Panel on Knowledge Repositories Organizer Chitta Baral chittaasu

Panel on Knowledge Repositories Organizer: Chitta Baral chitta@asu. edu Panel members: Michael Gelfond Vladimir Lifschitz

What do I mean by a knowledge repository? • A collection of knowledge modules. – That can be used by knowledge base developers. • Similar to Java or C++ libraries. • But building a knowledge repository – poses a lot more challenges; • to be discussed in later slides. – and will be much larger in size.

Why do we need knowledge repositories? • Reasoning with Knowledge and learning knowledge is the essence of AI. – Evident from the meaning of “intelligence” in a dictionary • Lot of progress in Knowledge Representation. • Especially with respect to Ans. Prolog (logic programming with answer set semantics) – – A core language with many suggested extensions A large body of theoretical results Many implementations Many applications

But we need to … • Go beyond – Writing knowledge axioms from scratch – Small knowledge bases • Be able to build large knowledge bases without starting from scratch. • Make it easier to build knowledge bases. • Reuse knowledge modules developed by others. • Make knowledge bases part of most AI systems.

Applications and Impacts of Knowledge Repositories • Question answering systems – Text: John took a flight from Rome to Paris 6 hours ago? – Question: Where is John now? Where is his wife who saw him off at the airport? • Any system that needs to use common-sense reasoning. • Any system that needs to reason with knowledge in one or many domains.

Is this a blue sky dream? • Not really? • Wordnet (http: //wordnet. princeton. edu/) – An electronic repository of words and their meanings has been very useful. • It took a lot of work to get built. • A knowledge repository will need a lot more work.

What does building a Knowledge Repository involve? • A large body of Knowledge modules possibly grouped in packages – Common sense modules – Domain specific modules – High level modules: actions, time, space, etc. • Methodology to facilitate building modules – Inheritance, encapsulation, modeling languages, etc. • Interface mechanisms similar (in functionality) to interface mechanisms in Java, C++ etc.

Existing efforts: CYC • CYC: a pioneer • Possible IP and legal issues. – subsets (Research. CYC) need signing of a lot of legal documents. • CYC’ s language is proprietary and untested outside of CYC. (mostly unpublished outside). • But if this can be overcome, then it could be a good starting or reference point.

Existing efforts: CYC • CYC : our effort : : Celera effort : Open Genomics effort • We would like the whole community to be involved in building. • Openmind: collects NL knowledge over the web.

Existing effort: SUMO and MILO • http: //www. ontologyportal. org/ • SUMO and MILO are freely available • SUMO – http: //cvs. sourceforge. net/viewcvs. py/*checkout*/sigm akee/KBs/Merge. kif? rev=1. 3 • MILO – http: //cvs. sourceforge. net/viewcvs. py/*checkout*/sigm akee/KBs/Mid-level-ontology. kif? rev=1. 2

SUMO • SUMO (Suggested Upper Merged Ontology) • Based on first-order logic. • It incorporates – – – elements of John Sowa's upper ontology Russell and Norvig's ontology PSL (Process Specification Language), Casati and Varzi's theory of holes, Allen's temporal axioms, etc. • It has a nice browsing and editing tools, and • Inference and Ontology management system – http: //sigmakee. sourceforge. net/

MILO (MId-Level Ontology) • Aim is to be a bridge between the abstract content of the SUMO and the rich detail of the various domain ontologies. • In progress, incomplete. • Contains a Description Logic Knowledge base – Class-subclass – Class-instances – Relations

Going beyond SUMO and MILO? • Why? – Both SUMO and MILO are based on firstorder logic. – Need ways to express defaults and exceptions, – need ways to express problem solving queries, such as planning, diagnosis, etc. –…

Recall: What do we need? • A large body of Knowledge modules possibly grouped in packages – Common sense modules – Domain specific modules – High level modules: actions, time, space, etc. • Methodology to facilitate building modules – Inheritance, encapsulation, modeling languages, etc. • Interface mechanisms similar (in functionality) to interface mechanisms in Java, C++ etc.

Coupling modules and inference mechanism • Ans. Prolog versus ASP – Ans. Prolog -- Programming in logic with answer sets – ASP – seems to be focused on the generate and test problem solving • Need modules of various kinds – Is ancestor(john, mary)? – Find a plan – Find a schedule (Prolog style) (ASP style) (CLP) • Different kinds of modules may need different inference mechanisms

Next Steps, challenges • Lets look at the AAAI’ 06 Spring Symposium CFP.

AAAI’ 06 Spring symposium • Title: Formalizing and Compiling Background Knowledge and its applications to Knowledge Representation and Question Answering. • Organizing Committee: – – – – Chitta Baral (chitta@asu. edu) Alfredo Gabaldon (alfredo. gabaldon@nicta. com. au) Michael Gelfond (mgelfond@cs. ttu. edu) Joohyung Lee (appsmurf@cs. utexas. edu) Vladimir Lifschitz (vl@cs. utexas. edu) Steve Maiorano (stevemai@mac. com) Sheila Mc. Ilraith (sheila@cs. toronto. edu) Leora Morgenstern (leora@steam. stanford. edu)

CFP: Requests contributions that are • A: formalizations (knowledge modules) of background knowledge in specific domains as well as, • B: papers addressing general challenges such as formalizing background knowledge for use by multiple users on multiple reasoning tasks. – Interface issues, reuse, etc.

A: Knowledge module papers • No restriction on the domain to be formalized or on the level of specificity • Suggested common format – A knowledge base (KB) written in English. – Examples of informal consequences of KB, preferably accompanied by some explanations, including defaults and other commonsense knowledge not directly mentioned in KB but needed to produce the desired consequence.

A: Knowledge module papers (cont. ) – Information about which logic/language is used in formalizing it. (Syntax, semantics, and where the reasoning system is available. ) – The formalization – Short description on how the formalization can be tested using the reasoning system.

Existing knowledge encoded in Ans. Prolog • Small Ans. Prolog programs (not quite modules – don’t have modular interface) – Knowledge Representation, Reasoning and Declarative Problem Solving. Baral – Various surveys: Niemela et al. ; Gelfond and Leone. Etc. • Larger programs – RCS-USA Advisor (http: //www. krlab. cs. ttu. edu///Software/) – www. baral. us/bookone/ – Vladimir is collecting a list of ASP applications.

Further ideas for submissions of type A. • At various abstractions – Actions, time, space, etc. • Various domains – Travel, terrorism, etc. • Further collections and catalogues of existing encoded knowledge.

B: Interface and Engineering issues • How to call a module from another module: interface syntax and semantics • Object oriented issues – Encapsulation – Classes, sub-classes, Inheritance – Polymorphism • Modeling language

Some initial steps on Interface issues – Towards an Integration of Answer Set and Constraint Solving. Baselice, Bonatti and Gelfond. ICLP’ 05 – A language for modular ASP. Tari, Baral, & Anwar. ASP’ 05. – Enhancing ASP with templates. Ianni et al. NMR’ 2004. – Personal communication. Lifschitz. – F-logic papers. Kifer et al.

Challenges vis-à-vis C++ and Java libraries • Number of modules could be much larger and much varied than classes and methods in Java libraries • Multiple Ans. Prolog sub-languages, each with a different reasoning mechanism • Various sources of knowledge – some would be learned • Initially a smaller number of developers • Language is still evolving (core is there)

More info on the symposium • Symposium Dates: – March 27 -29 2006. • AAAI site: – http: //www. aaai. org/Symposia/Spring/2006/sssparticipation 06. pdf • Symposium cite – http: //www. public. asu. edu/~cbaral/aaai 06 -ss/ • Deadlines: – – Submission: October 7, 2005 (extended to October 21 st) Response: November 4, 2005 Camera ready due at AAAI: January 27, 2006 Symposium date: March 27 -29 2006.

Thank You