[PPT] - I ndexi ng strategi es for goal speci fi c retri eval of PowerPoint Presentation

SLIDE 1

I ndexi ng strategi es for goal speci fi c retri eval

cases M i chael J. Pazzani D epar tm ent

I nfor m ati

and Com puter Sci ence U ni versi ty

Cal i forni a, I rvi ne, CA 92717 pazzani @ i cs. uci . edu O ne

the m ost curi

features

the hi story

econom i c sancti

has been the the extent to w hi ch the exper i ence

past cases has been

i gnored-

n Renw i ck

SLIDE 2

O utl i ne I . Probl em : retri eval

i ndi vi dual cases for m ul ti pl e purposes A . Expl anati

cti

I I . W hy i ndexi ng i s the ri ght w ay to thi nk about retri eval A . Psychol

cal

ve m em ory

ng speci fi ci ty & l evel s

processi ng

Retri evi ng rel evant rather than si m i l ar cases I I I . I ndexi ng strategi es for goal speci fi c retri eval

cases A . Retri eval

i ndi vi dual and general i zed cases

pr

ng done at stor age ti m e

retri eval cue

retri eval

ndi ces

(rather than abstract) features

c goal (expl ai n vs. predi ct)

e to be expl ai ned

pr edi cted

SLIDE 3

Question: What would happen if the United States refused to sell computers to South Korea unless South Korea stopped exporting automobiles to Canada?

The goal of the United States that South Korea not sell automobiles to Canada will fail and South Korea will purchase computers from a country which exports

States did not sell grain to the Soviet Union after the Soviet Union invaded

Soviet Union. Also, Australia did not sell uranium to France after France exploded nuclear weapons in the South Pacific. South Africa sold uranium to France.

SLIDE 4

Question: What would the United States do if Turkey invaded Cyprus?

The United States would refuse to sell a commodity to Turkey if Turkey invaded

sell food to Greece. Also, when the Soviet Union invaded Afghanistan, the United States refused to sell grain to the Soviet Union.

SLIDE 5

W hy i ndexi ng i s the ri ght w ay to thi nk about retri eval : Reconstructi ve m em ory (Bartl ett, 1932; Schank 1980; Kol

1984)

case i s

zed i n m em ory by the know l edge structures that gui ded the com pr ehensi

case. Corol l ary: A case w i l l be i ndexed i n m ul ti pl e pl aces i f m ore than

know l edge structure i s accessed duri ng com prehensi

y those aspects

a case w hi ch di ffer from those

the know l edge structure used for com prehensi

are stored. D i fferi ng features becom e i ndi ces that al l

retri eval

the case w hen that feature i s present i n the retri eval cue (or i nferred from i nform ati

i n the retri eval cue)

m ore cases are added to m em ory, addi ti

know l edge structures are created to

ze new experi ences.

SLIDE 6

W hy i ndexi ng i s the ri ght w ay to thi nk about retri eval : Experi m ental Evi dence

c m em ory (Freedm an & Loftus, 1971) Categor y (e. g. vegetabl e) & attr i bute (e. g. , gr een) Subj ects presented w i th category before attri bute retri eve exem pl ar m ore rapi dl y than attri bute before category.

sodi c M em ory (Rei ser, Bl ack & A bel son, 1987) A cti vi ty (e. g. Eati ng at a Restaurant) & acti

(di dn’ t get w hat you w ant) Subj ects presented w i th acti vi ty before acti

retri eve exem pl ar m ore rapi dl y than acti

before acti vi ty.

SLIDE 7

The effect

processi ng task

s

processi ng (Crai k & Lockhart, 1972) Retri evabi l i ty affected by depth

processi ng duri ng storage Jacoby & D al l as (1981): Subj ects study w ord l i st: Subj ects answ er questi

1. A ppearance: I s i t i n capi tal l etters 2. Sound: D oes i t rhym e w i th “teach” 3. M eani ng: I s i t a form

com m uni cati

I n test phase: subj ects veri fy that they studi ed w ord. Success Rate: M eani ng 50% < Sound 63% < M eani ng 86%

ng Speci fi ci ty (Tul vi ng, 1983) Retri eval affected by si m i l ari ty betw een storage and retri eval context Recogni ti

fai l ure

recal l abl e w ords. Sei fert (1988) found pri m i ng

story recal l rel ated to processi ng goal at storage and retri eval

SLIDE 8

W hy i ndexi ng i s the ri ght w ay to i m pl em ent CBR: Retri eve the m ost rel evant rather than the m ost si m i l ar case

i ndex i s a feature

a case i denti fi ed at storage ti m e that descri bes the si tuati

i n w hi ch the case shoul d be retri eved.

s by W al tz and Thagard & H ol yoak cannot account for si tuati

i n w hi ch there i s a uni que i tem speci fi ed by the retri eval cue i n m em ory, but i t i s not retri eved. “Can you thi nk

an i nci dent i n w hi ch the US threatened a thi rd-w orl d country and a U S adversary hel ped

“Can you thi nk

an i nci dent i n w hi ch an Engl i sh speaki ng country threatened an A fri can country and a country that exports pal l adi um hel ped

by W al tz assum es that the m em ory processes cannot di sti ngui sh m ost si m i l ar case (accordi ng to the surface features) from the m ost rel evant.

SLIDE 9

Experi m ental Com pari son

nearest nei ghbor vs. EBL I nput: D escri pti

10

15 econom i c sancti

i nci dents. Test: Predi ct success/fai l ure

fi ve hypotheti cal i nci dents. 1961 USSR vs. Albania 1976 US vs. Ethiopia threat: Refuse to sell grain threat: Stop aid (57 million) demand: Stop economic ties with China demand: Stop human rights violations

Canadian wheat at lower price 1980 US vs. USSR 1983 Australia vs. France threat: Cut off grain sales threat: Not sell uranium demand: Withdraw troops from demand: Stop nuclear tests Afghanistan

South Africa at higher price Argentina at a higher price

N um ber

Exam pl es A ccuracy

SLIDE 10

W hat m akes a good i ndex? I .Rel evant surface features A . H undred step rul e- D eep f eatur es (e. g. , m or al s) cannot be com puted i n ti m e

s and effects are the

the retri eval process, not i ts i nputs”

evant surface features are those surface features that al l

the sol uti

to a probl em to be i nferred. I I . A feature rel evant for

task m ay not be rel evant for a di fferent task. Processi ng task at storage ti m e becom es part

i ndex A . Expl anatory I ndi ces “What could cause the price of oil to rise?”

cti ve I ndi ces “What would happen if the US refused to sell com- puters to South Korea if South Korea continues to export automobiles to Canada?”

SLIDE 11

W hat m akes a good i ndex? I I I . D i fferent features

the sam e case can be expl ai ned

pr edi cted. The featur e that needs to be expl ai ned

pr edi cted becom es par t

the i ndex. A . Predi cti ng the resul t

an acti

What would happen if the United States refused to sell computers to South Korea unless South Korea stopped exporting automobiles to Canada?

cti ng w hat acti

w i l l

What would the United States do if Turkey invaded Cyprus? Sum m ary: A n i ndex consi sts

a tri pl e

e and val ue

a case (or gener al i zed case)

pti

task :explain or :predict

to be expl ai ned

predi cted.

SLIDE 12

D eri vi ng I ndi ces I .Feature and val ue (W hat features are predi cti ve) A . A nal yti cal l y: EBL

ri cal l y: SBL I I . Task A . A nal yti cal l y: antecedent/consequent

dom ai n know l edge

ri cal l y: predi cti veness vs predi ctabi l i ty I I I . Featur e to be pr edi cted (O r expl ai ned) D eterm i ned by processi ng goal

SLIDE 13

D eterm i ni ng I ndi ces for US-USSR grai n em bargo: Processi ng goal : Expl ai n resul t (USSR buys grai n from A rgenti na) 1. Threat

I ncreased dem and 2. I ncreased dem and

W i l l i ngness to pay hi gher pri ce 3. Purchase

Possess

SLIDE 14

I ndexi ng i n m em ory by rel evant features

SLIDE 15

Resul ts after 15 cases

SLIDE 16

Concl usi

ndexi ng expl ai ns desi rabl e aspects

hum an m em ory retri eval

eval

rel evant exem pl ars by surface features

ng speci fi ci ty (goal speci fi c retri eval )

ti ci sm by W al tz i ndexi ng caused by m i sunderstandi ng:

“occasi

, breakthrough i nsi ght” that requi res m ore than 100 steps can resul t i n a structure that i s retri eved i n l ess than a 100 steps.

ndexi ng by those surface features w hose presence i m pl i es a deep feature constructs effi ci ent recogni zers.

erarchi cal m em ory search i s not i nherentl y seri al

vel y paral l el m em ory searches (even w i thout i ndexi ng) can m ake use