What is Knowledge Representation and Reasoning? Tommie Meyer - PowerPoint PPT Presentation

What is Knowledge Representation and Reasoning? Tommie Meyer Department of Computer Science University of Cape Town, and Centre for Artificial Intelligence Research (CAIR) ICTAC 2018 tmeyer@cs.uct.ac.za http://www.cs.uct.ac.za/˜tmeyer 1 / 43

Artificial Intelligence 2 / 43

Knowledge Representation and Reasoning 3 / 43

Knowledge Representation and Reasoning 4 / 43

Knowledge Representation and Reasoning 5 / 43

Knowledge Representation and Reasoning 6 / 43

Outline of my talk ◮ Knowledge Representation and Reasoning ◮ An example of a KRR system ◮ Description logics ◮ Ontology engineering ◮ KRR and Deep Learning ◮ Conclusion 7 / 43

Knowledge Representation and Reasoning Knowledge Representation and Reasoning ◮ Finding appropriate ways of representing knowledge about a given domain of interest ◮ Developing methods to derive implicit consequences from explicitly represented information. The dominant approach is logic-based 8 / 43

Implicit consequences Person { disjoint } Italian English { disjoint,covering } Lazy LatinLover Gentleman Hooligan Example due to Enrico Franconi 9 / 43

Implicit consequences Simple reasoning example Person { disjoint } Italian English { disjoint,covering } Lazy LatinLover Gentleman Hooligan There aren’t any Latin lovers Therefore all Italians are lazy 10 / 43

Knowledge Representation and Reasoning All these sub-topics involve logic-based reasoning ◮ Even semantic networks 11 / 43

Example: Open domain question answering IBM’s Watson ◮ “An application of advanced Natural Language Processing, Information Retrieval, Knowledge Representation and Reasoning, and Machine Learning technologies to the field of open domain question answering” ◮ It comprehensively defeated the all-time best players in the American TV quiz show Jeopardy! 12 / 43

Example: IBM’s Watson Ken Jennings, famous for winning 74 games in a row on the TV quiz show, borrowing a line from a “Simpsons” episode: “I, for one, welcome our new computer overlords.” 13 / 43

Description Logics 14 / 43

Description Logics 15 / 43

Description Logics A logic-based formalisation of ontologies ◮ Decidable fragments of first-order logic ◮ Concepts: unary predicates; Roles: binary predicates ◮ A formal model-theoretic semantics Reasoning Services ◮ Instance checking ◮ Concept subsumption and equivalence ◮ Concept satisfiability and knowledge base consistency ◮ Worst case complexity terrible but performs well in practice 16 / 43

A partial description of the DL ALC Concept construction ◮ ¬ C ; C ⊓ D ; C ⊔ D ; ( ∀ R . C ) ; ( ∃ R . C ) TBox ◮ Subsumption: C ⊑ D ; Equivalence C ≡ D ABox ◮ C ( a ) and R ( a , b ) 17 / 43

A partial description of the DL ALC TBox ◮ Person ⊑ Animal ⊓ Biped; Person ≡ Man ⊔ Woman ◮ Mother ≡ Woman ⊓ ( ∃ ParentOf.Person ) ◮ MotherWithDaughters ≡ Mother ⊓ ( ∀ ParentOf.Woman ) ABox ◮ MotherWithDaughters(Mary); ParentOf(Mary,Sam) Implicit consequence: Instance checking ◮ Woman(Sam) 18 / 43

A partial description of the DL ALC TBox ◮ Person ⊑ Animal ⊓ Biped; Person ≡ Man ⊔ Woman ◮ Mother ≡ Woman ⊓ ( ∃ ParentOf.Person ) ◮ MotherWithDaughters ≡ Mother ⊓ ( ∀ ParentOf.Woman ) ABox ◮ MotherWithDaughters(Mary); ParentOf(Mary,Sam) Implicit consequence: Subsumption checking ◮ Man ⊑ Person 19 / 43

A partial description of the DL ALC TBox ◮ Human ⊑ ∃ hasHeart. ∃ hasPosition.Left ◮ Human ⊓ SitusInversus ⊑ ∃ hasHeart. ∃ hasPosition.Right ◮ ∃ hasHeart. ∃ hasPosition.Left ⊓ ∃ hasHeart. ∃ hasPosition.Right ⊑ ⊥ Implicit consequences: Concept unsatisfiability ◮ Human ⊓ SitusInversus ⊑ ⊥ ◮ Add Human(Sam) and SitusInversus(Sam): Inconsistent 20 / 43

Ontology Engineering 21 / 43

Ontology Engineering A formal way to conceptualise an application domain ◮ Concepts of interest in an application domain ◮ Properties describing concept features and attributes ◮ Perhaps some restrictions on the attributes and features Ontology Engineering ◮ The construction and development of ontologies 22 / 43

A visual representation of an ontology 23 / 43

Prot´ eg´ e: An Ontology Editing Tool 24 / 43

A reasoner at the backend of Prot´ eg´ e 25 / 43

Ontology Engineering ◮ Tools such as Prot´ eg´ e and the reasoners underlying it are possible because of the Web Ontology Language OWL 2 ◮ OWL 2 is the official ontology language endorsed by the World Wide Web Consortium (W3C) ◮ OWL 2 is based on the description logic known as SROIQ 26 / 43

OWL 2 OWL 2 Profiles OWL 2 DL OWL 2 QL OWL 2 EL RDFS OWL 2 RL Picture courtesy of Roman Kontchakov 27 / 43

OWL 2 QL and Ontology Based Data Access q ontology-based data integration ontology conceptual layer sources sources data layer sources Picture courtesy of Riccardo Rosati 28 / 43

OWL 2 QL and Ontology Based Data Access Reasoning Schema / Ontology Logical Query Result Schema Data Source Picture courtesy of Riccardo Rosati 29 / 43

OWL 2 QL and Ontology Based Data Access Reasoning Schema / Query Result Ontology Logical Schema Data Source Picture courtesy of Riccardo Rosati 30 / 43

OWL 2 QL and Ontology Based Data Access Reasoning Reasoning Schema / Query Result Ontology Logical Rewritten Schema Query Data Source Picture courtesy of Riccardo Rosati 31 / 43

Artificial Intelligence 32 / 43

AI vs Consciousness 33 / 43

Doomsday AI 34 / 43

Deep Learning 35 / 43

The Success of Deep Learning ◮ Sophisticated techniques for function fitting coupled with Big Data ◮ Successful identification of a class of practical applications ◮ We’ve shifted the goal posts a bit 36 / 43

The Real Surprise Underlying Deep Learning Some tasks, typically associated with perception or cognition can be approximated to a reasonable extent by fitting functions to data. 37 / 43

Deep Learning: Successful but Brittle 38 / 43

AI and the Game of Go 39 / 43

AI and Poker 40 / 43

Today, machines lack contextual reasoning capabilities, and their training must cover every eventuality, which is not only costly, but ultimately impossible. We want to explore how machines can acquire human-like communication and reasoning capabilities, with the ability to recognize new situations and environments and adapt to them. 41 / 43

Conclusion ◮ Deep Learning is function-fitting ◮ KRR is model-fitting ◮ The question is not whether to use Deep Learning, or KRR, or some other AI technique, but how to combine these techniques ◮ The best way to do so is for AI researchers to be exposed to all these techniques 42 / 43

Acknowledgement 43 / 43