Knowledge Patterns for the Web: extraction, transformation, and reuse

STLab Università di Bologna
!
Knowledge Patterns for the Web:
extraction, transformation, and reuse

!
! Ph.D. candidate

Andrea Giovanni Nuzzolese

nuzzoles@cs.unibo.it

!
!
!
!
19 May 2014 - Bologna
Supervisor

Paolo Ciancarini

!
Tutors

Aldo Gangemi

Valentina Presutti

• Problem statement

• Knowledge Patterns (KPs)

• Methods and case studies of KP extraction from the Web

• K~ore: a software architecture for experimenting with KPs

• Aemoo: a KP-aware application for entity summarization and
exploratory search on the Web

• Conclusion
Outline
2

3
Problem statement

4
The Linked Data cloud
• The Web is evolving from a global information space of linked
documents to one where both documents and data are linked,
known as Linked Data

5
The knowledge soup and
the boundary problem
• What is the information in the Web that provides the relevant
knowledge about Barack Obama as a Nobel Prize laureate?

• Interoperability problem: the Web is a knowledge soup because of
the heterogeneity of formats, representation schemata and
languages

• Relevance problem: It is hard to draw meaningful boundaries
around data in order to extract relevant contextual knowledge

6
What do we need?
• We need structures that organize entities (e.g., Barack Obama)
and concepts (e.g., Nobel Prize laureate) according to a unifying
view

!
• We need methods for extracting these structures from the Web

7
Knowledge Patterns

• Frames

“…any system of concepts related in such a way that to understand any one of them you have
to understand the whole structure in which it ﬁts; when one of the things in such a structure is
introduced into a text, or into a conversation, all of the others are automatically made
available…” [Fillmore 1968]

“…a remembered framework to be adapted to the reality by changing details as necessary. A
frame is a data-structure for representing a stereotyped situation, like being in a certain kind of
living room, or going to a child’s birthday party…” [Minsky 1975]

!
• Semantic Web

“…a KP is a formal schema for organizing concepts and relations that are relevant in a speciﬁc
context…” [Gangemi and Presutti 2010]
8
KPs across disciplines

9
A KP for OfﬁceHolder

9
Formal represenation

9
Access to data

9
Textual grounding
From wikipedia.org

• To identify methods for the extraction of KPs from the Web

!
!
!
!
!
• To design a software architecture for KP extraction

• To evaluate the effectiveness of KPs in a knowledge interaction
task, e.g., entity summarization and exploratory search
10
My thesis objectives

11
Knowledge Pattern transformation

• To increase syntactic and semantic interoperability, hence to
decrease the soup problem

• By homogenizing existing KP-like artefacts expressed in heterogeneous
formats, representing them as OWL 2 KPs
12
Motivations


12
Motivations
FrameNet• Examples are

ontologydesignpatterns.org

12
Motivations
• Examples are

The Component Library

12
Motivations
• Examples are

13
The KP transformation method: Semion

14
KPs from FrameNet
Syntactic reengineering

14
KPs from FrameNet
ABox refactoring

14
KPs from FrameNet
TBox refactoring

• A lexical dataset in Linked Data

• Provides frames as RDF

• Accessible via SPARQL endpoint

• A set of 1024 KPs

• Conceptually equivalent to FrameNet frames, but with
explicit formal semantics

• Published on ontologydesignpatterns.org

• Evaluation

• Based on the demonstration of the isomorphism of each
transformation step
15
Results

16
Knowledge Pattern extraction
from data

17
KP extraction: intuition

!
• Motivation

• To address the knowledge boundary problem

!
• Hypothesis

• The linking structure of Linked Data resources conveys a rich
knowledge that can be used for KP extraction

• Patterns observed over Linked Data links can be used for drawing
meaningful boundaries around data
18
Motivation and hypothesis

19
Method: key concepts
1. Collect RDF links

2. Index links

3. Collect statistics on indexed links

4. Induce boundaries around data

5. Formalize the KP

dbpedia:War_in_Afghanistan
20
Indexing RDF links: the Type Paths
rdf:property
A Type Path Pi,k,j is a
property path, whose
occurrences have the
same rdf:type for their
subject nodes and the
object nodes
dbpedia:Washington
dbpedia:Barack_Obama

20
rdf:type
rdf:property
object nodes
dbpedia:Washington
owl:Thing
dbpo:Event dbpo:MilitaryConﬂict
owl:Thing
dbpo:Person
dbpo:OfﬁceHolder
dbpo:Country dbpo:Place
owl:Thing

20
rdf:type
rdf:property
rdfs:subClassOf
object nodes
dbpedia:Washington
owl:Thing
dbpo:Event dbpo:MilitaryConﬂict
owl:Thing
dbpo:Person
dbpo:OfﬁceHolder
dbpo:Country dbpo:Place
owl:Thing

20
rdf:type
rdf:property
rdfs:subClassOf
object nodes
dbpedia:Washington
dbpo:MilitaryConﬂict
dbpo:OfﬁceHolder
dbpo:Country

20
rdf:property
Type Path
Type Path
object nodes
dbpo:MilitaryConflict
dbpo:OfficeHolder dbpo:Country
dbpo:OfficeHolder
rdf:property

• A KP is a set of type paths, such that

Pi,k,j ∈ KP ⟺ pathPopularity(Pi,k,j) ≥ t

• t is a threshold, under which a type path is not included in an
KP

!
• The pathPopularity is the ratio of how many distinct resources of
a certain type participate as subject in a path to the total number
of resources of that type. E.g.:

• POfficeHolder,wikiPageWikiLink,MilitaryConflict counts of 2500 occurrences in DBpedia

• 20555 individuals belongs to OfficeHolder in DBpedia

• pathPopularity(POfficeHolder,wikiPageWikiLink,MilitaryConflict) = 0.12

!
21
Boundaries of KPs

• Wikipedia contains a lot of knowledge

• It is a collaboratively edited, multilingual, free Internet encyclopaedia

• It is a peculiar source for KP extraction

• It has an RDF dump in Linked Data, i.e., DBpedia, grounded in a large
corpus

• The following design constraints that make KP investigation
easier

• Each wiki page describes a single topic, which corresponds to a single
resource in DBpedia;

• Wikilinks relate wiki pages. Hence each wikilink links two DBpedia
resources, which are typed with DBPO classes
22
Case study: extracting KPs from
Wikipedia links

23
Boundary induction
1. For each path, calculate the pathPopularity

2. Apply multiple correlation between the paths of all
subject types by rank, and check for homogeneity of
ranks across subject types (Pearson ρ = 0.906)

3. Create a prototypical distribution of the pathPopularity
for all the subject types

4. Decide the threshold t by applying clustering on the
prototypical distribution of the pathPopularity

23
Boundary induction



k-means (4 clusters):

• 3 small clusters with ranks above 27,67%

• 1 big cluster with ranks below 18,18%

23
Boundary induction



k-means (6 clusters):

• 1 big cluster with ranks below 11,89%

• the 9th rank of pathPopularity is at 11,89% and 9 is
the average number of frame elements in FrameNet

• Results

• Discovered 184 KPs formalized as OWL 2 ontologies

• KPs from Wikipedia links are called Encyclopaedic KPs (EKPs) as
they capture encyclopaedic knowledge
24
Results and evaluation

• Results

• Discovered 184 KPs formalized as OWL 2 ontologies

• KPs from Wikipedia links are called Encyclopaedic KPs (EKPs) as
they capture encyclopaedic knowledge
24
Results and evaluation
• Evaluation

• We conducted a user study asking 17 users to judge how relevant were
a number of (object) types (i.e., paths) for describing things of a certain
(subject) type, for a sample of 12 DBPO classes

• We compared average multiple correlation (Spearman’s ⍴ ~0.75 on a
range [-1, 1]) between users' assigned scores (Kendall’s W among
users ~0.68 on a range [0, 1]), and pathPopularity based scores.

25
Source enrichment

• Motivations

• Most of the Web links are untyped and unlabelled hyperlinks

• In many cases RDF statements do not provide typed entities
(e.g., 33% of DBpedia entities are untyped)

• The Web knowledge is mainly expressed by means of
natural language

• Hypothesis

• Natural language text can be used for generating RDF data
suitable for KP extraction

• E.g., a text surrounding anchors in Web pages or annotations in RDF
graphs
26
Motivations and hypothesis

• Using natural language deﬁnitions available in DBpedia abstracts
in order to type DBpedia entities
27
Automatic typing of DBpedia entities

27
Natural language deep parsing
(FRED - http://wit.istc.cnr.it/stlab-tools/fred)

27
Graph-based pattern matching

27
Word-sense disambiguation

27
Ontology Alignment

28
Results
• ORA: the Natural Ontology of Wikipedia

• Typed 3,023,890 entities with associated taxonomies of types

• Evaluation against a golden standard of the accuracy of types
assigned to a sample set of 318 Wikipedia entities

• User study for evaluating the soundness of the induced
taxonomy of types for each DBpedia entity

• Kendall’s W: 0.79

29
Source enrichment: general approach

29
Source enrichment: general approach
• Based on this approach other applications have been developed
so far

• CiTalO: automatic identiﬁcation of the nature of citations with
respect to the CiTO ontology [Di Iorio et al.]

• Sentilo: a semantic sentiment analysis tool [Reforgiato et al.]

• Legalo: automatic uncovering of the semantics of hyperlinks

30
K~ore

31
Architecture

31
Architecture
Transformation (knowledge soup problem)
Extraction
(knowledge boundary problem)
Reuse

32
K~tools

33
Aemoo

• Aemoo is a KP-aware application

• A KP-aware application is a system which

• Beneﬁts from KPs for addressing knowledge interaction tasks

• Uses KPs as the basic unit of mean for representing, exchanging, as
well as reasoning with knolwedge

• Aemoo exploits EKPs for

• Entity summarisation and Exploratory search

• Distinguishing between core and peculiar knowledge

• The data sources are Wikipedia, DBpedia,Twitter, and
GoogleNews
34
Aemoo in a nutshell

35
Aemoo UI
http://aemoo.org

• We asked to 83 users to use Aemoo, RelFinder and Google for
tasks of

• Summarization

• Lookup

• Exploratory search

36
Evaluation

37
Conclusion
• We have provided methodologies for

• KP transformation

• KP extraction

• Source enrichment

• We have designed a software architecture which implements such
methodologies

• We have developed a KP-aware application:Aemoo

• We are contributing to the realization of the Semantic Web as an
empirical science

• We have generated KPs and published them into a repository for
reuse

• 16 peer reviewed articles in international conferences and
workshops

• V. Presutti, D. Reforgiato A. Gangemi,A. Nuzzolese, S. Consoli. Sentilo: Frame-based Sentiment Analysis. Cognitive
Computation, to appear.

• Paolo Ciancarini,Angelo Di Iorio,Andrea Giovanni Nuzzolese, Silvio Peroni, FabioVitali: Evaluating Citation
Functions in CiTO: Cognitive Issues. In Proceedings of the 11th Extended Semantic Web conference (ESWC 2014).
Springer, pp 580-594, Heraklion, Greece, 2014

• A. G. Nuzzolese,V. Presutti,A. Gangemi,A. Musetti, P. Ciancarini.Aemoo: exploring knowledge on the web , In:
Proceedings of the 5th Annual ACM Web Science Conference .ACM, pp. 272-275, Paris, France, 2013.

• A. Gangemi,A. G. Nuzzolese,V. Presutti, F. Draicchio,A. Musetti, P. Ciancarini.Automatic typing of DBpedia entities .
In: J. Hein,A. Bernstein, P. Cudre-Mauroux, editors, Proceedings of the 11th International Semantic Web Conference
(ISWC2012). Springer, pp. 65-91, Boston, Massachusetts, US, 2012.

• A. G. Nuzzolese. Knowledge Pattern Extraction and their usage in Exploratory Search. In: J. Hein,A. Bernstein, P.
Cudre-Mauroux, editors, Proceedings of the 11th International Semantic Web Conference (ISWC2012) . Springer,
pp. 449-452, Boston, Massachusetts, US, 2012.

• A. G. Nuzzolese,A. Gangemi,V. Presutti, P. Ciancarini. Encyclopedic Knowledge Patterns from Wikipedia Links . In: L.
Aroyo, N. Noy, C.Welty, editors, Proceedings of the 10th International Semantic Web Conference (ISWC2011) .
Springer, pp. 520-536, Bonn, Germany, 2011.

• A. G. Nuzzolese,A. Gangemi, andV. Presutti. Gathering Lexical Linked Data and Knowledge Patterns from
FrameNet . In M. Musen, O. Corcho, editors, Proceedings of the 6th International Conference on Knowledge
Capture (K-CAP) , pp. 41-48.ACM,Alberta, Canada, 2011.
38
Publications

39
Thank you

40

• FrameNet is an XML lexical knowledge base

• Cognitive soundness

• Grounded in a large corpus

• It consists of a set of frames, which have

• Frame elements

• Lexical units, which pair words (lexemes) to frames

• Relations to corpus elements

• Each frame can be interpreted as a class of situations

41
FrameNet

42
Natural Language Enhancer

43
Refactor

44
Knowledge Pattern Extractor

45
Boundary induction
Step Description
1 For each path, calculate the path popularity
2
For each subject type, get the N top-ranked path popularity
values
3
Apply multiple correlation (Pearson ρ) between the paths of all
subject types by rank, and check for homogeneity of ranks
across subject types
4
For each of the N path popularity ranks, calculate its mean
across all subject types
5 Apply clustering (e.g., k-means) on the N ranks
6
Decide threshold(s) based on the clustering as well as other
indicators (e.g., FrameNet roles distribution)

46
Contextualized views
• What is the information in the Web that provides the relevant
knowledge about Barack Obama as a Nobel Prize laureate?
From the Google Knowledge Graph
From wikipedia.org

• Linked Data is a breakthrough in Semantic Web for the creation
of the Web of Data

• The Web of Data offers large datasets for empirical research

• For the ﬁrst time in the history of knowledge engineering we
have datasets

• Created by large communities of practice

• With a lot of realistic data

• On which experiments can be performed

• The Semantic Web can be founded as an empirical science

• In our vision KPs are the research objects of the Web as an
empirical science
47
The Web of Data

• They are archetypal solutions to common and frequently
occurring design problems

• They were introduced in the seventies by the architect and
mathematician Christopher Alexander.

“a good architectural design can be achieved by means of a set of rules that are packaged in
the form of patterns, such as “courtyards which live”, “windows place”, or “entrance
room” [Alexander 1979]

• They enable design based on reuse

• Software Engineering has eagerly borrowed design patterns

“. . . designers […] look for patterns to match against plans, algorithms, data structures,
and idioms they have learned in the past. . .” [Gamma et al. 1993]
48
Design Patterns

• Ontologies are artefacts that encode a description of some
world

• Like any artefact, they have a lifecycle: they are designed, implemented,
evaluated, ﬁxed, exploited, reused, etc.

• An Ontology Design Pattern (ODP) [Gangemi and Presutti
2009] is a modeling solution to solve a recurrent ontology
design problem

• Reusability in Ontology Engineering
49
Ontology Design Patterns

• A Knowledge Pattern is a small, well connected and recurrent
unit of meaning, which provides a semantic interpretation for
a symbolic schema. It is

• task based: a KP is associated to an explicit task typically
expressed by means of competency questions

• well-grounded: a KP enables access to big data
• cognitively sound: a KP closely mirrors the human ways of
organizing knowledge

50
A deﬁnition for KP

Knowledge Patterns for the Web: extraction, transformation, and reuse

More Related Content

What's hot (20)

Similar to Knowledge Patterns for the Web: extraction, transformation, and reuse (20)

More from Andrea Nuzzolese (7)

Recently uploaded (20)

Knowledge Patterns for the Web: extraction, transformation, and reuse