questions traditional view of language languageski.clps.brown.edu/cogsim/cogsim.12lang.prt.pdf ·...
TRANSCRIPT
1Lan
guag
e
How
canwepossib
lysim
ulate
languag
eab
ilitiesin
neu
ralterm
s?...
•(W
ecan
’t...Butman
yasp
ectsoflan
guag
eare
not“sp
ecial”):
•Justan
other
setofinput/outputpath
s.
•Lev
els:phonem
es/letters,
words,p
hrases,sen
tences,
parag
raphs,an
dbey
ond..
•Hugecombinato
rialpower:
distrib
uted
repsover
time!
2Biological
Substrates
ofLan
guag
e
Broca’s
Wernicke’s
FrontalP
arietal
Occipital
Temporal
-Broca’s
=sp
eechoutput,sy
ntax
,gram
mar
(surface
productio
n):
activemain
tenan
ceofcontex
tto
perfo
rmsyntactic
processin
g
-Wern
icke’s
=sem
antic
compreh
ensio
n+output(deep
):
interco
nnected
overlap
pingdistrib
uted
info
aboutsem
antics
3Trad
itional
view
oflan
guag
e
•Lan
guag
ecompeten
cedefi
ned
byknowled
geofrules
andexceptions
(eg.ibefo
reeexcep
tafter
c)
•Knowled
geab
outwordsissto
redin
acen
tralmen
tallexicon(dictio
nary
)
•Each
word
has
alex
icalrep
resentatio
nthat
islin
ked
to
inform
ationab
outits
orth
ograp
hy,phonology,sem
antics
4Neu
ralnet
/Connectio
nist
View
ofLan
guag
e
•Lan
guag
eisan
other
setofinput-o
utputmap
pings(eg
orth
ograp
hyto
phonology,o
rthograp
hyto
seman
tics)
•These
map
pingsare
trained
upusin
gthesam
elearn
ing
algorith
msused
elsewhere
(e.g.,v
ision)
•Thesam
epath
way
shan
dle
both
rules
andexcep
tions
•Hard
totell
what
is“reg
ular”
vs“ex
ceptio
nal”
–reg
ular:
clown,d
own
–excep
tion:blown..
butblowngoes
with
grown
•Distrib
uted
lexico
n:Knowled
geab
outwordsisem
bodied
in
reciprocal
map
pingsbetw
eenphonology,orth
ograp
hy,
seman
tics–
thereis
nocentral“w
ordrepresentation”
5Questio
ns
•What
gen
eralprocesses
areinvolved
inread
ing,an
dhow
do
these
sometim
esfail
(e.g.,in
dyslex
ia)?
•How
areweab
leto
read“cat”,“y
acht”,an
d“n
ust”?
•Whydokidssay
“Igoed
tosch
ool”
afterfirst
saying“I
wen
t”?
•How
dowordscometo
mean
anything?
•How
dowegobey
ondwordsto
senten
ces?
6Distrib
uted
Lexico
nModel
Orthography
Phonology
Sem
antics
Hidden
Hidden
Hidden
7Distrib
uted
Lexico
nModel
&Read
ing
Orthography
Phonology
Sem
antics
Hidden
Hidden
Hidden
TwoRoutes
•Direct
route:
orth
ograp
hyto
phonology
•Indirect
route:
orth
ograp
hy→
seman
tics→
phonology
8Sim
ulatin
gDifferen
tKindsofDyslex
ia
Orthography
Phonology
Sem
antics
Hidden
Hidden
Hidden
Phonological:
nonwords(“n
ust”)
impaired
.
Deep:
phono+sem
antic
errors
(“dog”as
“cat”)+
visu
alerro
rs(“d
og”as
“dot”)
+more
errors
with
“truth”(ab
stract)than
“chair”
(concrete)
Surface:
nonwordsOK+sem
antic
accessim
paired
+diffi
culty
readingexcep
tionwords(“y
acht”)
+visu
alerro
rs.
9TheModel
tw
y
ad
eg
kn
rs
il
no
pr
sv
lo
rt
ac
eg
pr
st
wa
ei
cd
ef
gh
ln
OS
_Hid
Sem
anticsSP
_Hid
w
st
pr
ln
hk
fg
−d
tw
pr
ln
hk
fg
−d
tw
pr
ln
hk
fg
−d
io
ae
OU
EI
@A
z
tv
rs
np
kl
gj
−d
z
tv
rs
np
kl
gj
−d
z
tv
rs
np
kl
gj
−d
Orthography
OP
_Hid
Phonology
Train
edonall
path
way
s(orth
o⇔
phonoetc)
for40
4-lettermonosyllab
icwords(eg
flag
,star)
Concrete
&ab
stractwordsuse
differen
tpoolsofsem
antic
units
Abstract
wordsactiv
atefew
ersem
antic
units
than
concrete
words
10Corpusan
dSem
antics
Concrete/A
bstract Semantics
Y
X0.00
5.00
10.00
15.00
20.00
25.00
30.00
35.00
40.00
0.0010.00
20.0030.00
40.00
tentflea
lassfacegrin w
avereedloon hare deer
hindtartflan
plea hint plan flaw lackneed
deedtactease past gain
lossfactrole
stay hire loan rent costw
ageropepostlockcase
starcoatflag
seman
ticrep
smad
eupofdistrib
uted
features
(e.g.concrete
=liv
ing;ab
stract=has-d
uratio
n)
11[dyslex
.proj]
12Sem
antic
Path
way
Lesio
ns,In
tactDirect
0.00.2
0.40.6
0.8L
esion Proportion
0 5 10 15 20
Errors
Visual
Vis + Sem
Semantic
Blend
Other
0 5 10 15 20
Errors
0.00.2
0.40.6
0.8L
esion Proportion
Concrete
Abstract
OS_H
id
SP_H
idSP
_Hid
OS_H
id
-visu
alerro
rswith
seman
ticpath
way
lesions;nosem
antic
errs!
-(m
ore
forconcrete
than
abstract..)
13Sem
antic
Path
way
Lesio
ns,In
tactDirect
Surface
Dyslex
ia
0.00.2
0.40.6
0.8L
esion Proportion
0 5 10 15 20
Errors
Visual
Vis + Sem
Semantic
Blend
Other
0 5 10 15 20
Errors
0.00.2
0.40.6
0.8L
esion Proportion
Concrete
Abstract
OS_H
id
SP_H
idSP
_Hid
OS_H
id
-visu
alerro
rswith
seman
ticpath
way
lesions;nosem
antic
errs!
-(m
ore
forconcrete
than
abstract..)
14Sem
antic
Path
way
Lesio
ns,L
esioned
Direct
0.00.2
0.40.6
0.8L
esion Proportion
0 5 10 15 20
Errors
Visual
Vis + Sem
Semantic
Blend
Other
0 5 10 15 20
Errors
0.00.2
0.40.6
0.8L
esion Proportion
Concrete
Abstract
OS_H
id
SP_H
idSP
_Hid
OS_H
id
-multip
leerro
rstypes
-more
abstract
seman
ticerrs
than
concrete
15Sem
antic
Path
way
Lesio
ns,L
esioned
Direct
Deep
Dyslex
ia
0.00.2
0.40.6
0.8L
esion Proportion
0 5 10 15 20
Errors
Visual
Vis + Sem
Semantic
Blend
Other
0 5 10 15 20
Errors
0.00.2
0.40.6
0.8L
esion Proportion
Concrete
Abstract
OS_H
id
SP_H
idSP
_Hid
OS_H
id
-multip
leerro
rstypes
-more
abstract
seman
ticerrs
than
concrete
16Direct
Path
way
Lesio
n
0.00.2
0.40.6
0.8L
esion Proportion
0 5 10 15 20
Errors
Visual
Vis + Sem
Semantic
Blend
Other
0 5 10 15 20
Errors
0.00.2
0.40.6
0.8L
esion Proportion
Concrete
Abstract
Full Sem
Full Sem
No Sem
No Sem
-minordirect
dam
age:
justviserro
rs-more
dam
age:
seman
ticerrs
⇒deep
dyslex
iaev
enwith
FullSem
1718
1920
Abstract
vsConcrete:
Summary
•Sem
antic
path
way
lesionshurtconcrete
wordsmore
than
abstract
words
•Concrete
wordsare
more
strongly
represen
ted(m
ore
units
active)
than
abstract
wordsin
thesem
antic
path
way
•Learn
ingisafunctio
nofactiv
ation,so
thesem
antic
path
way
learnsmore
aboutconcrete
words
•Themore
seman
ticpath
way
learnsab
outconcrete
wordsthe
lessdirect
path
way
learns
•Theless
thedirect
path
way
learns,th
eless
itisab
leto
support
perfo
rman
ceonits
own
•With
fulldirect
path
way
lesions,themodel
mak
esmore
seman
ticerro
rsforab
stractthan
concrete
•Abstract
wordshav
eless
distin
ctivesem
antic
repsthan
concrete
words
•Themodel
ismore
likely
tofall
into
wrongsem
antic
attractor
forab
stractwords
21Read
ing:Distrib
uted
Lexico
nModel
Orthography
Phonology
Sem
antics
Hidden
Hidden
Hidden
•Distrib
uted
reps(notlocalized
toonereg
ion).
•Interactiv
e(notmodules),lead
sto
interestin
gdivisio
nsof
labor.
22Questio
ns
•What
gen
eralprocesses
areinvolved
inread
ing,an
dhow
do
these
sometim
esfail
(e.g.,in
dyslex
ia)?
Distrib
uted
lexico
n(orth
o,p
hono,sem
)
•How
areweab
leto
read“cat”,“y
acht”,an
d“n
ust”?
•Whydokidssay
“Igoed
tosch
ool”
afterfirst
saying“I
wen
t”?
•How
dowordscometo
mean
anything?
•How
dowegobey
ondwordsto
senten
ces?
23Distrib
uted
Lexico
nModel
Orthography
Phonology
Sem
antics
Hidden
Hidden
Hidden
24Reg
ularities
&Excep
tions:A
Contin
uum
Reg
ularities
inpronunciatio
nare
often
partial,co
ntex
tdep
enden
t:
bint
iin
mint,h
int,stin
t,...(reg
ular)
vspint(ex
ceptio
n)
butalso
:mind,fi
nd,h
ind,...
(regular)
mine,fi
ne,d
ine,...
(regular)
Pronunciatio
ndep
endsoncontex
t.
Excep
tionsare
extrem
eofcontex
tdep
enden
t.
Need
aran
geofcontex
tdep
enden
cyforreg
ulars
andexcep
tions.
25
2627
28Phonological
Rep
resentatio
ns
•Sam
e7slo
tvowel-cen
teredrep
resentatio
nsas
befo
re:
–face
=fffA
sss
–grin
=grrin
nn
–star
=sttarrr
–post
=pppOstt
•Excep
tinstead
ofusin
galocalist
repofeach
phonem
e,weuse
adistrib
uted
rep
•Thisallo
wsusto
represen
tthefact
that
phonem
esvary
intheir
similarity
toonean
other...
29
30Read
ingModel
Ortho
Ortho_C
ode
Hidden
Phon
detailed
model
ofthe“d
irect”read
ingpath
way
(orth
o→
phono)
-train
edto
pronounce
largeset
ofreg
ular
&excep
tionwords
-gen
eralizationtestin
g:nonwords(eg
,nust)
3132
Nonword
Perfo
rman
ce
Reg
ularity
tests(G
lush
ko):bint→
/bint/
Pseu
do-homophones
(McC
ann&
Besn
er):
phoyce
→/fYs/
,choyce
→/CYs/
Match
edreg
ularity
/excep
tioncases
(Tarab
an):
Highfreq
:poes
→/pOz/
,goes
→/gOz/
,does
→/dˆz/
Low
freq:mose
→/pOs/
,poes
→/pOz/
,lose
→/lU
z/
Nonword
Set
Model
PMSP
Peo
ple
Glush
koreg
ulars
95.397.7
93.8Glush
koexcep
tionsraw
79.072.1
78.3Glush
koexcep
tionsalt
OK
97.6100.0
95.9McC
ann&
Besn
erctrls
85.985.0
88.6McC
ann&
Besn
erhomoph
92.3n/a
94.3Tarab
an&
McC
lelland
97.9n/a
100.01
33Read
ingSummary
•Onenetw
ork
canlearn
both
regular
pronunciatio
nsan
d
excep
tions,an
ditcan
gen
eralizeproperly
tononwords
•Netw
ork
learnsagoodmix
ofcontex
t-dep
enden
tan
d
contex
t-invarian
trep
resentatio
nsonits
own
34Questio
ns
•What
gen
eralprocesses
areinvolved
inread
ing,an
dhow
do
these
sometim
esfail
(e.g.,in
dyslex
ia)?Distrib
uted
lexico
n
(orth
o,p
hono,sem
)
•How
areweab
leto
read“cat”,“y
acht”,an
d“n
ust”?
Ran
geof
contex
tdep
enden
trep
s&
contin
uum
ofreg
ularity
-excep
tion
•Whydokidssay
“Igoed
tosch
ool”
afterfirst
saying“I
wen
t”?
•How
dowordscometo
mean
anything?
•How
dowegobey
ondwordsto
senten
ces?
35Past
Ten
seSim
ple
model
Input:Someunits
represen
tword
iden
tity,others
inflectio
n
Output:Phonology
Sem
antics
Phonology
pasttense
ed(strong correlation)
36Learn
ingthePast
Ten
seofVerb
sRumelh
art&
McC
lelland(1986)
Child
renexhibitthree
stages
ofdev
elopmen
t:
•Stag
e1:
Small
number
ofverb
sin
past
tense
–Very
highfreq
uen
cy
–Majo
rityare
irregular
–Correct
perfo
rman
ce
–Exam
ples:
came,g
ot,g
ave,lo
oked
,need
ed,took,w
ent
•Stag
e2:
Larg
ernumber
ofverb
s
–Mostly
regular
–Exam
ples:
wiped
,pulled
–Can
gen
eratepast
tense
forinven
tedverb
s(rick
→rick
ed)
–Over-reg
ularize
wordsthat
were
correct
instag
e1(goed
)
•Stag
e3:
Reg
ular
andirreg
ular
form
scoexist
–Reg
ained
use
ofcorrect
irregular
form
s
37Past
Ten
se:U-Shap
edCurve
Thisistheinterestin
gtarg
etdev
elopmen
talphen
omen
on:
23
45
67
89
10Y
ears of Age
0.90
0.92
0.94
0.96
0.98
1.00
1−Overreg Rate
Overregularization in A
damEventualcorrect
performance
assumed
38U-Shap
edHisto
ry
Initially
:explain
edin
termsofsep
arate,overzealo
usrule
system
.
Then
:Rumelh
art&
McC
lelland,U-sh
aped
curvebased
onslo
w
netw
ork
processin
gofreg
ularities.
BUT:train
edirreg
ulars
first,th
enreg
s.
(much
contro
versy
ensu
es)
Later:
Plunkett
etal.,etc,m
anipulate
enviro
ingrad
edway
(contin
uously
addreg
verb
sto
trainingset
instead
ofall
atonce).
39U-Shap
edModel
inLeab
ra
Can
weget
aU-sh
aped
curvewith
outbuild
ingtheexplan
ation
into
theen
viro
nmen
t?
•Key
:all
prev
iousconnectio
nist
accounts
used
feedforw
ard,
back
propnets
–noattracto
rdynam
ics.
•Problem
:back
prop(pure
error-d
riven
learning)lead
sto
steady
decrease
inerro
r;hard
toexplain
increase
inerro
r...
•Interactiv
ity,competitio
n&
Heb
bian
learningproduce
netw
ork
that
isin
dynam
icbalan
cebetw
eenreg
&irreg
map
pings.
•Small
tweak
scan
shift
itoneway
ortheother
(prim
ingmodel).
Strin
gofreg
ular
trialswill
leadto
overreg
ularizatio
n...
40ThePast
Ten
seModel
Sem
antics
Hidden P
honology
41Phonological
Rep
resentatio
ns
•Sam
e7slo
wvowel-cen
teredrep
sas
befo
re:
–face
=fffA
sss
–grin
=grrin
nn
–star
=sttarrr
–post
=pppOstt
•Small
additio
n:8th
slotto
represen
textra
inflectio
n
–started
-sstarttD
–startin
g=sstarttG
42ThePast
Ten
seModel
Inflectio
nReg
sfxReg
ular/
Irregular
exam
ples
Base
–Iwalk
tothesto
redaily.
Igoto
thesto
redaily.
Past
-edIwalk
edto
thesto
reyesterd
ay.Iwen
tto
thesto
reyesterd
ay.3rd
pers
sing
-sShewalk
sto
thesto
redaily.
Shegoes
tothesto
redaily.
Progressiv
e-in
gIam
walk
ingto
thesto
renow.
Iam
goingto
thesto
renow.
Past
particip
le-en
Ihav
ewalk
edto
thesto
rebefo
re.Ihav
egoneto
thesto
renow.
43Past
Ten
seResu
lts
025000
5000075000
Num
ber of Words
0.0
0.2
0.4
0.6
0.8
1.0
1−OR, Responses
Overregularization in B
p
OR
Responses
b)
a)
025000
5000075000
Num
ber of Words
0.0
0.2
0.4
0.6
0.8
1.0
1−OR, Responses
Overregularization in Leabra
OR
Responses
44Past
Ten
seResu
lts
Bp
L H0
L H001
L H005
0 25 50 75
100
Responses
Early C
orrect Responding
To 1st O
RT
o 2nd OR
Bp
L H0
L H001
L H005
0 50
100
150
200
Overregularizations
Total O
verregularizations
45Past
Ten
se:Summary
•Leab
rapast
tense
model
showsthat
youcan
get
U-sh
aped
pattern
from
amodel
with
outman
ipulatin
gthetrain
ing
enviro
nmen
t
•Ach
ieves
asu
bstan
tiallev
elofcorrect
respondingprio
rto
onset
ofoverreg
ularizatio
n
•Overreg
ularizatio
ncontin
ues
atalow,sp
orad
icrate
over
an
exten
ded
perio
doftim
e.(does
thisev
entually
goaw
ay?)
46
Past
Ten
se:Summary
•Thecontro
versy
contin
ues!
•Rules
accountpred
ictsthat
past
tense
acquisitio
nwill
be
sudden
andthat
itwill
beinsen
sitiveto
seman
ticfacto
rs...
•Counterex
ample
from
Ram
scar(2002):
–“frin
k”in
thecontex
tof“d
rink”=fran
k
–“frin
k”in
thecontex
tof“b
link”=frin
ked
Ram
scar(2002)
47Past
Ten
se:Summary
•Thecontro
versy
contin
ues!
•Rules
accountpred
ictsthat
past
tense
acquisitio
nwill
be
sudden
andthat
itwill
beinsen
sitiveto
seman
ticfacto
rs...
•Counterex
ample
from
Ram
scar(2002):
–“frin
k”in
thecontex
tof“d
rink”=fran
k
–“frin
k”in
thecontex
tof“b
link”=frin
ked
•Neu
roim
agingirreg
®
past
tenses
show
overlap
pingneu
ralactiv
ation(Jo
anisse
&Seid
enberg
,2005)
–nottw
osystem
s.
•See
alsoMcC
lellandvsPinker
smack
downin
Tren
dsin
Cognitiv
eScien
ces(2002)
48Questio
ns
•What
gen
eralprocesses
areinvolved
inread
ing,an
dhow
do
these
sometim
esfail
(e.g.,in
dyslex
ia)?
Distrib
uted
lexico
n(orth
o,p
hono,sem
)
•How
areweab
leto
read“cat”,“y
acht”,an
d“n
ust”?
Ran
geof
contex
tdep
enden
trep
s&
contin
uum
ofreg
ularity
-excep
tion
•Whydokidssay
“Igoed
tosch
ool”
afterfirst
saying“I
wen
t”?
Dynam
icbalan
cebetw
eenreg
ular
&excep
tionmap
ping
•How
dowordscometo
mean
anything?
•How
dowegobey
ondwordsto
senten
ces?
49How
DoWordsCometo
Mean
Anything?
•What
Gives
WordsTheir
Mean
ing?
•Where
Does
thisMean
ingComeFrom?
50What
Gives
WordsTheir
Mean
ing?:
Distrib
uted
Sem
antics
orthographyphonology
action oriented
tactile
visual
auditory
kinestheticform
color
3−D
telephone
footbrake
kettle
velvet
cloudsthunder
Sem
antics
isdistrib
uted
across
specialized
processin
gareas.
51Where
Does
thisMean
ingComeFrom?:
Correlatio
nal
Sem
antics
Heb
bian
learningen
codes
structu
reofword
co-occu
rrence.
Sam
eidea
as:
•V1recep
tivefield
learning:learn
thestro
ngcorrelatio
ns.
•Sim
ilarto
Laten
tSem
antic
Analy
sis(LSA)
Input
Hidden
52[sem
.proj]
53Multip
le-Choice
Quiz
0.neu
ralactiv
ationfunctio
n5.
attentio
nA
spikingrate
codemem
bran
epoten
tialpt
Acompetitio
ninhibitio
nselectio
nbinding
Binteractiv
ebidirectio
nal
feedforw
ardB
grad
ual
feature
conjunctio
nsp
atialinvarian
ceC
languag
egen
eralizationnonwords
Csp
ikingrate
codemem
bran
epoten
tialpoint
1.tran
sform
ation
6.weig
htbased
prim
ing
Aem
phasizin
gdistin
ctionscollap
singdiffs
Alongterm
chan
ges
learning
Berro
rdriv
enheb
bian
taskmodel
based
Bactiv
emain
tenan
cesh
ortterm
residual
Csp
ikingrate
codemem
bran
epoten
tialpt
Cfast
arbitrary
details
conjunctiv
e2.
bidirectio
nal
connectiv
ity7.
hippocam
puslearn
ing
Aam
plifi
cationpattern
completio
nA
fastarb
itrarydetails
conjunctiv
eB
competitio
ninhibitio
nselectio
nbinding
Bslo
winteg
rationgen
eralstru
cture
Clan
guag
egen
eralizationnonwords
Cerro
rdriv
enheb
bian
taskmodel
based
3.cortex
learning
8.dyslex
iaA
errordriv
entask
based
heb
bian
model
Asu
rfacedeep
phonological
readingproblem
Berro
rdriv
entask
based
Bsp
eechoutputhearin
glan
guag
enonwords
Cgrad
ual
feature
conjunctio
nsp
atialinvar
Ccompetitio
ninhibitio
nselectio
nbinding
4.object
recognitio
n9.
past
tense
Agrad
ual
feature
conjunctio
nsp
atialinvar
Aoverreg
ularizatio
nsh
aped
curve
Berro
rdriv
entask
based
heb
bian
model
Bsp
eechoutputhearin
glan
guag
enonwords
Cam
plifi
cationpattern
completio
nC
fastarb
itrarydetails
conjunctiv
e
54Sen
tences:
Bey
ondjustsem
antics
Trad
itional
approach
:S
(subject)
Art
N
The
boy
VN
P(direct object)
chasesthe cats
NP
VP
Altern
ativeap
proach
:
Distrib
uted
repsofsen
tence
mean
ing:Thesen
tence
Gestalt!
(gestalt
=unified
configuratio
nofelem
ents
that
can’tbedescrib
ed
merely
asasu
mofparts)
55Sen
tence
Compreh
ensio
n
•Wewan
tto
build
anintern
almodel
ofthesitu
ation
•e.g
.,“Theteach
erdran
kPep
siin
theclassro
om”
–Who/what
istheag
ent?
teacher
–What
isthepatien
t(object)?
Pep
si
–What
did
theag
entdo?drin
k
–Where?
classroom
(andso
on)...
•Goal:
Teach
amodel
tounderstan
dsen
tences
•Presen
toneword
atatim
e
•Wan
tthemodel
tobeab
leto
answ
erquestio
ns,
e.g.,W
hoistheag
ent?
(has
tobeab
leto
dothisev
enifag
ent
notcu
rrently
ininput)
56ToyWorld
Peo
ple:
busd
river
(adultmale),teach
er(ad
ultfem
ale),schoolgirl,
andpitch
er(boy).
Actio
ns:eat,d
rink,stir,sp
read,k
iss,give,h
it,throw,d
rive,rise.
Objects:
spot(th
edog),steak
,soup,ice
cream,crack
ers,jelly,iced
tea,koolaid
,spoon,k
nife,fi
nger,ro
se,bat
(anim
al),bat
(baseb
all),
ball,b
all(party
),bus,p
itcher,
andfur
Locatio
ns:kitch
en,liv
ingroom,sh
ed,an
dpark
.
Syntax
:Activ
e&
Passiv
e,phrases.
Someev
ents
more
probab
lethan
others
(egbusd
rivers
eatsteak
more
often
than
teachers)
57Netw
ork
Input
Encode G
estaltG
estalt Context
Decode
Role
Filler
Toan
swer
questio
nsat
theen
dofsen
tence,
net
need
sto
actively
main
taininfo
aboutwordsithas
seen...
SRN
58Train
ing
•Presen
twords&
their
roles,o
neat
atim
e;aftereach
word/role
pair,q
uiz
thenet
onwhat
ithas
seenupto
that
point
•Thebusd
river
stirredKool-A
id
•Presen
t“b
usd
river”
+ag
ent
–Whoistheag
ent?
busd
river
•Presen
t“stirred
”+actio
n
–What
istheactio
n?stirred
–Whoistheag
ent?
busd
river
•Presen
t“K
ool-A
id”+patien
t
–What
isthepatien
t?Kool-A
id
–Whoistheag
ent?
busd
river
–What
istheactio
n?stirred
59Tests
Task
Sen
tence
Role
assignmen
tActiv
esem
antic
Thesch
oolgirl
stirredthekool-aid
with
asp
oon.
Activ
esyntactic
Thebusd
river
gav
etherose
totheteach
er.Passiv
esem
antic
Thejelly
was
spread
bythebusd
river
with
theknife.
Passiv
esyntactic
Theteach
erwas
kissed
bythebusd
river.
(contro
l)Thebusd
river
kissed
theteach
er.Word
ambiguity
Thebusd
river
threw
theball
inthepark
.Theteach
erthrew
theball
intheliv
ingroom.
Concep
tinstan
tiation
Theteach
erkissed
someo
ne(m
ale).Role
elaboratio
nThesch
oolgirl
atecrack
ers(w
ithfinger).
Thesch
oolgirl
ate(so
up).
Onlin
eupdate
Thech
ildate
soupwith
dain
tiness.
(contro
l)Thepitch
erate
soupwith
dain
tiness.
Conflict
Thead
ultdran
kiced
-teain
thekitch
en(liv
ing-ro
om).
60Gestalt
Rep
resentatio
ns
SG G
estalt Patterns
Y
X0.00
1.00
2.00
3.00
4.00
5.00
6.00
7.00
8.00
9.00
10.00
11.00
12.00
13.00
14.00
15.00
16.00
0.002.00
4.006.00
8.00
sc_at_so.
bu_at_so.
te_at_so.
pi_at_so.
sc_at_st.
bu_at_st.
te_at_st.
pi_at_st.
pi_dr_ic.
sc_dr_ic.
bu_dr_ic.
te_dr_ic.
pi_st_ko.
bu_st_ko.
te_st_ko.
sc_st_ko.
61Problem
swith
thestatistical
approach
?
•Themodel
mak
esmistak
esforinfreq
uen
tan
d/orirreg
ular
senten
ces
•Exam
ple:
busd
river
atesoup;resp
ondswith
steakas
patien
t
•Explan
ation:Net
sawbusd
river
eatingsteak
7xmore
than
soup
•Statistical
model
overrid
esreality...
•Peo
ple
suffer
from
similar
biases!
(How
man
yan
imals
did
Moses
brin
gontheark
?)
62Questio
ns
•What
gen
eralprocesses
areinvolved
inread
ing,an
dhow
do
these
sometim
esfail
(e.g.,in
dyslex
ia)?
•How
areweab
leto
read“cat”,“y
acht”,an
d“n
ust”?
•Whydokidssay
“Igoed
tosch
ool”
afterfirst
saying“I
wen
t”?
•How
dowordscometo
mean
anything?Statistics
ofword
co-occu
rrences.
•How
dowegobey
ondwordsto
senten
ces?Sen
tence
gestalt
63 Applicatio
n:NetT
alk!(Sejn
owksi&
Rosen
berg
,1986)
•Learn
sto
read&
pronounce
english
text
•Inputsare
oneof29
chars
(26letters
+sp
ace,comma,fu
llsto
p)
•7letter
window
(provides
contex
t).total
=29x
7=203
inputs.
•Hidden
layer
of80
units.
•Outputgen
eratesoneof60
phonem
es,rep
resented
by21
articulatio
nunits
and5units
forstress/
syllab
leboundary
info.
64NetT
alk:Resu
lts
•Learn
sreg
ularities
ofen
glish
speech
•Gen
eralizesto
novel
wordsnotin
trainingset
with
78%
accuracy
65NetT
alk:Resu
lts
•Knowled
geisdistrib
uted
:relearn
ingafter
dam
agemuch
faster
than
orig
inal
training
•Distrib
uted
(spaced
)practice
more
effectiveforlongterm
retentio
nthan
massed
practice
66NetT
alk:Im
pressiv
e,But...
•Solves
readingan
dsp
eakingat
once
(unlik
epeo
ple)
•Doesn
’tad
dress
specializatio
nofdifferen
tbrain
areasin
languag
eprocessin
g.
•Uses
biologically
implau
sible
“errorback
propag
ation”meth
od
fortrain
ingweig
hts
•Explicit
“teacher”
provides
correct
inform
ationonoutput
articulato
ryunits
(instead
oftrial
anderro
rlearn
ingthat
we
hav
eto
do)
•Req
uires
man
ypasses
throughexact
sametrain
ingset
(rather
than
natu
rallan
guag
eexperien
ces).