Information Ontologies for the Intelligence Communities A Survey of DCGS-A Ontology Work

1
Information Ontologies for the Intelligence
CommunitiesA Survey of DCGS-A Ontology Work

• Ron Rudnicki
• November 12, 2013

2
Topics

• The DCGS-A ontology suite
• Standard operating procedures and ontology
  quality assurance
• Annotation vs. Explication
• How the DCGS-A ontologies are being used for the
  explication of data models

3
The DCGS-A Ontology Suite
4
Motives for Ontology Development

• Part of a Big Data solution

• Multiple formats including free text,
  semi-structured and structured
• Some surprise data sets are made available a
  short time prior to system testing
• Data sets will change along with domain of
  interest
• Data can not be collected into a single store
• Provide cross-source searching and analytics
• Need to maintain the provenance of data

5
Contribution of the Ontologies

• Design choices affect the outcome

• Common Upper Level Ontology The ontologies
  extend from a common upper level ontology
• Delineated Content - Each ontology has a clearly
  specified and delineated content that does not
  overlap with any other ontology
• Composable Content Classes in the ontologies
  represent entities at a level of granularity that
  can be composed in various ways to map to terms
  in sources

6
Integration Through a Common Upper Level Ontology

• Encourages uniform representations of domains

Entity
Object
Quality
bearer_of
Organization
Quality of Physical Artifact
Quality of Organization
Physical Artifact
has_quality
has_quality

• Provides common patterns within the target
  ontology for mappings from the sources
• Easier to include new sources of data
• Enables more uniformity between queries
• Easier to transition to domains of interest

CUBRC - Proprietary
7
Integration Through Delineated Content

• Each class in the target ontologies is defined in
  one place

Entity
Object
Organization
Physical Artifact
Spatial Location
located_at
located_at

• Facilitates locating a class within the target
  ontologies
• Provides better recall in queries
• Less likely to overlook relevant data

CUBRC - Proprietary
8
Integration Through Composition of Classes
Data Source 2

• Granular classes better accommodate mappings
  from various perspectives on the same domain
  without loss of information

prescribes
Model
Car
has quality
Full Size
manufactures
Length of Wheelbase
Manufacturer
Mid Size
is nominally measured by
Compact
CUBRC - Proprietary
9
High Level Depiction of Domain

• Provides Coverage of Domain of Human Activity

Actions
to perform
People Organizations
Artifacts
that take place in
use
Natural Artificial Environments
are distinguished by
Time
Attributes
10
Developed Using a Top-Down Bottom-Up Strategy

• Partial List of Data Sources Used

• Treasury Office of Foreign Assets Control
  Specially Designated Nationals and Blocked
  Persons
• NCTC Worldwide Incidents Tracking System
• UMD Global Terrorism Database
• RAND Database of Worldwide Terrorism Incidents
• LDM version .60 (TED)
• VMF PLI
• DCGS-A Global Graph
• DCGS-A Event Reporting
• BFT Report (CCRi test data)
• Cidne Sigact (CCRi test data)
• Long War Journal
• Harmony Documents from CTC at West Point
• Threats Open Source Intelligence Gateway

11
Based Upon Standards

• Partial List of Doctrine and Standards Used

• DOD Dictionary of Military and Associated Terms
  (JP 1-02)
• JC3IEDM
• Counterinsurgency (FM 3-24)
• Operations (FM 3-0)
• Multinational Operations (JP 3-16)
• International Standard Industrial Classification
  of all Economic Activities Rev.4 (ISIC4)
• Universal Joint Task List (CJSCM 3500.04C)
• Weapon Technical Intelligence (WTI) Improvised
  Explosive Device IED Lexicon
• Information Artifact Ontology (IAO)
• Phenotype and Trait Ontology (PATO)
• Foundational Model of Anatomy (FMA)
• Regional Connection Calculus (RCC-8)
• Allen Time Calculus
• Wikipedia

12
Current DCGS-A Ontology Architecture
13
Ontology Metrics
Ontology Name Number of Classes Number of Relations Equivalent Class Axioms Subclass Of Axioms
Agent Ontology 986 71 378 1004
AIRS Emotion Ontology 73 88
AIRS Mid-Level Ontology 516 8 221 641
Artifact Ontology 298 3 310
Event Ontology 409 2 423
Extended Relation Ontology 45
Geospatial Ontology 297 14 13 316
Information Entity Ontology 83 29 21 83
Quality Ontology 681 2 681
Relation Ontology 20
Time Ontology 16 22 30
Totals 3359 209 640 (19) 3576 (106)
14
Standard Operating Procedures and Ontology
Quality Assurance
15
Semantic Conformance Testing

• Semantic Smuggling

• An importing ontology reuses a term from another
  and adds to its content in some way
• adds an axiom to some upper-level term.
• the imported class inherits content from parent
  classes of the importing ontology
• Corrective action
• request that the curators of the ontology that is
  the source of the class add the content
• If not possible, then plan for revision of import
  architecture
• the importing ontology should introduce a subtype
  of the term to which the content could then be
  added.

16
Semantic Conformance Testing

• Multiple Inheritance

• Defining a class to be a subtype of more than one
  superclass
• Corrective action
• remove any subclass assertions that are false
  (e.g. Bank subClassOf Organization, Bank
  subClassOf Facility)
• refactor superclasses into disjoint classes
• write axiom so that the multiple inheritance
  exists in the inferred hierarchy rather than the
  asserted hierarchy

17
Semantic Conformance Testing

• Taxonomy Overloading

• Extending an ontology by introducing terms as
  child terms of a higher-level ontology using
  another relation (e.g. part of, is narrower in
  meaning than)
• Corrective action
• Place the terms into their appropriate place in
  the taxonomy

18
Semantic Conformance Testing

• Containment

• a term from a lower level is not a subclass of
  any class of the ontologies it imports
• containment requires that the domain covered by a
  lower-level ontology be circumscribed by the
  domain covered by the higher-level ontology from
  which it extends.
• Corrective action
• Add the class (or an appropriate superclass) to
  the appropriate higher-level ontology
• Import a higher-level ontology that does provide
  a superclass

19
Semantic Conformance Testing

• Conflation

• an ontology includes information model assertions
  that are not true of the domain
• e.g. carrying over a not null constraint as in
  every person must have an email address
• Corrective action
• Make needed modifications to axiom (generally the
  source of such violations) so that it conforms to
  the domain
• e.g. every person that has purchased from
  amazon.com must have an email address

20
Semantic Conformance Testing

• Logic of Terms

• a class is a set-theoretic combination of other
  classes
• Corrective action
• Add the class as a new type (College or
  University gt Higher Education Organization)

21
Calculating Value of Ontology Terms

• Provide some basis for class inclusion/exclusion

• The content of ontologies used in an enterprise
  will be the subject of debate and possibly,
  disagreement
• Having one or more metrics that are proven
  measures of value would help resolve such
  disagreements
• Current methods are often applied to ontologies
  in their entirety (e.g. Swoogle), fewer are
  designed to evaluate value of ontology classes
  and properties

22
Calculating Value of Ontology Terms

• Statistical Methods Supplemented by Weightings

• A purely statistical method applied to an
  ontology as a graph will undervalue isolated
  terms that are of importance in a domain
• Importance, is at least a function of amount of
  use and criticality
• Usage is tractable to definition, criticality
  less so

23
Annotation vs. Explication
24
Mappings

• Value and Assessment

• Many of the purposes for which ontologies are
  built will be realized only to the degree to
  which they are linked to data
• One component of mapping is an act of translation
  and should be assessed on the degree of
  equivalence between source and target
• Another component of mapping is an implementation
  and should be assessed on performance criteria
  such as costs and scalability
• Techniques and technologies vary

An introductory overview can be found at
http//www.w3.org/2005/Incubator/rdb2rdf/RDB2RDF_S
urveyReport_01082009.pdf
25
Mappings

• Subtypes

• Hashtags the subjective assignment of uncurated
  keywords to a source
• Annotations rule based assignment of curated
  terms to a source
• Machine maps automated, structure-based
  translation of source into target vocabulary
• Definitions rule based expansion of source
  terms into types and differentiating attributes
• Explications rule based translation of all
  semantic content (including that which is
  implicit) of a source using terms and relations
  of the ontology

Term Mappings
Assertion Mappings
26
Mappings

• Pros and Cons

• Term mappings
• Can be automated
• Enable faceted queries (Select JFK as type
  Airport)
• Can result in significant loss of information
• Not reuseable
• Assertion mappings
• Manual process that does not scale
• Requires extensive knowledge of the target
  ontology
• Enables navigational queries
• Improves integration of data sources
• Can result in significant carry over of source
  information
• Not reuseable

27
Assessing Current Mapping Methods
No Ideal Instances
Low
Hashtags
Annotations
Machine Maps
Definitions
Time/Money
Explications
High
Translation
Lossy
Lossless
28
Examples of Mappings
A Source of Data About Cities
CityId Name State IncorporationDate Area Coordinates
1 Tampa Florida July 15, 1887 170.6 sq mi. 27 5650 N 82 2731 W
2 Boston Massachusetts March 4, 1822 89.63 sq. mi. 42 2129 N 71 0349 W
3 Dallas Texas February 2, 1856 385.8 sq. mi. 32 4658 N 96 4814 W
4 Los Angeles California April 4, 1850 503 sq. mi. 34 03 N 118 15 W
29
Explication of the Source as an End Point
Coordinates
City Name
City
designated_by
designated_by
part_of
Area
delimits
State
has_quality
City Government
designated_by
State Name
participates_in
Act Of Incorporation
occurs_on
Incorporation Date
30
Explication Implementation Example

• A Portion of a D2RQ File Mapping Birth Place and
  Date

• mapPersonBirth
• rdftype d2rqClassMap
• rdfslabel "Person Birth"
• d2rqclass eventBirth
• d2rqclassDefinitionLabel "Treasury OFAC
  Person Birth"
• d2rqdataStorage mapKDD-02-B-Treasury-SDN
  
• d2rquriPattern "treasurydata_PersonBirth/_at__at_
  TreasuryPerson.idurlify_at__at_" .
• mapPersonBirthTemporalInterval
• rdftype d2rqClassMap
• rdfslabel "Person Birth Temporal Interval"
  
• d2rqclass spanTemporalRegion
• d2rqclassDefinitionLabel "Treasury OFAC
  Person Birth Temporal Interval"
• d2rqdataStorage mapKDD-02-B-Treasury-SDN
  
• d2rquriPattern "treasurydata_PersonBirthTem
  poralIdentifier/_at__at_TreasuryPerson.idurlify_at__at___at__at_Tre
  asuryPerson.dateofbirthlist_uidurlify_at__at_" .
• mapPersonBirthTemporalIntervalIdentifier
• rdftype d2rqClassMap
• rdfslabel "Person Birth Temporal Interval
  Identifier"

31
Explication Current Method

• The full mapping of birth place and date consists
  of 16 such blocks
• The full mapping of the entire table consists of
  150 such blocks
• If the ontologies change, so must the mappings
• Common patterns in the ontologies make some
  re-use possible by adding placeholders to
  portions of maps and replacing them with specific
  values for the source at hand.
• Applications exist or are under development to
  auto-generate initial mappings that a human can
  then edit

32
Explication Current Method

• The improvements are source and implementation
  specific
• What works for structured sources mapped in D2RQ
  cant be reused in structured sources mapped in
  other languages (R2RML, EDOAL)
• Separate mappings would be needed for sources
  expressed in XML, HTML or free text
• Another solution is needed

33
How the DCGS-A Ontologies are Being Used for the
Explications of Data Models
34
Start with Machine Made Assertion Mappings

• Type to Type mapping (e.g. table column to class)
• Relationships between types expressed using a
  default generic object property
• Meta-data about the source entity (e.g. table
  name, column name, element name) is mapped to
  annotation properties (rdfslabel)

35
Machine Made Assertion Mapping as a Starting Point
Name
Coordinates
has_coordinates
has_name
City
has_area
Area
has_state
State
has_incorporation_date
Class mappings created by associating the
container with the components with a generic
property
Incorporation Date
36
Current Content of Ontologies is not Well Used

• Ontologists are trained to associate subclass and
  equivalence axioms to classes
• OWL reasoners dont expand the graph by creating
  instances based upon these axioms
• OWL reasoners are resource expensive and often
  result in unimpressive output
• Not much control can be exerted upon which
  inferences an OWL reasoner performs

37
Create a Library of Rules

• Change the relationship and type of the name of a
  city

• CONSTRUCT ?city exdesignated_by ?cityname .
• ?cityname rdftype exCityName .
• WHERE
• ?city rdftype exCity .
• ?cityname rdftype exName .
• ?city ?related_to ?cityname .
• NOT EXISTS ?city exdesignated_by ?cityname .

38
Create a Library of Rules

• Delete the original relationship and type

• DELETE ?city ?related_to ?name .
• ?name rdftype exName .
• WHERE
• ?city ?related_to ?name .
• ?name rdftype exName .
• ?city exdesignated_by ?name .
• ?name rdftype exCityName .

39
The Affect of Such Rules on Translated Data
city_name_1
coordinates _1
has_text_value
Tampa
designated_by
has_value
designated_by
27 5650N 82 2731W
city_1
part_of
area_1
state_1
has_quality
designated_by
has_value
state_name_1
delimits
170.6 sq. mi.
has_value
city_government_1
Florida
is_output_of
act_of_incorporation
has_incorporation_date
act_of _incorporation_1
has_value
occurs_on
July 15, 1887
act_of _incorporation_2
40
Benefits of Rule Library

• No need to write different rules for different
  source formats
• Changes to the ontology affect a single rule
  rather than some (possibly large) number of
  mappings
• Allows mappings from source to target to be
  simple and possibly fully automated
• Writing of rules can be performed by SMEs
• Fine grained control of which rules are executed
• by user group
• above a stated level of priority (weighting)