deer - rdf data extraction and enrichment framework

Motivation Approach Evaluation Conclusion and Future Work

DEERRDF Data Extraction and Enrichment Framework

Mohamed Ahmed Sherif

September 15, 2015

Mohamed Ahmed Sherif — DEER 1/31


Outline

1 Motivation

2 Approach

3 Evaluation

4 Conclusion and Future Work



Outline

1 Motivation

2 Approach

3 Evaluation




RDF Extraction & Enrichment

Need for enriched datasets

TourismQuestion AnsweringEnhanced Reality...

RDF extraction and enrichment

Triples to be added to the originalKB and/orTriples to be deleted from theoriginal KB



Example

GeoSuPhat

mo:MusicArtist

... artist name of George (Pete) Peterson. Born and raisednear the shores of Lake Michigan and its cold windy wintershe was Influenced by bands like Pink Floyd, Kraftwerk, ToddRundgren, ELP, Hawkwind and his love of computers andsynthesis. After serving in the military and a few yearsgoing to college in Colorado where he studied electronicsand computer programming. George moved to Europe andsettled down in Germany where he had met his wifeChristine. ...

jamendo:336993

foaf:name

a

mo:biography



DEER Enrichment Functions

Idea

Generate enrichment data based on existing data

Input/output are RDF datasets

Current Enrichment Functions

Dereferencing

Linking

NLP

Conformation

Filter



DEER Enrichment Operators

Idea

Create complex workflows by connecting modules

Input/output are RDF datasets

Current Enrichment Operators

Clone

Merge



DEER Specification Paradigm

d1

Dereferencing

d2

cloned3

NLP

d4

Filter

d5 d6Merge

d7

AuthorityConform

d8



Manual DEER Specification

1 @p r e f i x : <ht tp : // geoknow . org / s p e c s o n t o l o g y/> .2 @p r e f i x r d f s : <ht tp : //www.w3 . org /2000/01/ rd f−schema#> .3 @p r e f i x geo : <ht tp : //www.w3 . org /2003/01/ geo/wgs84 pos#> .4 : d1 a : Datase t ;5 : ha sU r i <ht tp : // dbped ia . org / r e s o u r c e / Be r l i n> ;6 : f romEndPoint <ht tp : // dbped ia . org / spa r q l> .7 : d2 a : Datase t .8 : d3 a : Datase t .9 : d4 a : Datase t .

10 : d5 a : Datase t .11 : d6 a : Datase t .12 : d7 a : Datase t .13 : d8 a : Datase t ;14 : o u t p u t F i l e ” De e rBe r l i n . t t l ” ;15 : outputFormat ” Tu r t l e ” .16 : d e r e f a : Module , : De r e f e r enc i ngModu l e ;17 r d f s : l a b e l ” De r e f e r e n c i n g module” ;18 : ha s I npu t : d1 ;19 : hasOutput : d2 ;20 : hasParameter : dere fParam1 .21 : dere fParam1 a : ModuleParameter , : De re f e r enc ingModu l ePa ramete r ;22 : hasKey ” i n pu tP r op e r t y 1 ” ;23 : hasVa lue geo : l a t .24 : c l o n e a : Operator , : C loneOpera to r ;25 r d f s : l a b e l ” Clone op e r a t o r ” ;26 : ha s I npu t : d2 ;27 : hasOutput : d3 , : d4 .28 : n l p a : Module , : NLPModule ;29 r d f s : l a b e l ”NLP module” ;30 : ha s I npu t : d3 ;31 : hasOutput : d5 ;32 : hasParameter : nlpPram1 , : nlpPram2 .33 : nlpPram1 a : ModuleParameter , : NLPModuleParameter ;34 : hasKey ” useFoxL igh t ” ;35 : hasVa lue ”OFF” .36 : nlpPram2 a : ModuleParameter , : NLPModuleParameter ;37 : hasKey ” askEndPoint ” ;38 : hasVa lue f a l s e .39 : f i l t e r a : Module , : F i l t e rModu l e ;40 r d f s : l a b e l ” F i l t e r module” ;41 : ha s I npu t : d4 ;42 : hasOutput : d6 ;43 : hasParameter : F i l t e rP r am1 .44 : f i l t e r P r am1 a : ModuleParameter , : NLPModuleParameter ;45 : hasKey ” t r i p l e s P a t t e r n ” ;46 : hasVa lue ”? s <ht tp : // dbped ia . org / on to l ogy / ab s t r a c t> ?o ” .47 : merge a : Operator , : MergeOperator ;48 r d f s : l a b e l ”Merge op e r a t o r ” ;49 : ha s I npu t : d6 , : d5 ;50 : hasOutput : d7 .51 : aconform a : Module , : Author i t yCon fo rmat ionModu le ;52 r d f s : l a b e l ” Au tho r i t y Conformat ion module” ;53 : ha s I npu t : d7 ;54 : hasOutput : d8 ;55 : hasParameter : aconformPram1 , : aconformPram2 .56 : aconformPram1 a : ModuleParameter , : NLPModuleParameter ;57 : hasKey ” s o u r c e S ub j e c tAu t h o r i t y ” ;58 : hasVa lue ” h t tp :// dbped ia . org ” .59 : aconformPram2 a : ModuleParameter , : NLPModuleParameter ;60 : hasKey ” t a r g e t S u b j e c tA u t h o r i t y ” ;61 : hasVa lue ” h t tp :// dee r . o rg ” .



Manual KB Enrichment

Manual customized enrichment pipelines

⊕ Leads to the expected results

Time consuming

Cannot be ported easily to other datasets



Automatic KB Enrichment

Enrichment pipeline M : K → K that maps KB K to anenriched KB K ′ with K ′ = M(K ).

M is an ordered list of atomic enrichment functionsm ∈M

M =

{φ if K = K ′,

(m1, . . . ,mn),where mi ∈M, 1 ≤ i ≤ n otherwise.

Research questions

1 How to create self-configuring atomic enrichmentfunctions m ∈M?

2 How to automatically generate an enrichment pipeline M?



Outline

1 Motivation

2 Approach

3 Evaluation




Running Example

Dataset DrugBank

Goal Gather information about companies related to drugs fora market study

:Aspirin

:Paracetamol

:Ibuprofen

:Quinine

:Druga

a

aa



Running Example

Dataset DrugBank

Goal Gather information about companies related to drugs fora market study

:Aspirin

:Paracetamol

:Ibuprofen

:Quinine

db:Ibuprofen

db:Aspirin

Ibuprofen was extracted by the research armof Boots company during the 1960s ...

:Druga

a

aa

owl:sameAs

owl:sameAs

rdfs:comment



Atomic Enrichment FunctionsI. Dereferencing atomic enrichment function

Datasets are linked (e.g., using owl:sameAs)

Deferences pre-specified set of predicates

Adds found predicates to source the dataset

:Aspirin

:Paracetamol

:Ibuprofen

:Quinine

:Druga

a

aa







:Aspirin

:Paracetamol

:Ibuprofen

:Quinine

db:Ibuprofen

db:Aspirin


:Druga

a

aa

owl:sameAs

owl:sameAs

rdfs:comment







:Aspirin

:Paracetamol

:Ibuprofen

:Quinine

db:Ibuprofen

db:Aspirin


:Druga

a

aa

owl:sameAs

owl:sameAs

rdfs:commentrdfs:comment







:Aspirin

:Paracetamol

:Ibuprofen

:Quinine

db:Ibuprofen

db:Aspirin



:Druga

a

aa

owl:sameAs

owl:sameAs

rdfs:commentrdfs:comment

rdfs:comment



Self-ConfigurationI. Dereferencing Enrichment Functions

Finds the set of predicates Dp from the enriched CBDsthat are missing from source CBDs

Non-enriched CBD of Ibuprofen

:Ibuprofendb:Ibuprofen :Drugaowl:sameAs

Enriched CBD of Ibuprofen

:Ibuprofendb:Ibuprofen


:Drug

:BootsCompany

a

:relatedCompany

owl:sameAs

rdfs:comment

Dp = {:relatedCompany, rdfs:comment}



Self-ConfigurationI. Dereferencing Enrichment Functions

Dereferences Dp = {:relatedCompany, rdfs:comment}

CBD of Ibuprofen

:Aspirin

:Paracetamol

:Ibuprofen

:Quinine

db:Ibuprofen

db:Aspirin


:Druga

a

aa

owl:sameAs

owl:sameAs

rdfs:comment

Finds only rdfs:comment, adds it to the source dataset

Dereferencing enriched CBD of Ibuprofen



:Drugaowl:sameAs

rdfs:comment



Atomic Enrichment FunctionsII. NLP atomic enrichment function

Datatype objects contain unstructured information

Uses Named Entity Recognition to extract implicit data

Adds extracted entities to the source datasets



:Drugaowl:sameAs

rdfs:comment



Atomic Enrichment FunctionsII. NLP atomic enrichment function

Datatype objects contain unstructured information

Uses Named Entity Recognition to extract implicit data

Adds extracted entities to the source datasets



:Drug

:BootsCompany

a

fox:relatedTo

owl:sameAs

rdfs:comment



Self-ConfigurationII. NLP Enrichment Function

Extracts all possible named entity types

Adds extracted entities to the source dataset

NLP enriched CBD of Ibuprofen



:Drugaowl:sameAs

rdfs:comment



Self-ConfigurationII. NLP Enrichment Function

Extracts all possible named entity types

Adds extracted entities to the source dataset




:Drug

:BootsCompany

a

fox:relatedTo

owl:sameAs

rdfs:comment



Atomic Enrichment FunctionsIII. Predicate conformation atomic enrichment function

Enriched datasets may contain diverse ontologies

Predicate conformation maps a set of a pre-specifiedpredicates to a target ontology



:Drug

:BootsCompany

a

fox:relatedTo

owl:sameAs

rdfs:comment



Atomic Enrichment FunctionsIII. Predicate conformation atomic enrichment function

Enriched datasets may contain diverse ontologies

Predicate conformation maps a set of a pre-specifiedpredicates to a target ontology



:Drug

:BootsCompany

a

fox:relatedTo:relatedCompany

owl:sameAs

rdfs:comment



Self-ConfigurationIII. Predicate conformation Enrichment Function

Finds list of predicates Ps and Pt from the source resp.target datasets with the same subject and objectsChanges each Ps with its respective Pt




:Drug

:BootsCompany

a

fox:relatedTo

owl:sameAs

rdfs:comment

Enriched CBD of Ibuprofen (positive example target)



:Drug

:BootsCompany

a

:relatedCompany

owl:sameAs

rdfs:comment



Self-ConfigurationIII. Predicate conformation Enrichment Function

Finds list of predicates Ps and Pt from the source resp.target datasets with the same subject and objectsChanges each Ps with its respective Pt




:Drug

:BootsCompany

a

fox:relatedTo:relatedCompany

owl:sameAs

rdfs:comment

Enriched CBD of Ibuprofen (positive example target)



:Drug

:BootsCompany

a

:relatedCompany

owl:sameAs

rdfs:comment



KB Enrichment Refinement Operator

Input

Set of atomic enrichment functions MSet of positive examples E

Refinement Operator

ρ(M) =⋃

∀m∈MM ++ m ( ++ is the list append operator)

Output

Enrichment pipeline M



Positive Example

:Ibuprofendb:Ibuprofen :Drugaowl:sameAs

Non-enriched CBD of Ibuprofen



:Drug

:BootsCompany

a

:relatedCompany

owl:sameAs

rdfs:comment

Enriched CBD of Ibuprofen



Learning Algorithm

1 Start by empty enrichment pipeline M = ⊥2 Self-configure all mi ∈M, add as child to ⊥3 Select most promising node

4 Expand most promising node

⊥



Learning Algorithm



⊥

(m1) (m2) (m3)



Learning Algorithm



⊥

(m1) (m2) (m3)

(m1,m2) (m1,m3)



Learning Algorithm



⊥

(m1) (m2) (m3)

(m1,m2) (m1,m3) (m3,m1) (m3,m2)



Learning Algorithm



⊥

(m1) (m2) (m3)

(m1,m2) (m1,m3) (m3,m1) (m3,m2)

(m3,m2,m1)



Most Promising Node Selection

Node complexity c(n)

Linear combination of the node’s children count and level

Node fitness f (n)

Difference between node’s enrichment pipeline F-measureand weighted complexity, f (n) = F (n)− ω.c(n)

ω controls the tradeoff between

Greedy search (ω = 0)Search strategies closer to breadth-first search (ω > 0).

Most promising node

The leaf node with the maximum fitness through thewhole refinement tree



Outline

1 Motivation

2 Approach

3 Evaluation




Experimental Setup

Datasets

1 manual experimental enrichment pipelines for Jamendo

2 manual experimental enrichment pipelines for DrugBank

5 manual experimental enrichment pipelines for DBpedia(AdministrativeRegion)

Learning Algorithm

6 atomic enrichment functions

Termination criterion:

Maximum number of iterations of 10Optimal enrichment pipeline found (F-score = 1)



Configuration of the Search Strategy

Node fitnessf (n) = F (n)− ω.c(n)

ω controls the tradeoff between

Greedy search (ω = 0)Search strategies closer tobreadth first search (ω > 0).

Result: ω = 0.75 leads to thebest results

ω P R F

0 1.0 0.99 0.990.25 1.0 0.99 0.990.50 1.0 0.99 0.990.75 1.0 1.0 1.01.0 1.0 0.99 0.99



Effect of Positive Examples

Manual Examples Size of Time Size of Time Learn IterationsM count M M(KB) learned

M′M′(KB) Time count F -score

M1DBpedia

1 1 0.2 1 1.6 1.3 1 1.02 1 0.2 1 1.8 1.3 1 1.0

M2DBpedia

1 2 23.3 1 0.1 0.2 1 0.992 2 15 2 17 0.3 9 0.99

M3DBpedia

1 3 14.7 3 15.2 6.1 9 0.992 3 15 2 15.1 0.1 9 0.99

M4DBpedia

1 4 0.4 2 0.1 0.7 2 0.992 4 0.6 2 0.3 0.9 2 0.99

M5DBpedia

1 5 22 2 0.1 0.7 2 1.02 5 25.5 2 0.2 0.9 2 1.0

M1DrugBank

1 2 3.5 1 4.1 0.1 10 0.992 2 3.6 1 3.4 0.1 10 0.99

M2DrugBank

1 3 25.2 1 0.1 0.1 10 0.992 3 22.8 1 0.1 0.1 10 0.99

M1Jamendo

1 1 10.9 2 10.6 0.1 2 0.992 1 10.4 2 10.4 0.1 1 0.99



Outline

1 Motivation

2 Approach

3 Evaluation




Conclusion and Future Work

Conclusion

Introduced DEER

Presented self-configuring atomic enrichment functions

Presented an approach for learning enrichment pipelinesbased on a refinement operator

Future Work

Parallelize the algorithm on several CPUs as well as loadbalancing

Support directed acyclic graphs as enrichmentspecifications by allowing to split and merge datasets

Pro-active enrichment strategies and active learning

Implement more enrichment function and operators



Thank You! Questions?

Mohamed SherifAugustusplatz 10D-04109 Leipzig

[email protected]://aksw.org/MohamedSherif

http://aksw.org/Projects/DEER

Automating RDF Dataset Transformation and Enrichment by Mohamed Ahmed Sherif, Axel-Cyrille Ngonga

Ngomo, and Jens Lehmann in 12th Extended Semantic Web Conference, Portoroz, Slovenia, 2015


http://aksw.org/MohamedSherif

http://aksw.org/Projects/DEER

deer - rdf data extraction and enrichment framework

Engineering