Kein Folientitel - Fachgebiet Wissensverarbeitung

Knowledge representation as surrogate
D. Wissensrepräsentation
„Knowledgeȱrepresentation is mostȱfundamentallyȱaȱsurrogate,ȱaȱsubstituteȱ
forȱtheȱthingȱitself,ȱusedȱtoȱenableȱanȱentityȱtoȱdetermineȱconsequencesȱbyȱ
thinkingȱratherȱthanȱacting.ȱ[...]ȱReasoning is aȱprocess that goes onȱ
internally [ofȱaȱperson or program],ȱwhile most things it wishes toȱreason
about exist only externally.“
Slide 1
Prinzipien der Wissensrepräsentation
Slide 3
Knowledge representation as as set ontological commitments
Davis, Shrobe, and Szolovits (1993) discussed five principles for a knowledge
representation:
Aȱknowledge representation “is a set ofȱontological commitments,ȱi.e.,ȱanȱ
answer toȱthe following question:ȱ‚Inȱwhat terms should we think about the
1.
2.
3.
4.
5.
A knowledge representation is a surrogate,
a set of ontological commitments,
a fragmentary theory of intelligent reasoning,
a medium for pragmatically efficient computation,
and a medium of human expression.
world?ȱ[...]ȱInȱselecting any representation,ȱwe are [...]ȱmaking aȱset ofȱ
decisions about how andȱwhat toȱsee inȱthe world.ȱ[...]ȱWe (andȱour
reasoning machines)ȱneed guidance inȱdeciding what inȱthe world toȱattend
to,ȱandȱwhat toȱignore.“
• The way a knowledge representation focuses on these principles characterizes its
spirit.
• Each knowledge representation is a trade-off between these principles.
Slide 2
Slide 4
Knowledge representation as theory of intelligent reasoning
Knowledge representation as medium of human expression
„The initial conception ofȱaȱknowledge representation is typically motivated
by some insight indicating how peopleȱreason intelligently,ȱor by some
beliefȱabout what it means toȱreason intelligently atȱall.ȱ[...]ȱȱ
„Knowledge representations are [...]ȱthe medium ofȱexpression andȱ
It is aȱfragmentary theory ofȱintelligentȱreasoning,ȱexpressed inȱterms ofȱthree
components:ȱ
(i)
communication inȱwhich we tell the machine (andȱperhaps one another)ȱ
about the world.ȱ[...]ȱ
the representation‘s fundamentalȱconception ofȱintelligentȱreasoning;
(ii)ȱthe set ofȱinferences the representation sanctions;
Knowledge representation is thus aȱmedium ofȱexpression andȱ
(iii)ȱandȱthe set ofȱinferences it recommends.
communication for the use by us.ȱ[...]ȱ
The authors consider five fields which have provided notions of intelligent
reasoning:
• Mathematical Logic (e.g., Prolog)
Aȱrepresentation is the language inȱwhich we communicate,ȱhence we must
• Psychology (e.g., frames)
be able toȱspeak it without heroic effort.“
• Biology (e.g., neural networks)
• Statistics (e.g., bayesian networks)
• Economics (e.g., rational agents)
(Siehe drei Folien weiter.)
Slide 5
Slide 7
Wissen
Knowledge representation as medium for efficient computation
Es gibt verschiedene Wurzeln der Wissensverarbeitung, die zu
unterschiedlichen Modellen von Wissen führen:
Knowledge representation „is aȱmedium for pragmatically efficient computation,ȱ
i.e.,ȱthe computational environment inȱwhich thinking is accomplished.ȱOneȱ
• Biologie: Vernetzung
Æ neuronale Netze
contribution toȱthis pragmatic efficiency is supplied by the guidance aȱ
representation provides for organizing information soȱasȱtoȱfacilitate
• mathematische Logik: Deduktion
Æ Logikkalküle, Prolog
making the recommended inferences.“
• Statistik: Unsicherheit
Æ Fuzzy-Logik, Bayessche Netze
• Psychologie: Begriffe
Æ wissensbasierte Systeme
Slide 6
• Ökonomie: Ziele, Bewertungen
Æ Agenten
Slide 8
Neuronale Netze
Neuronale Netze
Vorbild Biologie
• Wir werden in dieser Vorlesung den Schwerpunkt auf die begriffliche
Wissensrepräsentation legen.
Gehirn
• Vorher aber ein kleiner Ausflug in die neuronalen Netze, um die
Verschiedenheit der Ansätze zu beleuchten.
In der Vergangenheit wurden
Forscher aus den verschiedensten
Fachgebieten durch das biologische
Vorbild motiviert.
Geflecht aus Neuronen
• Neuronale Netze werden detailliert in der Vorlesung „Knowledge
Discovery“ besprochen.
Das Gehirn besteht aus
• ca. 1011 Neuronen, die mit
• ca. 104 anderen Neuronen
• ca. 1013 Synapsen
verschaltet sind.
Slide 9
Konnektionistische Wissensrepräsentation
Slide 11
Neuronale Netze
Vorbild Biologie
(Künstliche) Neuronale Netze
sind massiv parallel verbundene Netzwerke aus
• einfachen (üblicherweise adaptiven) Elementen in
• hierarchischer Anordnung oder Organisation,
die mit der Welt in der selben Art wie biologische Nervensysteme
interagieren sollen.
[Kohonen 1984]
Neuronen
Synapse
Slide 10
Slide 12
Neuronale Netze
x1
v11
x2
...
v12
y1
• Forschung:
• Modellierung & Simulation biologischer neuronaler Netze
• Funktionsapproximation
• Speicherung von Informationen
• ...
vnm
y2
w1
Wofür nutzt man sie?
xm
w2
y3
w3
...
yn
wn
• Anwendungen (z.B.):
• Interpretation von Sensordaten
• Prozeßsteuerung
• Medizin
• Elektronische Nase
• Schrifterkennung
• Risikomanagement
• Zeitreihenanalyse und -prognose
• Robotersteuerung
o
Slide 13
Slide 15
Neuronale Netze
Begriffliches Wissen
• Wir werden in dieser Vorlesung den Schwerpunkt auf die begriffliche
Wissensrepräsentation legen.
• Die psychologische Perspektive führt zu begrifflichen Modellen von
Wissen:
–
–
–
–
–
–
Semantische Netze
Frames
Begriffliche Graphen
Formale Begriffsanalyse
Beschreibungslogiken
Ontologien/Semantic Web
Figure 1: Neural Network learning to
steer an autonomous vehicle (Mitchell 1997)
Slide 14
Slide 16
Begriffliche Wissensrepräsentation
Begriffliche Wissensrepräsentation
Wir stellen nun kurz ein paar Formalismen zur Repräsentation begrifflichen
Wissens vor:
Was ist ein Begriff?
Begriff. Denkeinheit,ȱdieȱausȱeinerȱMengeȱvonȱGegenständenȱunterȱErmittlungȱ
derȱdiesenȱGegenständenȱgemeinsamenȱEigenschaftenȱmittelsȱAbstraktionȱ
gebildetȱwird.ȱȱ
•
•
•
•
•
•
•
Benennung. AusȱeinemȱWortȱoderȱmehrerenȱWörternȱbestehendeȱBezeichnung.
Definition. BegriffsbestimmungȱmitȱsprachlichenȱMitteln.
Letzteren werden wir im Verlauf der Vorlesung wiederbegegnen als formale
Grundlage für Wissensrepräsentation im WWW.
[DIN 2342 Teil 1, Begriffe der Terminologielehre - Grundbegriffe, 10.92]
The set ofȱcharacteristics that comeȱtogether asȱaȱunit toȱformȱthe concept is called
the intension.ȱThe objects viewed asȱaȱset andȱconceptualized into aȱconcept are
known asȱthe extension.ȱThe two,ȱextension andȱintension,ȱare interdependent.ȱForȱ
example,ȱthe characteristics making upȱthe intension ofȱ‚lead pencil‘ determines the
extension,ȱthose objects that qualify asȱlead pencils andȱvice versa.
[ISO 704: Terminology Work: Principles and Methods, ISO 2000]
Natural Language
Concept hiararchies
Existential Graphs
Semantic Networks
Frames
Conceptual Graphs
Description Logics
(siehe u.a. J. Sowa: Semantic Networks. http://www.jfsowa.com/pubs/semnet.htm )
Slide 17
Begriffliche Wissensrepräsentation
Slide 19
Knowledge Representation by Natural Language
DIN 2330: Begriffe und Benennungen. Allgemeine Grundsätze
Advantages of natural language:
Definition
Benennung
• Very expressive.
• Easy to understand for humans.
Begriff
Begriffsebene
Merkmal a
Merkmal b
Merkmal c
Problems with natural language (for automatic processing):
• Natural language is often ambiguous.
• Semantics are not formally defined.
• There is little uniformity in the structure of sentences.
Gegenstandsebene
Gegenstand 1
Gegenstand 2
Gegenstand 3
Eigenschaft A
Eigenschaft B
Eigenschaft C
Eigenschaft D
Eigenschaft A
Eigenschaft B
Eigenschaft C
Eigenschaft E
Eigenschaft A
Eigenschaft B
Eigenschaft C
Eigenschaft F
Slide 18
Slide 20
Existential Graphs
Knowledge Representation by Natural Language
Excursion: Workaround for Information Retrieval and Text Mining:
• “Bag of Words” Model (see lecture on Information Retrieval)
• (Is not considered as ‘Knowledge Representation’ in the narrow sense.)
• Introduced by Charles S. Peirce in 1897(!).
• Advantages:
– intuitive reading
– intuitive calculus
– formal semantics (but hard to extract from Peirce’s manuscripts)
•
•
• Disadvantage:
– non-linear notation difficult to represent in the computer
Advantages:
– easy computation in a vector space
– “similarity” between documents can be measured
Disadvantage:
– internal structure and meaning of the text gets lost
Slide 21
Concept hierarchies
•
going back to Aristotle/Porphyry
•
Advantages:
– easy to understand
– widely used
Disadvantage:
– not very expressive (but sufficient for many applications – eg,
the file system on your PC)
•
Slide 23
Semantic Networks
Slide 22
•
"Semantic Nets" were first invented for computers by Richard H. Richens of
the Cambridge Language Research Unit in 1956 as an „interlingua“ for
machine translation of natural languages.
•
They were developed by Robert F. Simmons in the early 1960s and later
featured prominently in the work of M. Ross Quinlan in 1966.
•
Did not provide clear, formal semantics. Æ Different implementations could
treat the same data differently!
Slide 24
Semantic Networks
Conceptual graphs
• Based on Peirce’s existential graphs
• Aim at providing a formal semantics, by mapping to predicate logic.
The example above shows default reasoning: Usually, elephants are grey.
• Problem: How are royal elephants treated? Æ The formalism is ambiguous.
Different implementations may treat the same data differently. A formal
semantic would answer this question.
•
• Advantages:
– intuitive reading
– expressive (beyond first order logic, see example above)
– provides linear notation for easier computation
Later, Description Logics where introduced to overcome this problem by
introducing such a semantic.
• Disadvantages:
– mapping to predicate logic is not (yet) well defined Æ formal semantics are still
missing
Slide 25
Frames
•
•
•
Slide 27
Description Logics
Much like a semantic network except each node represents prototypical
concepts and/or situations.
Each node has several property slots whose values may be specified or
inherited by default.
Frames are similar to object-oriented modelling
•
•
•
Advantages:
– is intuitive for many users
•
Disadvantages
– formal semantics are missing
•
•
Slide 26
Beschreibungslogiken sind eine Familie von entscheidbaren
Fragmenten der Prädikatenlogik.
Im Gegensatz zu Frames und semantischen Netzen bieten sie eine
formale Semantik.
Erster Vertreter: KL-ONE (1985)
Heutzutage Grundlage für das Semantic Web (Æ Kapitel E & F)
Slide 28
Conclusion
• There are many formalisms for representing conceptual knowledge.
• Depending on the purpose, a different formalism might be the most
suitable.
• In this light, we will discuss (in the next section) three formalisms
in more details that were established for the (Semantic) Web: XML,
RDF, OWL.
Slide 29