik 20

21. Top-k Query Processing
Result candidate c=<(x1,…,xk),score>
 complete when all variables are bound to some resources
 xi =* indicates unbounded
Threshold
Binding operator 0.47
 c’=(c,xiri)
Threshold determines upper bound for unseen resources
 Scheduling between SA and RA
 Tight bound is desired
Priority Queue
<(*,*,m2),0.50>
Person Character Movie <(p1,c1,*),0.49>
type type type
pid_fk mid_fk <(p3,*,*),0.48>
?p ?c ?m title
name
Hepburn Holiday
Person Character 0.10 Movie
id name S(r) id name S(r) id title S(r)
p1 Audrey Hepburn 0.20 c1 Princess Ann 0.10 m2 The Holiday 0.19 Output K=1
p3 Katharine Hepburn 0.18 c2 Katharine Hepburn m1 Roman Holiday 0.18
p5 Philip Hepburn 0.13 c3 Iris Simpkins m3 Holiday Blues 0.09
p6 Anna Hepburn 0.12 c4 Louise m4 Family Holiday 0.08
© FZI Forschungszentrum Informatik 21

22. Top-k Query Processing
Result candidate c=<(x1,…,xk),score>
 complete when all variables are bound to some resources
 xi =* indicates unbounded
Threshold
Binding operator 0.46
 c’=(c,xiri)
Threshold determines upper bound for unseen resources
 Scheduling between SA and RA
 Tight bound is desired
Priority Queue
<(p1,c1,*),0.49>
Person Character Movie <(p3,*,*),0.48>
type type type
pid_fk mid_fk <(*,c3,m2),0.44>
?p ?c ?m title
name
Hepburn Holiday
Person Character 0.09 Movie
id name S(r) id name S(r) id title S(r)
p1 Audrey Hepburn 0.20 c1 Princess Ann 0.10 m2 The Holiday 0.19 Output K=1
p3 Katharine Hepburn 0.18 c2 Katharine Hepburn m1 Roman Holiday 0.18
p5 Philip Hepburn 0.13 c3 Iris Simpkins 0.05 m3 Holiday Blues 0.09
p6 Anna Hepburn 0.12 c4 Louise m4 Family Holiday 0.08
© FZI Forschungszentrum Informatik 22

23. Top-k Query Processing
Result candidate c=<(x1,…,xk),score>
 complete when all variables are bound to some resources
 xi =* indicates unbounded
Threshold
Binding operator 0.46
 c’=(c,xiri)
Threshold determines upper bound for unseen resources
 Scheduling between SA and RA
 Tight bound is desired
Priority Queue
<(p3,*,*),0.48>
Person Character Movie <(*,c3,m2),0.44>
type type type
pid_fk mid_fk
?p ?c ?m title
name
Hepburn Holiday
Person Character 0.09 Movie
id name S(r) id name S(r) id title S(r)
p1 Audrey Hepburn 0.20 c1 Princess Ann 0.10 m2 The Holiday 0.19 Output K=1
p3 Katharine Hepburn 0.18 c2 Katharine Hepburn m1 Roman Holiday 0.18 <(p1,c1,m1),0.48>
p5 Philip Hepburn 0.13 c3 Iris Simpkins 0.05 m3 Holiday Blues 0.09
p6 Anna Hepburn 0.12 c4 Louise m4 Family Holiday 0.08
© FZI Forschungszentrum Informatik 23

24. INFORMATIK
FZI FORSCHUNGSZENTRUM
Experiments
© FZI Forschungszentrum Informatik 24

25. Experiments
Datasets: Subsets of Wikipedia, IMDB and Mondial Web
databases
Queries: 50 queries for each dataset including “TREC style”
queries and “single resource” queries
Metrics: Three metrics are used: (1) the number of top-1 relevant
results, (2) Reciprocal rank and (3) Mean Average Precision
(MAP)
Baselines: BANKS , Bidirectional (proximity) , Efficient , SPARK,
CoveredDensity (TF-IDF).
RM-S: Our approach
© FZI Forschungszentrum Informatik 25

26. Experiments
MAP scores for all queries
Reciprocal rank for single
resource queries
© FZI Forschungszentrum Informatik 26

27. Experiments
Precision-recall for TREC-style queries on Wikipedia
© FZI Forschungszentrum Informatik 27

28. INFORMATIK
FZI FORSCHUNGSZENTRUM
Application
© FZI Forschungszentrum Informatik 28

29. Large amount of environmental data
Environmental issues stir public interests
 Increase transparency, awareness, responsibility, protection
Growing amount of data
 Public access through EU directive 2003/4/EC
 PortalU (Germany) http://www.portalu.de/
 EDP (UK) http://www.edp.nerc.ac.uk
 Envirofacts (USA) http://www.epa.gov/enviro/index.html
Linking data in international context
 Local government databases of environmental part of LOD cloud
 Linked environment data for the life sciences
© FZI Forschungszentrum Informatik 29

30. Opportunity: mass dissemination and
consumption of environmental data
The percentage of people who actively find environmental
information is significantly lower than those who have those with
frequent access to it!
Complex results
 CO emission values around Karlsruhe area in Germany
Analytics
 CO emission values around Karlsruhe area in Germany
 Sorted by year
 Bar chart
 Emission values of US and Germany
 Compare average
 Timeline visualization
© FZI Forschungszentrum Informatik 30

31. KOIOS – Overview
A semantic search system
 Exploit semantics in the data for keywords interpretation to hide
complexity of query languages and data representation
 Keyword search for searching structured data
 Lower access barriers while enabling richness of data to be fully
harnessed
Contribution
 Transfer research results to commercial EIS
 Selector mechanism
Process
 Input: keywords
 Facet-based refinement
 Selector (result and view template) initialization
 Output: query results embedded in specific views
© FZI Forschungszentrum Informatik 31

32. KOIOS – Architecture
© FZI Forschungszentrum Informatik 32

33. Facets generation
Derive facets from query results (not from query!) for refinement
 Attributes serve as facet categories
 Attribute values as facet values
E.g. for ?s
 Statistics.description: “CO-Emission , PKW”, “CO-Emission , LKW”…
 Value.year: 2005,2006,…
© FZI Forschungszentrum Informatik 33

34. Selectors
Selector: parameterized, predefined result and view templates
 Data parameters: specify scope of information need, initialized to a
particular values based on facet categories and values
 Query parameter: additional data processing for analysis tasks
(GROUP-BY, SORT, MIN, MAX, AVERAGE etc.)
 Presentation parameter: visualization types (data value, data series,
data table, map-based, specific diagram type, etc.)
© FZI Forschungszentrum Informatik 34

35. Selector initialization
Selectors
 capture templates for information needs and presentation of their
results
Map facets to selectors and initialize them
 Applicable selectors: cover facet categories
 Initialize selectors based on facet values
 Initialized values are captured in the WHERE clause
 Non-initialized parameters are included in the SELECT clause
© FZI Forschungszentrum Informatik 35

36. Deployment
Hippolytos project (Theseus)
 Easy access to spatial data
warehouse (disy Cadenza) built for
domain of environmental
administration
Data about
 Emission and waste
 From the Baden-Württemberg
 Provided by:
Umweltinformationssystem (UIS)
Baden-Württemberg, Landesamt für
Geoinformation und
Landentwicklung (LGL) Baden-
Württemberg and Statistisches
Landesamt Baden-Württemberg
© FZI Forschungszentrum Informatik 36

37. Facets and selectors
© FZI Forschungszentrum Informatik 37

38. Chart-based visualization

39. Map-based visualization

40. Conclusions
Keyword search on structured data is a popular problem for
which various solutions exist.
We focus on the aspect of result ranking, providing a principled
approach that employs relevance models.
Experiments show that RMs are promising for searching
structured data.
Top-k Query processing proposed to get only most relevant
results
Application on environmental data enables intuitive
 Access
 Visualization
 Analysis of environmental information!
© FZI Forschungszentrum Informatik 40

41. INFORMATIK
FZI FORSCHUNGSZENTRUM
Thank you for your attention!
Questions?

42. Opportunity: mass dissemination and
consumption of environmental data
Increase transparency, awareness, responsibility, protection
© FZI Forschungszentrum Informatik 42

43. Challenges: intuitive access and visualization of
structured environmental data and analytics
The percentage of people who actively find environmental
information is significantly lower than those who have those
with frequent access to it!
Complex structured queries
Knowledge of the underlying data /
query language
Complex structured data
Heterogeneity and distribution of
environmental data is overwhelming
Complex structured results
Understanding results and
extracting relevant information /
analytics are difficult tasks
© FZI Forschungszentrum Informatik 43

44. KOIOS
Semantic search system, KOIOS, for intuitive access, analysis,
and visualization of structured environmental information
Overview and architecture
Structured query generation
from keywords
Facet-based browsing and
refinement
Selector initialization for final
result and view construction
Implementation and deployment
Conclusions
© FZI Forschungszentrum Informatik 44

45. Conclusions
Replace predefined forms and hard-coded visualization
Semantic search using lightweight semantics in data and
schema to dynamically
 Translate keywords to queries
 Generate facets for results
 Initialize result and presentation templates
Enables intuitive
 Access
 Visualization
 Analysis of environmental information!
© FZI Forschungszentrum Informatik 45

46. Inverted Index
princess  m1, c1
breakfast  m3
hepburn  m3,p1,p4,c2
melbourne  p2
iris  c3
holiday  m1,m2,m3
breakfast  m3
ann  m1,c2
………. … …….
04.04.2011 © FZI Forschungszentrum Informatik 49

47. Ranking Schemes
Proximity between keyword nodes
 EASE:
 XRank:
 w is the smallest text window in n that contains all search keywords
2012-4-10
SIGMOD09 Tutorial 50

48. Ranking Schemes
Based on graph structure
 BANKS
 Nodes:
 Edges :
 PageRank-like methods
 XRank [Guo et al, SIGMOD03]
 ObjectRank [Balmin et al, VLDB04] : considers both
Global ObjectRank and Keyword-specific
ObjectRank
2012-4-10
SIGMOD09 Tutorial 51

49. Ranking Schemes
1 ln(1 ln(tf )) N 1
Score(n, Q) ln
w Q n (1 s ) s dl / avdl df
TF*IDF based:
 Discover/EASE
 [Liu et al, SIGMOD06]
 SPARK
 but not at the node level
2012-4-10
SIGMOD09 Tutorial 52

50. Relevance Models
Relevance sample probabilities
Model q1 P(w|Q) w
israeli
.077 palestinian
M q2 palestinian .055 israel
.034 jerusalem
M q3 raids .033 protest
M .027 raid
w ??? .011 clash
P(q | w) .010 bank
.010 west
P( w) .010 troop
P( w | q1...qk ) P(q | M ) P( M | w) …
P(q1...qk ) q M
P(q1...qk | w)