INTRODUCTION TO ARTIFICIAL INTELLIGENCE Massimo Poesio LECTURE 4: Semantic Networks and Description Logics SEMANTIC NETWORKS • Around end of the ’70s researchers in Psychology thought they had found evidence that knowledge was organized more or less as expected on the basis of ideas about taxonomies • This led to the development of formalisms for knowledge representation in AI called SEMANTIC NETWORKS • Soon researchers like Schubert (1975) and Hayes (1979) demonstrated that these formalisms were just notational variants of logic • The ‘logical’ approach to semantic networks has however led to the development of so-called DESCRIPTION LOGICS, a family of logics which also includes logics with better computational properties than first order logic SOME RESULTS FROM COGNITIVE PSYCHOLOGY: SEMANTIC NETWORKS • Collins & Quillian, 1969: knowledge appears to be organized around objects and in a taxonomic way – A canary is yellow – A canary has feathers – A canary eats food • Haviland & Clark 1974, Sanford & Garrod 1979: ‘associated’ knowledge available when concepts are mentioned – I looked around the house. – The lounge was very spacious. SEMANTIC NETWORKS • Hypothesis: commonsense knowledge is organized in networks whose nodes are types and instances of types, and whose relations encode – Taxonomic relations (as in Aristotle) – Attributes • The key inference that such theories want to model: INHERITANCE – Semantic networks also called INHERITANCE NETWORKS AN EXAMPLE OF SEMANTIC NETWORK ANIMAL eats food can move can fly has feathers BIRD scales CANARY sings yellow FISH swims AN EXAMPLE OF INHERITANCE BIRD has FEATHERS CANARY IS-A BIRD ∴ CANARY has FEATHERS DESCRIPTION LOGICS • Brachman & Levesque (1985) proposed a formal approach to knowledge bases organized as semantic networks, encoding inheritance reasoning SPECIFYING A KNOWLEDGE BASE: TBOX AND ABOX • According to Description Logics, a knowledge base contains two types of knowledge: – Generic knowledge about concepts, contained in the TBOX (ie: SEMANTIC MEMORY) • Bicycles have two wheels • Parents have children – Knowledge about the instances of these concepts, contained in the ABOX (ie: EPISODIC MEMORY) • Massimo’s bicycle is grey • Distinct logical languages for each of them EXAMPLE OF TBOX KEY TERMS • Nodes: CONCEPTS • Subtype relation: ISA • Properties: ATTRIBUTES or ROLES CONCEPT DEFINITION SYNTAX • Intersection of concepts: C ∩ D – E.g., ANIMAL ∩ FLY – Interpretation: ANIMAL(x) ∩ CANFLY(x) • Attributes: ∃R.C – E.g., ∃hasFeather.FEATHER • Value restriction: ∀R.C – E.g., ∀hasWheel.WHEEL • Number restriction: (≤ n R), (≥ n R) – E.g., (≤ 2 hasWheel) • Negation: ¬ C – E.g., ¬ FEMALE EXAMPLES OF COMPLEX CONCEPTS • • • • • • BIRD ∩ YELLOW ∩ SINGS ANIMAL ∩ RATIONAL PERSON ∩ ¬ FEMALE VEHICLE ∩(≤ 2 hasWheel) VEHICLE ∩∃hasEngine.ENGINE TBOX DEFINITIONS • NECESSARY AND SUFFICIENT – CANARY ≡ BIRD ∩ YELLOW ∩ SINGS – HUMAN ≡ ANIMAL ∩ RATIONAL – WOMAN ≡ PERSON ∩ FEMALE – MALE ≡ PERSON ∩ ¬ FEMALE – BICYCLE ≡ VEHICLE ∩(≤ 2 hasWheel) ∩ ¬∃hasEngine.ENGINE • PRIMITIVE – BEAR ⊂ ANIMAL SEMANTICS • TBOX concepts denote SETS – ∩ denotes INTERSECTION – ¬ denotes COMPLEMENTATION – Etc • The resulting language is a subset of FOL ABOX DEFINITIONS • PERSON ∩ FEMALE (Maria) • hasCHILD(Maria,Gigino) INFERENCE IN DL • Description Logics were developed to model inheritance reasoning • In fact, they model a more complex form of reasoning: SUBSUMPTION • They are intended to be the COMPUTATIONALLY LEAST EXPENSIVE logics in which such reasoning is possible SUBSUMPTION • Concept D subsumes concept C, written C⊆D • If D is MORE GENERAL than C, i.e., if the set denoted by C is a subset of the set denoted by D EXAMPLE OF (TBOX) SUBSUMPTION: INHERITANCE BIRD≡ ANIMAL ∩ CANFLY ∩ ∃hasFeather.FEATHER CANARY ≡ BIRD ∩ YELLOW ∩ CANSING ∴ CANARY ⊆ CANFLY EXAMPLE OF (TBOX) SUBSUMPTION ENGINED_OBJECT ≡∃hasEngine.ENGINE CAR≡ VEHICLE ∩(= 4 hasWheel) ∩ ∃hasEngine.ENGINE ∴ CAR ⊆ ENGINED_OBJECT EXAMPLE OF (ABOX) SUBSUMPTION WOMAN ≡ PERSON ∩ FEMALE PERSON ∩ FEMALE (Maria) ∴ WOMAN(Maria) SUBSUMPTION AND MODERN PSYCHOLOGICAL THEORIES OF CONCEPTUAL KNOWLEDGE • As we will see in the next lectures, modern theories of concepts in cognitive science (since Rosch) have abandoned the position that conceptual knowledge is organized taxonomically in favour of the `clustering’ views from PROTOTYPE THEORY • Subsumption is a modern approach to inheritance that does NOT depend on the existence of special ISA links UNDECIDABILITY, COMPLEXITY, and LOGIC • One would want to have a logic as expressive as possible – ideally, as expressive as natural language • But there is a tight connection between the expressive power of a logic and the cost of reasoning with that logic • It is known from Goedel and Turing that FOL is undecidable • Even the propositional calculus is NP-complete THE COMPLEXITY OF DESCRIPTION LOGICS • The simplest form of DL is DECIDABLE and POLYNOMIAL (i.e., relatively efficient) • But even minor additions result in exponential complexity • DL-FOL is undecidable READINGS • Nardi & Brachman, An introduction to Description Logics, ch. 1 of Handbook of Description Logics (on the site)