KNOWLEDGE MANAGEMENT AND APPLICATIONS David Snchez Department of - - PowerPoint PPT Presentation

KNOWLEDGE MANAGEMENT AND APPLICATIONS

April 2013

David Sánchez

Department of Computer Science and Mathematics

Tarragona

2

The university

3

 Created in 1991  52 programmes of study  Over 12,000 students

The faculty

4

 Engineering degress

 Computer science  Telematics

 Masters

 Computer Security and

Intelligent Systems

 Artificial Intelligence  Security of the Information and

Communication technologies

 Doctoral program

 Computer Engineering

Research group

5

 Data privacy and electronic commerce  Privacy and security in mobile environments  Private information recovery and codes  9 professors and lecturers  6 post doctoral researchers  7 Ph.D. students  7 Research assistants

Contents

6

 Introduction  Knowledge acquisition  Semantic operators  Applications to privacy

Motivation

7

 Numerical data is easy to manage and

transform

 3<4 = true  (1+2)/2 = 1.5  {3, 2, 5} -> {2, 3, 5}

 A plethora of algorithms rely on aritmetical

functions to deal with numerical data

Motivation

8

 What about text?

 Car ¿>? bike  (apple + orange) / 2 = ??  {flu, cold, pneumonia} -> {?, ?, ?}

 Arithmetical functions do not make sense  Text (words, noun phrases) refers to concepts  Concepts should be managed according to

their formal semantics

Ontologies

9

 Provide a structured representation of a

shared conceptualization

 Elements

 Classes (concepts)  Instances (individuals)

 Semantics

 Properties (semantic relationships)  Restrictions (logical definition of meanings)

Contents

10

 Introduction  Knowledge acquisition  Semantic operators  Applications to privacy

Creating ontologies

11

 Manually

 Knowledge formalization is challenging  Knowledge can be subjective  Time consuming

 Assisted

 Proactive knowledge modelling tools

 Wizards  Reasoners to check knowledge consistency

 Knowledge engineering methods

 101, METHONTOLOGY, On-To-Knowledge

Ontology learning

12

 Semantics are implicitly referred in text  Textual corpora can be analysed to acquire

knowledge

 Discover concepts and individuals  Discover and label relations

 Taxonomic (cancer is a disease)  Non-taxonomic (cancer is treated with radiotherapy)  Attributes (cancer is non-contagious)

 Discover restrictions

 Axioms (Spain borders France -> France borders Spain)

Ontology learning from the Web

13

 Corpora: the Web

 The largest electronic repository  Heterogenous  It approximates the distribution of information at a

social scale

 Availability of massive IR tools: Web search

engines

Knowledge discovery from text

14

 NL processing tools to identify nouns, noun phrases

and named entities

 Concepts and individuals  Linguistic patterns to discover semantics  Taxonomic  “cities such as (Nimes)”, “cancers likes (melanoma)”  Non taxonomic  “cancer is treated with (surgery)”  Attributes  “camera has (10MP resolution)”, “camera features (3x zoom)”  Axioms (functionality, transitivity, symmetry, reflexibity, etc.)  “Spain borders France”, “France borders Spain” -> Symmetry

Retrieval of suitable corpora

15

 Create appropriate web search queries

 Taxonomic: “cities such as” […]  Non taxonomic: “cancer is treated with” […]  Attributes: “camera features” […]  Axioms: “Spain borders” & “France borders”

Statistical assessment

16

 Statistical assessor

 WSE page count approximates query

probabilities at a social scale

 Use an association score to filter noisy

extractions

 Point-wise mutual information

References

17  Taxonomic learning

 David Sánchez, Antonio Moreno: Pattern-based automatic taxonomy

learning from the Web. AI Commununications 21(1): 27-48 (2008)

 Non-taxonomic learning

 David Sánchez, Antonio Moreno: Learning non-taxonomic relationships

from web documents for domain ontology construction. Data & Knowledge Engineering 64(3): 600-623 (2008)

 Attribute learning

 David Sánchez: A methodology to learn ontological attributes from the

• Web. Data & Knowledge Engineering 69(6): 573-597 (2010)

 Axiom learning

 David Sánchez, Antonio Moreno, Luis Del Vasto Terrientes: Learning

relation axioms from text: An automatic Web-based approach. Expert Systems with Applications 39(5): 5792-5805 (2012)

Contents

18

 Introduction  Knowledge acquisition  Semantic operators  Applications to privacy

Exploiting ontologies

 Structured knowledge enables a

semantically-coherent interpretation of textual data by

 Defining semantically-grounded operators  Semantic similarity is the most basic operator

 Similarity(apple, orange) > Similarity(apple, bike)

Semantic similarity

20

 Semantic similarity  Degree of taxonomical resemblance

 e.g., dogs and cats are similar as they are mammals

 Semantic relatedness  Other non taxonomic relationships are also considered

 e.g., car and wheel or pencil and paper

 Similarity measures can be grouped in several

families according to

 the type of knowledge exploited  the principles in which similarity estimation relies

Ontology-based similarity

21

Edge-counting measures

22

( , ) | min_ ( , ) | Distance a b path a b =

IC-based measures

23

Least Common Subsumer (LCS)

( , ) ( ( , )) Sim a b IC LCS a b =

IC-based semantic similarity

24

 IC calculus relies on probability

assessments

Based on corpora

Requires general and heterogeneous

corpora

Language ambiguity hampers results Data sparseness produce weak statistics

( ) log ( ) IC c p c = −

Ontology-based IC computation

25

 Assumption: concepts with many hyponyms in

an ontology are more probable to appear in corpora

 Concept probabilities are intrinsically

approximated according to taxonomic knowledge

 Number of hyponyms

( ) ( ) log hyponyms c IC c

• ntology_size

= −

Feature-based measures

26

( , ) common_features(a,b) Sim a b disjoint_features(a,b) =

References

27



Feature-based similarity measures

 Montserrat Batet, David Sánchez, Aïda Valls: An ontology-based measure to

compute semantic similarity in biomedicine. Journal of Biomedical Informatics 44(1): 118-125 (2011)

 David Sánchez, Montserrat Batet, David Isern, Aïda Valls: Ontology-based

semantic similarity: A new feature-based approach. Expert Systems with Applications 39(9): 7718-7728 (2012)



IC-based similarity mesures

 Based on corpora

 David Sánchez, Montserrat Batet, Aïda Valls, Karina Gibert: Ontology-driven web-based

semantic similarity. Journal of Intelligent Information Systems 35(3): 383-413 (2010)

 Based on ontologies

 David Sánchez, Montserrat Batet, David Isern: Ontology-based information content

• computation. Knowledge-Based Systems 24(2): 297-303 (2011)

 David Sánchez, Montserrat Batet: A New Model to Compute the Information Content of

Concepts from Taxonomic Knowledge. International Journal on Semantic Web and Information Systems 8(2): 34-50 (2012)

 David Sánchez, Montserrat Batet: Semantic similarity estimation in the biomedical

domain: An ontology-based information-theoretic perspective. Journal of Biomedical Informatics 44(5): 749-759 (2011)

Other semantic operators

28

 Semantic similarity/distance is the base to

develop other semantically-grounded

• perators over a sample of textual data

 Aggregation (mean/centroid)

1 2 1

( , ,..., ) arg min ( , )

n n c i i

Mean x x x distance c x

=

  =    

∑

Aggregation

29

Sample colic lumbago lumbago migraine pain appendicitis gastritis lumbago migraine

Mean candidates colic (1) lumbago (3) migraine (2) appendicitis (1) gastritis (1) pain (1) Sum dist colic 3 3 4 4 1 24 lumbago 3 2 5 5 2 19 migraine 3 2 5 5 2 21 appendicitis 4 5 5 2 3 34 gastritis 4 5 5 2 3 34 pain 1 2 2 3 3 17 ache 2 1 1 4 4 1 16 inflammation 3 4 4 1 1 2 27 symptom 2 3 3 2 2 1 22

Sorting algorithm

30

• Algorithm. Sorting procedure

Inputs: P (dataset) Output: P’ (P sorted) 1 Compute the mean of all values in P 2 Consider the most distant value f to the mean 3 Add f to P’ and remove it from P 4 while (|P| > 0) do 5 Obtain the least distant value r to f 6 Add r to P’ and remove it from P 7 Output P’

References

31

 Sergio Martínez, Aïda Valls, David Sánchez: Semantically-

grounded construction of centroids for datasets with textual

• attributes. Knowledge-Based Systems 35: 160-172 (2012)

 Sergio Martínez, David Sánchez and Aida Valls: A

semantic framework to protect the privacy of electronic health records with non-numerical attributes. Journal of Biomedical Informatics 46(2): 294-303

 Josep Domingo-Ferrer, David Sánchez, Guillem Rufian-

Torrell: Anonymization of Nominal Data Based on Semantic Marginality. Information Sciences. To Appear

Contents

32

 Introduction  Knowledge acquisition  Semantic operators  Applications to privacy

Statistical Disclosure Control

33

Publish anonymized data

Remove or mask data that could reveal personal identities

Collect data

Statistical Offices Identifiers Quasi-identifying attributes Confidential attributes D.N.I.

Birth place Birth year Occupation

Income

12345678 Valls 1962 Lawyer 2,300 87654321 Cambrils 1965 Judge 5,000

 Statistical Disclosure Control (SDC)

discipline aims at protecting statistical data in a way that:  Can be released and exploited  Without publishing information that could identify a concrete individual

Statistical Disclosure Control

34 34

K-Anonymity

For each combination of quasi-identifiers values, at least k records exist in the dataset with the same combination

Information loss

Statistical Disclosure Control

35

 Many masking methods have been designed to

build groups of k-anonymous datasets

 Focusing mainly on numerical attributes

 Our goal

 To apply semantically-grounded operators to existing

anonymization algorithms

 Recoding  Microaggregation  Resampling

 The anonymized output would preserve the meaning

• f the input

 Retain the analytical utility

Anonymization via microaggregation

36

holidays sports nature nature sports nature relaxation nature nature surfing sports relaxation nature 1 2 2 5 relaxation holidays holidays 2 activity

Centroid:

nature 5

Clusters:

event relaxation 2 holidays 2 surfing 1 sports 2 sports

Centroid:

nature calmness sports

Original dataset Anonymised dataset

surfing calmness sports nature nature sports nature calmness nature nature calmness calmness sports

Document sanitization

37

 Should

 Avoid disclosure  Preserve semantics

 We rely on the

Information Content

• f terms

 Highly informative

terms (high IC) are removed

Document sanitization

38

 Given an input document

 Ex: “Peter Greenow, from United States, suffers

from pancreatic cancer”

 Fix a sanitization threshold

 The most concrete concept to be known. (Ex:

“Cancer”)

 Extract noun-phrases from text

 “Peter Greenow”, “United States”, “Pancreatic

Cancer”

Document sanitization

39

 Compute the IC of each noun phrase and

detect sensitive ones

 IC(Peter Grenow) > IC(Cancer)  IC(United States) < IC (Cancer)  IC(Pancreatic Cancer) > IC (Cancer)

 Replace sensitive terms by generalized

versions (from an ontology) fulfilling the treshold

 Ex: “[Person], from United States, suffers from

[cancer]”

Publications

40  Data anonymization

 Sergio Martínez, David Sánchez, Aïda Valls: Semantic adaptive

microaggregation of categorical microdata. Computers & Security 31(5): 653-672 (2012)

 Sergio Martínez, David Sánchez, Aïda Valls, Montserrat Batet: Privacy

protection of textual attributes through a semantic-based masking method.Information Fusion 13(4): 304-314 (2012)

 Sergio Martínez, David Sánchez and Aida Valls: A semantic framework

to protect the privacy of electronic health records with non-numerical

• attributes. Journal of Biomedical Informatics. 46(2): 294-303

 Montserrat Batet, Arnau Erola, David Sánchez, Jordi Castellà-Roca:

Utility preserving query log anonymization via semantic resampling. Information Sciences. To appear.

 Document sanitization

 David Sánchez, Montserrat Batet and Alexandre Viejo. Automatic

general-purpose sanitization of textual documents. IEEE Transactions on Information Forensics and Security. To Appear.