On Compositionality of ISO 25964 Hierarchical Relations (BTG, BTP, BTI)

Vladimir Alexiev, Ontotext Corp
Jutta Lindenthal, Consultant to digiCULT-Verbund eG
Antoine Isaac, Europeana and VU Amsterdam


Networked Knowledge Organization Systems (NKOS 2014) Workshop

DL2014 Conference, London, 12 Sep 2014


2D interactive version, pdf, slideshare.

Press O for overview, H for help.

Proudly made in plain text with reveal.js, org-reveal, org-mode and emacs.

Table of Contents

BTG, BTP, BTI

KOS have used different kinds of hierarchical relations for a long time:

Relation Abbr Name Example
broaderGeneric BTG Genus/Species Relation mineral BTG inorganic material (AAT)
      Iceland spar BTG calcite (AAT)
broaderPartitive BTP Part/Whole Relation Tuscany BTP Italy (TGN)
broaderInstantial BTI Kind/Instance Relation Rembrandt van Rijn BTI person (ULAN)
      SG Dynamo Dresden BTI football clubs (GND)

Use of BTG, BTP, BTI in Thesauri

Examples:

Definitions in ISO 25964

  • BTG: amenable to logical all-and-some test
    • Children should all be a type, or kind of the parent
    • From the parent's point of view, it encompasses only some of any given child
    • Can be concluded it's comparable to rdfs:subClassOf (isA): transitive
  • BTP: part of entity/system belongs uniquely to particular possessing whole in any context
    • The part may not belong to more than one whole, and BTP has to be universally valid
    • AAT Guidelines: "Each child should be part of the parent and all ancestors above it" (transitive)
  • BTI: individual instance to general class
    • Instances often represented by proper name (also called "classes of one")
    • Instances may not have further BTI nor BTG
    • But may be further subdivided: use custom relationship (BTX), eg BTS (subdivision)

GVP Hierarchical Relation Counts

hier-rel-per-type.png

GVP Hierarchical Relations

  • AAT: most are BTG, but there is a variety of BTP:
    • (C) BTP (C): calendars of relics BTP cabinets of relics
    • (C) BTP (G): anvil components BTP
    • (G) BTP (C): BTP jewelry
    • (G) BTP (G): BTP
    • (C) BTP (H): building divisions BTP Single Built Works
  • TGN: all are BTP
    • placeType: in the current TGN LOD (2.0) has no relation to BTI
    • May reconsider and make it subprop of BTI, see TGN Place Type Relation discussion paper
  • ULAN: most are BTI, e.g. Rembrandt (ULAN) Persons facet (ULAN)
    • May consider more specific, eg Rembrandt (ULAN) BTI Painters (AAT)
    • Some BTP, e.g. corporate body BTP another corporate body

GVP Hierarchy Structure

  • Subjects include (C)oncepts; but also: (F)acets, (H)ierarchies, (G)uide Terms
    • Not for indexing, only to structure. Implemented as iso:ThesaurusArray
    • G and C can be intermixed: F>H>(G|C)
007-subject-hierarchy.png

SKOS/ISO vs GVP Impedance Mismatch

  • SKOS and ISO define Standard Hierarchical Relations
    • Only between Concepts
    • skos:broader, iso:broaderGeneric, etc
  • We define custom GVP Hierarchical Relations
    • Connect the hierarchy uniformly
    • gvp:broader, gvp:broaderGeneric, etc
  • We infer appropriate standard relations when they connect concepts directly
    • Notice the "thread-through" skos:narrower in the prev diagram

Problem Statement

What are the appropriate combinations (compositions) of BTG, BTP, BTI?

  • Matters with respect to appropriate closure for information retrieval
  • It's a prerequisite for sensible search expansion
  • Has not been systematically analyzed to date

The problem with broaderTransitive

ISO 25964 formalized BTG, BTP, BTI as sub-properties of skos:broader

  • skos:broader contributes unconditionally to skos:broaderTransitive
    (broaderGeneric|broaderPartitive|broaderInstantial) => broader => broaderTransitive
  • Ambiguities in representing thesauri using extended SKOS - examples from real life (NKOS 2012):
    broaderTransitive should be established only for BTG and BTP, but not for BTI, nor mixed paths BTG+BTP
  • skos:broaderTransitive may include semantically inappropriate statements
    • a place inherits all place types of its parents
    • eg: Sofia BTP Bulgaria BTI country => Sofia BTI country
  • Lively discussion at SKOS mailing list from Nov 2013 to April 2014

The Time Has Come!

We have to resolve this issue for GVP LOD representation:

  • To infer appropriate thread-through standard relations
  • Makes sense to represent TGN place types and ULAN actor roles as BTI
    • But skos:broaderTransitive causes confusion and bloat (100M=>400M statements)
  • So we want to infer only appropriate compositions (see BTG, BTP, BTI Inference)
    • BTGE, BTPE, BTIE (gvp:broaderGenericExtended, gvp:broaderPartitiveExtended, gvp:broaderInstantialExtended)
    • Their disjunction gvp:broaderExtended
    • (Also gvp:broaderPreferred and gvp:broaderPreferredExtended)

BTG, BTP, BTI Compositionality

  • Basic decision table. BT*x means BT*|BT*E
  BTGx BTPx BTIx
BTGx BTGE BTPE no
BTPx BTPE BTPE no
BTIx BTIE no no
  • BTG=>BTGE, BTP=>BTPE, BTI=>BTIE: basic inferences

BTG, BTP, BTI Compositionality (2)

  • BTGx/BTGx=>BTGE
    • If X is kind of Y and Y is kind of Z then X is kind of Z
    • Eg: racehorses BTG BTG Equus caballus => racehorses BTGE Equus caballus
  • BTGx/BTPx=>BTPE
    • If X is kind of Y and Y is part of Z then X is part of Z (X can play the role of Y)
    • Eg: beak irons BTG anvil components BTP beak irons BTPE
  • BTGx/BTIx=>n/a
    • A generic concept may not be hierarchically subordinate to an instance.
    • The understanding of instance, as described in ISO, excludes this composition

BTG, BTP, BTI Compositionality (3)

  • BTPx/BTGx=>BTPE.
    • If X is part of Y and Y is kind of Z then X is part of Z (Z can play the role of Y)
    • Eg: anvil components BTP BTG => anvil components BTPE
  • BTPx/BTPx=>BTPE
    • If X is part of Y and Y is part of Z then X is part of Z
    • Eg: Sofia BTP Bulgaria, Bulgaria BTP Europe, so Sofia BTP Europe
    • But see mereological exceptions/imprecisions below!
  • BTPx/BTIx=>no
    • Counter-example: Sofia BTP Bulgaria BTI country. But Sofia is no country

BTG, BTP, BTI Compositionality (4)

  • BTIx/BTGx=>BTIE
    • If X is instance of Y and Y is kind of Z, then X is instance of Z (Z can play the role of Y)
    • Eg: Mt Athos BTI orthodox religious center BTG Christian religious center => Mt Athos BTIE Christian religious center
  • BTIx/BTPx=>no
    • Counter-example: Statue of Liberty pedestal BTI pedestals BTP statues. That particular pedestal is neither BTI nor BTP statues in general
    • But see "beyond paths" below
  • BTIx/BTIx=>n/a
    • An instance as a class of one cannot have instances
    • But see BTI Elaborations below

Usage: Inferring ISO relations

anvils-components.png

Inferring ISO relations

Secretariat-inference.png

Inference Dependencies

hierarchicalRelations-simplified.png (A bit simplified, see GVP Hierarchical Relations Inference)

Usage 2: Query Expansion in Information Retrieval

The main purpose of a proper broader relation is to enable query expansion in information retrieval, eg:

  • Sofia BTP Bulgaria BTP Europe => Sofia BTPE Europe
    • Enables a search for places in Europe to also find Sofia
  • Mt Athos BTI orthodox religious centers BTG Christian religious centers BTG religious centers => Mt Athos BTIE religious centers
    • Enables a search for religious centers to also find Mt Athos

Usage 3: Beyond Chain Inferences

If X necessary BTP Y and Z BTI X and T BTI Y then Z BTP T

statue-pedestal.png statue-pedestal.png

Usage 3: Beyond Chain Inferences

If X necessary BTP Y and Z BTG Y then X BTP Z

keyboard-instrument.png

Usage 4: Quality Checking

swell-boxes.png
  • "swell boxes" BTG "organ components" BTP "organs (aerophones)" => BTPE
  • "swell boxes" BTG "organs (aerophones)" is asserted in error
  • Can catch it if we declare BTGE & BTPE as owl:disjointProperty
  • But is this true in all cases?

BTP Imprecisions

  • Mixing partial vs full inclusion; and physical vs administrative:
    Netherlands Antilles BTP Netherlands BTP Europe ??
NetherlandsAntilles.png

BTP Imprecisions (2)

  • Mixing Partial vs full inclusion:
    Istanbul BTP Turkey BTP Asia

Turkey.png How about Istanbul BTP Europe? It does straddle the Bosphorus strait:

Istanbul_and_Bosporus.jpg

BTP Imprecisions (3)

  • Mixing member vs substance meronym:
    chicken feet BTP chicken BTP chicken soup ??
  • Mixing intrinsic vs extrinsic BTP; and categories (person vs group):
    Mick Jagger's BTP Mick Jagger BTP The Rolling Stones ??

Mereology is a complex topic spanning: philosophy, mathematical logic, theoretical computer science, physics, Sheaf, Topos, or Category Theory, object-oriented programming.

BTI Elaborations

  • Metaclasses in OOP and Punning in OWL allow classes of classes, and use them profitably
  • ISO: instance may have parts/subdivisions, recommends custom relation BTX (eg BTS=subdivsion)).
    Eg "BMW E87" BTS "BMW 1 Series" BTI "Automobiles"
  • Biological classification: concepts belong to different levels (taxonomic ranks).
    Eg Secretariat (ULAN ) BTI racehorses BTG Equus caballus BTI species

BTI in Taxonomic Relations

Secretariat.png

Do Individuals Belong in a Thesaurus?

  • TGN gave up placeType
  • If you exclude BTI then broaderExtended coincides with broaderTransitive: BTG* | (BTG|BTP)* = (BTG|BTP)*
  • Some CRM SIG members: "Individuals don't belong to a thesaurus. Mixing individuals and generics is logically inconsistent"
  • Eg in Getty LOD: tgn:7009977 London is gvp:adminplaceconcept, gvp:subject, skos:concept

Do Individuals Belong in a Thesaurus? (2)

We think yes: main role of a thesaurus is a list of fixed values (concepts, people, etc)

  • Eg GND mixes 10M things: materials, subjects, football clubs, deities, ghosts
  • Eg British Museum LOD: London England is ecrm:E53_Place, skos:Concept (but latter may be removed)
  • Eg LoC MARC Relators: Author is skos:Concept, rdf:Property, owl:ObjectProperty !

Thanks for your time!

QuestionMark.jpg