Knowledge Representation and Reasoning
Concept-based versus linguistic representation
AI knowledge representation systems are usually concept-based
and thus independent of specific natural languages.
Semantics
A knowledge representation system should have a well-defined
semantics which identifies the "meaning" of concepts, relations and
statements.
A concept or relation is satisfiable if it
has a model. For example, an existing white swan is a model for
"Swans are white".
A statement is valid if it is satisfiable for all models.
For example, an existing black swan shows that "All swans are white"
is not valid.
Satisfiability and validity depend on the context of statements.
For example, "Unicorns are animals" can be valid in a fairy tale
context but is not considered valid in normal everyday life.
Relations
Relations can be transitive, for example, "Snoopy ISA Dog",
"Dog ISA Mammal" therefore "Snoopy ISA Mammal". The ISA relation
should always be transitive, other relations, such as "HAS" or part-whole
relations are not always transitive.
Relations require quantification. For example, does "Dogs
eat dogfood" mean "all dogs eat some dogfood" or "some dogs eat
all dogfood" or "some dogs eat only dogfood" or ...?
Description Logics
Description Logics is an improved version of the knowledge representation
language KL-ONE. It distinguishes between terminologic knowledge
(TBox),
such as "poodles are dogs" or "people eat food", and assertional
knowledge (Abox), such as "Snoopy is a dog" or "Snoopy is eating that
piece
of meat". Semantic networks can be formalized in description
logics.
Tasks for knowledge representation systems
Acquisition: new information is integrated into the system
Retrieval: existing information is retrieved, query answering
Reasoning: checking whether concepts, relations and assertions
are satisfiable/valid; checking whether the knowledge base is
consistent, which means that all concepts and relations are
satisfiable and all statements are valid.
Reasoning is used for acquisition because, for example,
new concepts must be integrated into the existing ISA hierarchy
without creating inconsistencies. Reasoning is also used for retrieval
to retrieve implicit knowledge.
Challenges for reasoning strategies and knowledge representation
Complexity: the reasoning tasks must not take too much time or
computing resources
Incomplete or uncertain knowledge must be handled by the system