Saltar al contenido ForoHackers

Knowledge representation: start here

Representación del conocimiento
#355 · Ophuscado · · Denunciar

Knowledge representation is how to write down what is known about a domain so that a program can use it and reason with it: facts, categories, relations and rules, in a form that is precise and can be queried. It is the symbolic side of AI, and it still matters wherever answers must be explained, checked against rules, or built from curated knowledge instead of learned patterns.

Key concepts

  • Logic: propositional and first-order logic give statements a precise meaning and a notion of valid inference.
  • Ontologies: shared vocabularies of classes, properties and constraints that define what exists in a domain.
  • Graphs of facts: subject, predicate and object triples that link entities, the basis of knowledge graphs.
  • Rules and inference: forward and backward chaining derive new facts from rules and old facts.
  • Description logics: restricted logics that keep reasoning decidable while describing classes and relations.

Vocabulary

  • Triple: one fact written as subject, predicate, object.
  • Reasoner: a program that derives the consequences of a knowledge base or detects contradictions.
  • Closed-world assumption: whatever is not known to be true is taken as false.
  • Taxonomy: a hierarchy of categories linked by is-a relations.
  • Query language: a declarative way to ask a knowledge base what it holds, such as SPARQL.

What to study, in order

  1. Propositional and first-order logic: truth tables, quantifiers, entailment.
  2. Write facts and rules in a logic programming language such as Prolog, and query them.
  3. Model a small domain as an ontology and as a table of triples, and compare how each is queried.
  4. Run a reasoner over your model, and see which facts it adds and which contradictions it finds.
  5. Study how symbolic knowledge and learned models can combine, and where each is more reliable.

Related tags

#logic #knowledge-graphs #databases #formal-methods #machine-learning

Inicia sesión o únete para responder.