Searching Linked Data

Thanh Tran, AIFB Institute, KIT, [email protected] KIT – University of the State of Baden-Wuerttemberg and National Laboratory of the Helmholtz Association1

Searching Linked DataFrom Finding Relevant Sources to Computing AnswersInvited Presentation @ International Workshop on Scalable Semantic Computing, Hangzhou, China, November 2010.

Thanh Tran, Günter Ladwig, Veli Bicer, Lei Zhang, Daniel Herzig, Yongtao Ma, Andreas Wagner, Rudi Studer from AIFB Institute, KIT

Thanh Tran, AIFB Institute, KIT, [email protected] KIT – University of the State of Baden-Wuerttemberg and National Laboratory of the Helmholtz Association

Agenda

Searching Linked Data

Opportunities & challenges

Keyword Query Routing

Problem Definition

Summary Models

Experiments

Linked Data Query Processing

Combining Top-down & Bottom-up

Stream-based Query Processing

Corrective Source Ranking

Conclusions

2


Linked Data

- 203 linked datasets serve 25 billion RDF triples interconnected by 395 million links- As of 09-2010 + other linked data not covered by LOD cloud

3

More Data

More Links


Opportunities

4

“Articles from awarded researchers at Stanford ”

Freebase contains data about people DBPedia contains information about awards DBLP contains bibliographic data

More Data

More Links

More complex information needs More precise results More integrated results


Problems“Articles from awarded researchers at Stanford ”

z) n(x,publicatio Stanford) name(y, y) worksAt(x, Award) Turing prizes(x,.,).( yxz

Formulating queries is a hard task!• Which data sources?• Which schema elements?

Processing queries is expensive!• Process against all data sources? • Explore all links to other sources?

Large number of unknown, unexplored & irrelevant sources! What is in there? What is out there? What is relevant?

USABILITY SCALABILITY

5



Given the needs (expressed as sets of keywords), are there answers in linked data? what combination of data sources produce them? how to incorporate related unexplored linked sources?

6

Identify valid combination of sources

Identify schema elements

Let user choose combination of sources

Focus on this combination of sources and explore related linked sources

Keyword Query Routing to Relevant Linked Data Sources

Focused, Adaptive and Stream-based Linked Data Query Processing (c.f. LARKC)


Agenda




Problem Definition

Summary Models

Experiments


Combining top-down & bottom-up

Stream-based query processing

Corrective source ranking

Conclusions

7


LOD Data Graph

8

per1

uni1

Stanford University

per2

JohnMcCarthy

JohnMccarthy

per3 prize1

Turing Award

JohnMcCarthy

author

name name name name label

employ

sameAs sameAs prizes

DBLPFreebase DBPedia

pub2

author

pub1 pub3

…John.

title

per4 prize2author

JohnSmith

Music Award

name label

prizes

Web data modeled as a set of interlinked data graphs Each data graph represent a source Data graph vs. schema graph vs. source graph


LOD Schema Graph

9

Author

University

Person Person Prize

authoremploy


Written Work

author

Article




LOD Source Graph

10



sames sameAs

author


Keyword Query Answers

11

), dD,Q,F,R(q ji

User information need award“„stanford article

per1

uni1

Stanford University

per2

JohnMcCarthy

JohnMccarthy

per3 prize1

Turing Award

JohnMcCarthy

author


employ



pub2

author

pub1 pub3

…John.

title

per4 prize2author

JohnSmith

Music Award

name label

prizes

Article

type


Problem Definition

Keyword query result (also called Steiner graph) is a subgraph of data graph that for every keyword, contains a matching data element (called keyword elements), and these elements are pairwise connected over a path.

12

d-max Steiner graph is a Steiner graph where paths between keyword elements is d-max or less.

Keyword query routing: compute valid set of data sources called keyword routing plan. A plan is valid if its union set of sources produces non-empty keyword query results.


A Valid Keyword Routing Plan

13

), dD,Q,F,R(q ji

User information need award“„stanford article

per1

uni1

Stanford University

per2

JohnMcCarthy

JohnMccarthy

per3 prize1

Turing Award

JohnMcCarthy

author


employ



pub2

author

pub1 pub3

…John.

title

per4 prize2author

JohnSmith

Music Award

name label

prizes

Article

type


Agenda




Problem Definition

Summary Models

Experiments





Conclusions

14


Keyword Sets

16

per1

uni1

Stanford University

per2

JohnMcCarthy

JohnMccarthy

per3 prize1

Turing Award

JohnMcCarthy

author

name name name label

employsameAs sameAs prizes


pub2

author

pub1 pub3

…John.

title

per4 prize2author

JohnSmith

Music Award

name label

prizes

Stanford

University

John

McCarthy John

McCarthy

McCarthy

John

Turing

Award

Smith Music

One keyword set for every data source Elements stand for distinct keywords mentioned in a source


Element-level Keyword-Element Relationship Graph (E- KERG)

17

per1

uni1

Stanford University

per2

JohnMcCarthy

JohnMccarthy

per3 prize1

Turing Award

JohnMcCarthy

author




pub2

author

pub1 pub3

…John.

title

per4 prize2author

JohnSmith

Music Award

name label

prizes

Stanford

University

John

McCarthy John

McCarthy

McCarthy

John

Turin

Award

Smith Music

A keyword-element captures a keyword k and the data element mentioning k A relationship between two keyword-elements exists iff there is a path between

their associated data elements In d-max KERG, the paths to be considered have length d-max or less

uni1 per2 per1 per3 prize1

per4

John

prize2

Award

John

pub4


Schema-level Keyword-Element Relationship Graph (S-KERG)

18

per1

uni1

Stanford University

per2

JohnMcCarthy

JohnMccarthy

per3 prize1

Turing Award

JohnMcCarthy

author




pub2

author

pub1 pub3

…John.

title

per4 prize2author

JohnSmith

Music Award

name label

prizes

Stanford

University

John

McCarthy John

McCarthy

McCarthy

John

Turin

Award

Smith Music

A keyword-element captures a keyword k and the schema element which contains some instances (date elements) mentioning k

A relationship between two keyword-elements exists if there is a path between some instances of their associated schema elements

Groups ele. (rel.) when they capture same keyword (rel. between same classes)


per4

John

prize2

Award

John

pub4

University Person Author

Article Person Prize


Data-Source-level Keyword-Element Relationship Graph (D-KERG)

19

per1

uni1

Stanford University

per2

JohnMcCarthy

JohnMccarthy

per3 prize1

Turing Award

JohnMcCarthy

author




pub2

author

pub1 pub3

…John.

title

per4 prize2author

JohnSmith

Music Award

name label

prizes

Stanford

University

John

McCarthy John

McCarthy

McCarthy

John

Turin

Award

Smith Music

A keyword-element captures a keyword k and the source which contains some instances (date elements) mentioning k

A relationship between two keyword-elements exists if there is a path between some instances of their associated sources

Groups ele. (rel.) when they capture same keyword (rel. between same sources)


per4

John

prize2

Award

John

pub4

University Person Author

Article Person Prize


Agenda




Problem Definition

Summary Models

Experiments





Conclusions

21


Experiments

Chunk of the BTC dataset containing 10M RDF triples from 154 sources, linked via 500K mappings

22

Manually crafted 30 keyword valid multi-data-source queries, i.e., produce non-empty keyword answers and involve more than 2 sources Town River America Beijing Conference Database 2007


Validity

P@k measure the percentage of plans that are valid out of the top-k plans P@5 for KS only 6%, P@5 up to 100% for E-KERG (dmax =4) More valid plans were computed when a higher value was used for dmax

dmax =3 seems to be a good tradeoff Queries with larger number of keywords resulted in lower precision

23

2 3 4 50.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0 E-KERG D-KERG

S-KERG KS

|K|

P@5

0 1 2 3 40.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0 E-KERG

D-KERG

S-KERG

KS

dmax

P@5


Performance

24

Times increased with higher values for dmax

Sharp for E-KERG and S-KERG Relatively stable for D-KERG

Times increase with number of keywords All other models had poor performance w.r.t complex queries but D-KERG E-KERG needed more than 100s for queries with more than 2 keywords

Time for D-KERG was no more than 10ms on average

0 1 2 3 41

10

100

1000

10000

100000

1000000

S-KERG D-KERG KS E-KERG

dmax

Que

ry P

roce

ssin

g Ti

me

(ms)

2 3 4 51

10

100

1000

10000

100000

1000000

S-KERG D-KERG KS E-KERG

|K|

Que

ry P

roce

ssin

g Ti

me

(ms)


Agenda




Problem Definition

Summary Models

Experiments





Conclusions

27


Mixed Query Processing Strategy

Combination of top-down and bottom-up strategies Top-down: partial local index of sources, not assumed to

be complete Bottom-up: new sources are discovered at run-time

Corrective Source Ranking Deal with heterogeneous source descriptions Adaptive re-ranking

Stream-based Query Processing Deal with unpredictable nature of Linked Data access

ISWC 2010, Shanghai, China


Agenda




Problem Definition

Summary Models

Experiments





Conclusions

29


Query Plan

Source Retrieval


Compile-time Construct query plan Probe local index for

sources Network latency

Do not block! Evaluation driven by

incoming data

Run-time Retrieve sources Push data into query plan Discover new sources Rank sources


Join

Join

worksAt(?x, dbpedia:KIT) knows(?x, ?y)

name(?y, ?n)

Results

Source Retriever 1

Source Retriever 2

...

Push

Source RankerRetrievesource

Sourcediscovered

Source 1 (score: 1.0)Source 2 (score: 0.7) ...

Samples

Local source index

Linked Data


Agenda




Problem Definition

Summary Models

Experiments





Conclusions

31



Prefer more relevant sources Relevancy of a source is based on

Current query Any available intermediate results Overall optimization goal

Define a set of source features and derive concrete source metrics Not all metrics are available for all sources (heterogeneity)

Refine previously computed metrics using newly discovered information (intermediate results, samples)



Evaluation

Three systems: top-down (TD), bottom-up (BU), mixed (MI) 8 queries over various datasets (DBpedia, Geonames, NYT) To make the approaches comparable, sources were restricted to

those discoverable by the BU approach ~6200 sources, containing ~500k triples

Sources hosted on local proxy server with artificial delay of 2 seconds 25% of sources were randomly chosen to construct index for MI



Results


Query 1 Query 6

BU MI TD BU MI TD

25% Results 24810.5 10300.0 11038.0 8222.5 4743.5 5545.0

50% Results 43464.5 40782.0 15787.0 10961.5 7650.5 5634.0

Total 84066.5 86895.5 44323.5 24086.0 20711.0 16469.0

Src. Selection 0.0 853.0 1444.5 0.0 1331.0 1863.5

Ranking 25.5 2404.0 411.5 23.5 292.5 335.0

Overall early result reporting25% results: MI 8.7s, BU 15.1s50% results: MI 12.8s, BU 22.0sImprovement of ~42%

Detailed results for two queries:


Result Arrival Times



Agenda




Problem Definition

Summary Models

Experiments





Conclusions

39


Conclusions

40

Keyword query routing Helps users without knowledge of linked data and schemas to

find combination of sources that contain answers corresponding to their needs

Focus on relevant combinations Summarizing at the level of sources (D-KERG) represents the

most practical trade-off, produces results in less than 10ms out of which every second one was valid

Stream-based query processing helps to deal with unpredictable nature of Linked data

Corrective, mixed strategy that incorporate new sources and knowledge at run-time for optimization (source ranking) helped to report early results 42% faster on average


Thanks for Your Attention!

Institute AIFB, KIT

[email protected]

41

mailto:[email protected]

Searching Linked Data

Education

Transcript of Searching Linked Data