the design of j-drew: a deductive reasoning engine for the web bruce spencer national research...

Post on 25-Dec-2015

217 Views

Category:

Documents

4 Downloads

Preview:

Click to see full reader

TRANSCRIPT

The design of j-DREW: a deductive reasoning engine

for the web

Bruce SpencerNational Research Council Canada

and University of New Brunswick

Fredericton, Canada

20-Sep-0220-Sep-02 2

NRC – A National Institution

NRC Institute / Innovation Centre

IRAP Office

Virtual Innovation Centres

20-Sep-0220-Sep-02 4

Motivation• Started as a bet: Can I teach deduction and advanced

internet systems in one course?– Students understand recursive search– basic propositional top-down prover

• Teaching a course at UNB on building Internet applications that employ deduction(in conjunction with Harold Boley, formerly DFKI)– Semantic web ontologies have rules– XML is declarative, so deduction is a basic operation

• Students have trouble building systems (variables)– unification, composition of substitutions

unbinding on backtracking• Specialized Prolog implementation courses focus on vars

– Warren and Meier: Computing with Logic (1988)• Can we hide the hard parts?

20-Sep-0220-Sep-02 5

Requirements on deduction system used in the course (and on web?)• First order logic with function symbols• Easily configured

– Forward or backward search or backward and iterative deepening

– Search strategy to be modified by student• Embeddable

– Supports calls to and from rest of system– Easy to add new primitives (iterators)

• Written in 100% Java– Access to libraries for networking,

cryptography, XML parsers, RDF, XSLT, …

20-Sep-0220-Sep-02 6

Propositional Proverinitially proofTree has an open Goalloop if(proofTree.hasNoOpenGoal()) halt('success'); else Goal g = proofTree.selectOpenGoal(); g.createMatchingClauseList(); if(g.hasMoreMatchingClauses()) DefiniteClause c = g.nextClause();

g.attachClause(c); choicePoints.push(g); else chronologicalBacktrack();

Java Iterator

chronologicalBacktrack while(choicePoints.nonEmpty()) Goal g = choicePoints.pop(); g.removeAttachedClause(); if(g.hasMoreClauses()) return; halt('failure')

Create bindings

Remove bindings

20-Sep-0220-Sep-02 7

Tree with initial goal p(Z)

p(X) :- q(X), r(X).q(g(X)).q(k).r(g(h))

p(Y1) :- q(Y1), r(Y1)

?:-p(Z)

a

a ChoicePoints

b

Z/Y1

q(g(Y2) r(g(h)b c

c

Y1/f(Y2)

Y2/hq(g(Y2)

q(k) Y1/k

b’

20-Sep-0220-Sep-02 8

Hiding the hard parts: Top-Down

• A Goal creates its list of matching clauses and maintains pointer to current selection– Java iterator

• Attaching a clause head to a Goal– imposes variable bindings across tree

• Removing the clause – relaxes variable bindings

• The tree provides access to all open Goals

20-Sep-0220-Sep-02 9

Deployed and Planned Prototypes

• Basic Prolog Engine– Negation as failure– RuleML input

• Iterative deepening search– Search complete, unlike Prolog– Dynamic additions to clause set

• allowed between depth limits (synchronous)

• Possible student exercises – delay selecting deferred goals– Negation of ground goals– Backtracking to deeper choicepoint

• Bottomup

} deployed

} planned

20-Sep-0220-Sep-02 10

Deduction Example: Semantic Web

• Semantics = meaning• Ontology

– hierarchy of terms (subset)– logic conditions on their usage

• Terms in a web page can be linked to their meanings in some ontology– Search with greater accuracy

20-Sep-0220-Sep-02 11

20-Sep-0220-Sep-02 12

•SmokedSalmon is the intersection of Smoked and Salmon

Smoked Salmon

20-Sep-0220-Sep-02 13

•Gravalax is the intersection of Cured and Salmon, but not Smoked

•SmokedSalmon is the intersection of Smoked and Salmon

Smoked Salmon

Gravalax

20-Sep-0220-Sep-02 14

•Lox is Smoked, Cured Salmon

•Gravalax is the intersection of Cured and Salmon, but not Smoked

•SmokedSalmon is the intersection of Smoked and Salmon

Smoked Salmon

Gravalax

Lox

20-Sep-0220-Sep-02 15

• A search for keywords

Salmon and Cured should return pages that mention Gravalax, even if they don’t mention Salmon and Cured

• A search for Salmon and Smoked will return pages with smoked salmon, should also return pages with Lox, but not Gravalax

Smoked Salmon

LoxGravalax

The Semantic Web vision is to make information on the web “understood” by computers, for searching,categorizing, …

20-Sep-0220-Sep-02 16

Smoked Salmon

LoxGravalax

20-Sep-0220-Sep-02 17

One possible encoding

Search criteria:retrieve(P) :-

mentions(P, cured),mentions(P, salmon).

Ontology:mentions(P, cured) :- mentions(P, gravalax).mentions(P, salmon) :- mentions(P, gravalax).

A search for keywords Salmon and Cured should return pages that mention Gravalax, even if they don’t mention Salmon and Cured.

mentions(p1 gravalax).

retrieve(p1) succeeds

20-Sep-0220-Sep-02 18

retrieve(P) :- mentions(P, smoked),mentions(P, salmon).

mentions(P, cured) :- mentions(P, lox).mentions(P, salmon) :- mentions(P, lox).mentions(P, smoked) :- mentions(P, lox).mentions(P, cured) :- mentions(P, gravalax).mentions(P, salmon) :- mentions(P, gravalax).

A search for Salmon and Smoked will return pages with smoked salmon, should also return pages with Lox, but not Gravalax.

mentions(p1 gravalax).mentions(p2, lox).

retrieve(p1) failsretrieve(p2) succeeds

20-Sep-0220-Sep-02 19

Bottom-Up / Forward Chaining• Set of support prover for definite clauses• Facts are supports• Theorem: Completeness preserved when definite

clause resolutions are only between first negative literal and fact.– Proof: completeness of lock resolution

(Boyer’s PhD)• Use standard search procedure to reduce

redundant checking (next)• Unlike OPS/Rete, returns proofs and uses first

order syntax for atoms

20-Sep-0220-Sep-02 20

Theorem Prover’s Search Procedure

• 3 Definite Clause Lists:– new facts

(priority queue)– old facts– rules

• 2 Discrimination trees: – used facts – rules, indexed on

first goal

loop select new fact for each matching rule resolve process new result add to used facts

process new result(C) if C is rule for each old fact matching first resolve process new result add C to rules else add C to new facts

20-Sep-0220-Sep-02 21

Summary• Teaching building deduction systems not hard

– use of abstractions to hide variable handling• Semantic Web needs deduction

– access to proof data structure (trust)• j-DREW deployed systems can use

– forward / backward / other searches– small footprint, 100% Java

• Take home messages:– not clear what form Semantic Web systems

will take, so flexibility is key– implementers should provide API’s

20-Sep-0220-Sep-02 22

References

• CS6999 Semantic Web Techniques http://www.cs.unb.ca/~bspencer/CS6999

• RuleML http://www.dfki.de/ruleml

• j-DREW open source?– Not yet

top related