I,AN(WA(;F, AC()t:ISITION AS I,\],;AtlNIN(; 
MIKIKO NISIIIKIMI, t\[II)I,:yt(K1 NAKASIIIMA AND IlITOStll MATSUIIAIIA 
1.',leer rotechldcal l,almralory 
Tsukuba, Japan 
nisikinliql)etl.go.,hl, nakashilntc~etl.g(i.jtl, lnalsul)ar(!~ell.go.jt) 
Abstract 
C, honasky's proposition that language is handled hy a 
language-specific laeulty needs more justi|ieation. In 
language acquisition ill i)artieular, it is still in ques- 
tion whether the faculty is necessary or aot. We suc- 
ceeded in explaining one eonstrainl oil language ac- 
quisition in terms of a general learning mechanism. 
This paper describes a machine learning system Rhea 
applied to the domain of language acquisition and 
shows that Rhea call learn the tendency which chil- 
dren confronting lie',',' words Seelll to llave. 
1 Introduction 
Chomsky proposed that language is handled by a 
language specilic faculty, but this proposition has 
not been verified, especially ill the area of lan 
guage acquisili(:)n. Although Berwick\[1\] showed 
tile existence of a special nlechanisnl sulfirielfl for 
the learning of syntax, there is still a question of 
whether or not the Illecha.tllslB iS lleCessary, l?lll "- 
therlnore, his model does not explain acquisition 
of setnantics or eoncepis. These were simply pre- 
supposed. 
We started froln a geimral k!arning mecha - 
niSlll alld succeeded ill explainiug a constraint 
on language acquisition. 
Children learldng their ill'st language face 
and solve a big prol)hun of induction. They lind 
(:)lit \]lOW words &l'e ased all(\] )'elated Ic, other 
words from lilni(ed informal\[on at a surprisingly 
ra.pid rate. Ill tim fieM of (leveli)pmental psychol- 
ogy, many kinds of constraints have been pro 
posed to accounl for this I>henonten(m. Most of 
these constraints (:olne froln the view that as 
sunles a specific fl'anlework far language acquisi 
tion, but there is altother view: language as an 
extension of other intellectuaJ faculties, and its 
acquisition as one resnlt of the universal learning 
process that leads to (mr acquisition of intellect. 
We want to explain the children's ability ill 
terms of tile latter view. '\['hus, we make a 
machine learning systenl, Rhea, which arcepts 
n-luple illplltS elms\[sting of instances froln it. 
donlains (one from each (IonlaJn) and creates 
the rules that delilnil the imssibh , torah\[nations. 
This framework b, very ~;eeera\[, and yet i\[" we 
choose (mi(:,r wc.'lds and lingl, islic des('riplions 
for thetil as tWO input donlains, it can be seen as 
a language arquisilion system without language 
specific constraints. 
In this paper, we describe lhe nm<'hine 
h,arnin~ syslenl Rhea an(l its applicali<n: to Ihe 
donlain of language artluisition. We show \[hal 
with(m! a pri(ll'i inl~.Jrulatio(I almut how outer 
worlds an, or,~anized, Rhea can learn lhe "'set 
ting for new word.~ '~, which children con\['rc, ldillg 
flew w(lrds s(!elll t(I possess. 
The imim is how the model aeqlfir~,~ and 
tormalizes the "meaning" of an expression. To 
achiew, this aut(:mOrlmusly, Rhea has its own rep 
resenlati(m language for outer-worlds. If one lin 
guistir expression is repeatedly given along wilh 
(tifli, renl otHer-worlds, it builds up one ('OmnlOtl 
r(,im~senlati(m tbr ;tl\] lhe (mt(,r-worlds. This in 
|el'llal I'e\[H'eselll;llioll that has a Olle Io ()tie cor- 
respondence to a linguistic expressi(m is regar(led 
as the "meaning" of the expression iu our model. 
2 Constraints 
Ill order to ehicidate the rhildren's rapid a('qui- 
sit\[on of wmahulal'y, constraints ou~ the possible 
hypotheses about I\[1(. meanings of linguistir e× 
pressions hay:' heen postulaled. ('lark\[2\] pro 
poses lh~ prhzciplf of eo'rdrasl whereby every two 
forlns ec.lltrast ill Illeallill~, anIl Marklnan\[:l\] sug- 
gests a Sll'Oll,l~el' ass~Hnplio~t of tft3"011Otllie tll'(\](l+ 
nizatiom 
The assunlp!ion of lax(re,male organizalioll 
cold\[lieS children Io assumil:g lhal a wc.r(l ,givell 
with all unknown objert refers Io a taxonomic 
class of the ob.iert. As ostensive d<,finitioll is 
the only way to acquire early wwabulary, the 
assumption reduces the possihle search space of 
area.sing. With this assulnlHion, if yoa see son.e- 
one point Io an un familiar objec! and say a word, 
VOII Call i)resll(lle that I It(, word is eil her ! he label 
of lhe object or the label of one of Ihe categories 
it belongs Io and can li~rget about the possibility 
of the word's I'(!l'('l'l'illg I(I O11(' (If its attribtlt('s or 
AcrEs DE COLING-92, NANTES, 23-28 ^ofrr 1992 7 0 7 I'ROC. OF COLING-92, NA~rrEs, AUG. 23-28, 1992 
( Represent -- Classify __ Generalize ) 
ins~ 
Figure I: ll,\[ormation flow 
its relation to other objects. 
Chil(lren seenl to (:onsJder the assunlpliOll 
of taxonolni(" organization. Markntan's exper- 
iment shows thai even though they are liahle 
to consider thematic relations in dmnains other 
than language acquisition, children hearing a 
new word attend Io taxonomic relations. This 
tendency is called the "setting for' new words". 
It is /tot cleat', however, if such constraints 
are innate or not, or more essentially if they call 
he derived froln restri<`tioas thai ally intelligent 
system should observe. One way to <`\]arify this 
point is to examine whether tit(, model that does 
not contain the constraint can acquire it during 
the learning process. 
8 An overview of Rhea 
3.1 Rhea as a machine learning sys- 
tem 
Fig.1 illustrates Rhea's learning process in 
two dif\[erent domains, A and B. The system's 
task is to find general rules that predict which in- 
stance from l)olltain A can at(pear with a certain 
instal'tee front B, and vice versa. 
Rhea accepts as input a \[)air of instances 
i = (a,b). One instance is fronl l)omain A an(1 
the other front l)omain B. One pair is given at 
a time. Rhea is equipped with an internal repre- 
sentation language for each domain, D.4 and De, 
and has predelined methods to extend the repro- 
sentation languages in case of nee(l. Sindlarities, 
generalization operations and specialization op- 
erations are defined upon ea<,h language. Rhea 
represents all input pair using these languages 
and their extensions, anti makes all internal l'et t- 
resentation D(i) - (1)A(a), De(b)}, which is a 
pab" of a representation of" I)olnain A instan(:e 
att(I that of a l)omaill H instance. More tha.ll one 
possible internal representation may exist for one 
input, but the one found first is stored. 
When represelltations are acctHlflulated, R hea 
is able to find out rules, h tirst sorts intert~al 
representations into classes based on similarities. 
Classes nlay or may not overlap. Then Rhea gen- 
eralizes representations of' each ('lass. This pro- 
cess of <`lassili('ation and generalizatiol~ is, done 
oil demand. 
When a partial input , (an instan<`e frotn l)o- 
main A) is given and its <`otlnterpart b (front l)o 
nmizl B) is to he predlcte(I, tilt, ntodel first (:las 
stiles the partial input isle a class N using tlt(' 
infol'nlalion about a, makes the gelmraLiz;ltion of 
l)omain |l part ttf all the other reltresentalions 
ill <'\]ass :',: and expects one of its spe('ializalions 
to he b's rt,presentalion l)B(b). 
The nmdel \[ortns classes so that reF.resenta- 
lions in ea(:h (:lass share sortie characteristics. 
Two internal representations, (l),t(a\]), l)h,(bj )) 
an(l (l)A(a;\])~ I)B(b2)), belong to ihe same class 
if DA(al) and DA(a2 ) are sinlilar in the trite+ 
rion defitmd in the representation language I) A. 
and l)B(bl) an(l I)B(b2) are also sindlar it\] the 
erileri(m defined in D~j. In the extreme t'ase, if 
l)A(al ) equals 1) 4( a2), then I)H( bl ) nlusl equal 
DMb2) and viee versa, which n,eans that when 
two instarlces frt:,nt Olle domain are represented 
as the same, instances from the other (lonlain 
that altpear with them l`tttlst also }lave the Sallte 
lilt el'l`( a.l ropresett t at ion. 
a.2 Rhea as a language acquisition 
model 
l'thea, when applied to the domain c,f outer- 
worMs ,5' and the domain of linguisl ic expressions 
L tha,t descril)e t he outer worlds, can I)e regarded 
as a language acquisition model. 
In these domains, Rhea learns the followings: 
1. \[",xtensions of the represet,tatioll language 
of linguistic expressions I)1, 
2. hlterlm\] representations (>f linguistic ex- 
pressions Dz(ll ) ..... DI.(I,~ ) 
3. l:\]xteilsions of tilt' representalion language 
of outer-worhls D,s" 
4. htternal represetltations of out(,n'-worhls 
l)s(+t) ..... l)s(,% ) 
5. Classification of inputs 
which resl)e<,tively can be seen as 
I. Syntacti(' rides 
2. Structures (:.flinguistic expressions 
3. (?c.It(:el)tS that delineate mealdngs 
4+ Meanings of linguistic exl)ressions derived 
from Oilier-worlds 
ACRES DE COLING-92, NANTES, 23-28 AOt~"r 1992 7 0 8 PRec. OF COLING-92, NANTES, AUO. 23-28, 1992 
Pigure 2: Rhea as a language aC(luisit, ion niii,h~t 
-(with imcchan 
(time-slices \[ (T~ T2 "m T4 ....... Tnl) I 
S ... , .: 
()Llie~ 
\["igure 3: Scene: a parl (if the inpul 
5. Categories of lhigui~li(: exlin'(,ssions. 
Fig.2 shows tire conliguration of lh(! language 
al:quisition nlodel, Rtiea. 11 rec(,iv(,s a pa_h" of Oil(, 
scelt~, and a linguistic expression thai describes 
ill() SC(Hi(L All OXpl'(!S."iiOll iS a, SPIILI(HI(:(! (if words 
and contahls IIO S~i'/iCtllr~t,I iilt(li'lliatioll. A si:(!Iio 
is tim equlva,lenl of SeilSOl'y hit)ul fronl Olilor 
worlds. Pig.3 shows all e×aliilile o\[ it ~(:(,ii(!, A 
,M:ellO is a Se(lil(~il(:(! of ,<;ilflp,~;hol',~ wliicl~ ar(, lists 
(if assertions thai, be(:onie t, rllo ()1' false at the tilliO 
when the sii~lmhots have been taken. Each as- 
sertiOll expresses a r(!la, tion between two iorlliS. 
The tornis llla)' lie olije('ts, attributes or wthies, 
which cannot tie distingulshed t)y Rllea. 
The parser makes tile iliterna\] rl!l)resollla 
Lions of \]lnguisli(: e×l)ressions, all(\] the Jilter 
finder makes those of scenem The elas,siJTer 
divides rel)resentalions into classes and nlakes 
rules. Sin\]c(' two inputs reprosexnh,(I as the sam(, 
lit (ill(! dOlllaill liltlSI Jlavo tllO Sill\[l(' F(~t)l'(~S(!ll| ;ttiC)l~ 
ill I IIO olh01' (IOlllaili, I\[\[101'o lll;,ly \[)U li(i S.Vllfiliyllls 
or l)olysemaIHs, which means tha! the model has 
"the i)rin(:il)k, of ('()ill l'asl ~' implanted from the 
beginning;. 
4 hlternal l'epresentatlons of inputs 
Th(' inlernal r(,pr(,setltatiim of an inllut is a pair 
of internal n'(,i)ros(mtati(ms of th(, inpul's con- 
Milu(!nls, which i~ a pair of one .~h'uclure and 
one jiltrr. 
4.l Internal representation of llnguis- 
tie expressions 
The ilH(q'nM repr(,s(mtatio, ofa lingHisllc eXl)Ves 
sion is the synla('ti(' structure ol Iho ,>xl)ri!ssioul. 
l.'or examplo, a \[inl~uimi(" expressi(m ''Kitty 
ato pancakos ~ is inlernmlly r(,I)U'es(,nt0d as a 
S: (S~ntence (Classl 'Kitty ~) 
(Class2 (Class3 'ato') 
(Class4 'pancakos '))) 
The first eh, menH in the list specifies t, lw name 
of the ('hL'~s Ih(, slrll('lUl'(~ b(qong~ to and lh(, resl 
ar(, ils c(mslitu(,nis. Each constituent in turn has 
i~S class llalllO alld C()llSlittl(!nls, 
The rellreslmtalioil languagl~ l)t, al the begin- 
ning ('OlllaiIIs supI)ositions lhat Oil(! input (,XFq'(,s 
sion \[i.'ms one s~,'uclur(, and cau hi, d(,scrib(,d 
with a l)hras~' slruclm'(' grammar, Th(' model 
a('c(,pts a now inlmt (,xpu'es.si(m I)rovided thai it 
can b(, descrihed by adding al in(is1 ol,e new ,'tlh,. 
~;hoii kllOWii i'tllOS (:allIiOl pal's( ~ ant oxprossi(m, 
Rhea l)al'sOs it from the bottom to tnp and flr(llil 
t}ne top to dowtl sinlultalleously and mak(,s par 
lial slruclur(m. \[f th(,y can \[m comllin(,d into one 
structure bv a(idhlg one rub,, l{liea adds the rulo 
1o lho llilqliOlV as all OXll,llSion of D L. lfoliO i'llit, 
('aiill()l c()illl(!Cl all of I holil, I ho model backl racks 
to find anolh(,r lnil'Silig of ailalldOliS f. tio ililllil. 
llhea sels Ih(' class (ll all illl\[<llfiWii v,,ol'(l C(ll/ 
si(lerhlg the ~(:('n(' ~iVlql with Ihe wor(l, ltxl:liei\] 
SOllle rule predh'is the ('lass flf Ill(, w(ird and tim 
~(:(,Ii(! i)r(,sonie(l with the word Call be giv('n all 
ilit(,rnal r(!t)r(~s(!llLali(in sitnilar to Ihos(, (if other 
words in the class, Ill(, w(ir(I is ad(le(l to the tire - 
dicled class. If nol, a iil,,w caleg(iry is a~signe(\] 
Io the word. 
4.2 Internal representation of scenes 
An internal r(,pi'esontallon of a scene I)rovides Ihe 
s(,ananlic~, of Ill(' linguimic expression that comes 
AcrEs DE COLING-92, NANTES, 23-28 ^oL~r 1992 7 0 9 PROC. OY COLING-92, NANTES, AUG. 23-28. 1992 
~e.e s2 ~.....--.--- ~ Focus of atleniion 
*,@~i** i ) f(s2) 
f"igure 4: Relalionship among Filter, Scenes and 
l"tlcllS Ill Attentions 
witii the scene. Lhlguistic expressions cilange or 
cotll tel tile listeners' illlerprelaliolls of tile outer 
world, and make speakers and listeners share lille 
focll,s of at(chiton (hereinafler, F()A). hi order 
to niodel this process, a scene is internally top 
resented a.s a prlwellure thai COllVel'IS the SCelle 
inlo ail FOA. We call this procedure a filhr. As 
stated before, a SCelle is il seq/lellce of lists of 
assertions, and so is all FOA. PeAs must con- 
lain .:it lea.st olie non-variable assertion lleeatlse 
there rnusl exist non-variable FOAs to tie shared 
arulnig speakers a, ll(I listeners. If a filter applied 
to s(tene s yiehls a non-variable sequence el lists 
of a~ssertions, the filter is valid for .s. 
Any valid filter for sceile .s can lie a reln'e- 
selitation of the sceile. \]\[?o1" t,×ainple, a scelie 
(liar COil(sins solneolle eating pancakes ilia)' lit. ' 
internally represented ill several ways. A llroee - 
dilre (hal focuses lhe listeners' aJtentioii Oil pal(- 
cakes and yiehls pallcakes as ;Ill VOA is valid 
for tim scene, and one that stresses lhe eating 
aclion call also lie all internal representation of 
the scelle, tlowevel', scelleS whicii appeared v.rith 
lilt' saliie expression milS1 have lhe sanie fiher 
t)ecallse iiiere lllay tie iio polysenlalltS. 
Fig. 4 shows the relationsl(ip auiolig fillers, 
sc0n0s alld f"()As. Sill(:(' the FOAs derived by 
fillet" f frOlll SCelle ,sl and scelle ,s~ both contain 
sonic objects, the filler is valid for both scenes. 
Thus two sceiies thai appear with linguistic ex 
presslon / are represented by tile filter. 
4.2.1 Represeutation language of filters 
l"ilters art, mappings from scenes 1o I"OAs. ilhea 
has 32 parameterized sinlple nlallpillgs as its 
representalioll language Ds al the start. Ii 
COlllliilleS lllappings and seaFches a given s£elle 
for values to instant(ate paralneters, %%:(' call 
these paralneterized lnapplngs Jill+r-primiliv6.s 
and ilistantialed niapphlgs fillfr-~l~ln~ll.,>. I;'or 
instance, alllOll~ the imssible conibinalions of 
filler-prinlitives is the olie 
(snap-remove no,i-include *vat±able*) 
which relilOVeS assei'liolls that do not COl(tail( ;4 
certain lerlil froill a stlapshol. !,Vlien a SCelle 
is giVell, the model selects Olle (iF Ihe terms in 
the seen0, nallleiy $3_oca~5±on, Io Sllbslilllle \['OF 
*variablG* alld lllilkes a filter-eleulenl 
(snap-romovo not-includo $1ocation) 
which exIracts assertions thai ('el(lain the terui 
$3.oca'cion fronl a snapshol in llle scene. A fil 
|el is il sequelice of olle or lllore titler-eleillelltS. 
Piiter-elenlents in Ihe seqllellce are applied i,o 
a scelle lille I)y olie and lhe reSllll becoul0s the 
I.'OA. 
4.2.2 Acquisition of filters 
Rhea shapes fihers lhrolngh lrial and error. 
~li~h£~lle%'t`'l' ;I llew scelle is given wilh all expres 
skin, the filler thai seems to correspon(I Io lhe 
e×pression is tested for its validiiy for the IleW 
sceile, and Rhea lhen elal)orales or corret`'ls Ihe 
filler depending on lhe result. 
VVheli the new input (l,s) is gi'¢ell~ Ill(, lllOde\] 
creales 1)i,(1), which is the representation of l 
by the \]aiigtlage \])l,, and searches lhrough the 
lileiilory tbr a ropreselilatioil lhal has (tie for(it 
(l)l,(lJ, f), where f is an internal represelilation 
of aii instance tl'Oill l)oiriain ,q'. 
\]f there is 11o r0presentaiion of lhe fornl 
(IlL(I), f), I is regarded a.s a new expression and 
Rhea builds a candidate lbr filler f. The can- 
didate consists of one filter elenielil made by st' 
letting one fiher-pl+iniitive randotnly and subsli- 
luting terms in the giveli scelle ,,; fol" parauielers 
of the fiiler-ilrinlilive. If the candidate is valid 
t(:ll' sc01ie ,+, ii is ilSOd as all h/terlial repr(!selita 
lion of the seell(!. If it is liot, ant`tiber candidate 
is erealelt alld lesled+ As t}lere IliliM lit, tie SVll- 
oily((is, a tilter lliliSl lie differenl frolil those of 
ot ht`,r exln'essiOllS. 
If tii,,a already knows the Ihiguislh" e×pres- 
sion l, thai is, if the represenlalioli of the fornl 
(liD(l), f) is iIi the nlelilory of" 11 liea, ii cliecks Ihe 
valldily of lilter f for stone <+. Rhea elaborates 
valid filters alid correcls invalid inle~-;. 
Elaboralion is to ulake filters Ill(ire specific by 
adding conditions. Rhea nlay either hlserl erie 
ran(Ionlly seh, cle(I filiel'+olenienl inlo I, he exist- 
ing filter of replace Olle \[i\]ier-elenielit by a lilOre 
sllecific olle. For each illplll, the niodel fail add 
Ollly one condition, st) leai'ning proceeds grallu- 
alll'. The IleW \[iher lllllSl i)e diltbrelll \[rolii the 
Acr~ DE COLING~92, NANTES. 23-28 AOi~r 1992 7 1 O PROC. OF COLING-92, N^~rEs. AUG. 23-28, 1992 
fillers of other extiresslolis alld IlltlSl extract ail 
\]~'()A whieit is (lifl'erenl \[I'o111 the {ill(! derived Ily 
the old filter. If Rhea cailliOt elaborale lhe fiher 
t(i (flake lip a ilew oiie, it keeps the ohl lille. 
Correct, ion of a filter is d(liie ll.y deleting COll- 
ditions. Rhea keeps a reli~/orl couulfr 17 for ev- 
ery internal represe/itatioli. 11 is tile lllllilt)er (if" 
sliccesslve scenes from which the filter CallllOI ex- 
h'act all I?OA and Rhea cannol correct it. T(Icor 
recta liil(;r, Rhea luay l'eniove j iilter.elenlelits, 
replace parameters (if k fiher-elenlelils with other 
values extracted l'roln sceiie .s or replace l filler 
eleuierits witii lllOl'e gelleral oties. The llllmller of 
changes j q- k + l, however, niusl lie( exceed lhe 
value of the revision c()Illllei'. ~Vhell the ct)rre(- 
lion succee(ls, lthea sets Ill(, revision eOllilter t(i 
zero. If the filter Ca, llliOl lie niade valid for .~cene 
s within the allowed nillliber of i:ha, iige~, Rhea 
keeps it and incrc:inents the revision counter by 
Olle. 
5 Classification and generalization of in- 
put 
Rhea divides internal representations into 
(:lasses. A (:lass contains representations lhal 
have both shriller str/ictlires alld similar filters. 
As classes niay overlall, all interlial rel)resellia 
t\[Oll Call be a llFleFtlber of two OF Ill(ire classes. 
5.1 Similarity of structures 
Two structures are similar if they are in iu- 
terchangeable positions wlthiil bigger structures. 
l'br e×ampl0, liavhig two sti'llClllres: 
S:t: (Sentence (Category1 'yellow') 
(Category2 'pancake')) 
52: (Sentence (Categoryl 'red') 
(Category3 ' raspberries ~ ) ) 
ilia)' trigger the making of a, class that 
COilta\]ns two r0l)reselltations whoso slrtletlil'0S 
are (Category2 <pancake') and (Category3 
'raspberries') resl)ectively. These struc- 
ttlres are ,'~{rnilor becallse lhey both have till(! 
Categoryl a~s their sister (:lass alld (el'ill Iilenl- 
bel',~ of tile Sentence (:lass. 
5.2 Similarity of filters 
Filters are lists of filter-elenienis. Two filters are 
sitnilar when they can be g~neralized into the 
same non-null and non-variable list. Rhea has 
the following genera\]izatlon (= dropping dOWll 
conditions) operations. 
I. deletion el lransforlnatloli iillO a variabh, 
of a filter elenient ala specified position hi 
the list 
7. delelion fll" lrails\[oriIlaliOll bile a wlriable 
of \[iller-elenlenls belweeu those thai nialch 
certaill patterns 
3. transforluation bile a variahh, (if a llarl (if a 
filter elenient ala ,~llecitied tiosithln hi the 
lisl 
If a seqnelice of olleraliolis is alipiled to a set of 
fillers aud yMiIs a COlliiilOll aud non-lrivlal re- 
sull, Ihe hilerual representalions Ihal have Itiose 
filters Call (el'Ill olle (:lass. 
l'Tli' exaiuph,, all internal i'elll'eselllalioli willi 
a tiber ((F x y) (G v)) aild aliolher reln'eseil- 
tatioii wiiose tiller i~ ((F x z)) lllay I)elolig lo 
lhe Salile (la~s because t If(, 11Oll trivial generaliza 
lion of the two filters ((F x *variable) ) exists. 
5.3 How classes can be used 
As described iu s/IbsecliOll 3.1, a rlass COllsl raillS 
ils lllelllbers Io a certain fornl of repre.'-;elilaliOll. 
Tiiere are two ways for lhe Illodel Io 11so Ibis re- 
striction. 
()lie way is based Oli Ihe class illS(aliCe r01a- 
tiOllS aiiiOlig repl'eselltali(lllS. ~¢Vo (!all deiilal'ca,|e 
life search space for the ltleallillg of the e×llres- 
sioa if it~ cla~s is knowu. 
Rhea, ill lleed of filldilig Ihe filler pah'ed 
wll h a strilClllre, \[irst deternihies lhe class oF the 
strllCllire, gelleralizes all Ill(, lillt,rs (if llielul)ers 
(if tilt, class aild e×pecls lhal the liher hi qiles- 
I, iou is flue (if the specializalioll~ of lhe general 
ized filters. Specializatiou is done tly subslillll- 
iug wihies for variables iu the generalized Iilter or 
ad(Ihlg lille (n' lllore liher-eh,uienls Io I, he filler. 
The oilier way utilizes inela4"elaiionships 
of i'elaiioiisliillS aill(lllg relu'esenlaliolis. The 
strucl fires deline whole llarl relal ilmsliips anioiig 
thenlselves. \[{etn'esenlations o\[ a class are ex 
\[)ected I(i share st)llle chal'aClel'iSlics (if thest, re- 
latioushills. ~'(' C~ill guess Ill(' llleallill~ oF a Sell 
telice Iti;il wa,'; ilever heard before. 'l'\[iis happens 
wllen we know all lilt' coiislilileiil words and how 
theh" iilealiings COlllributl, Io lhe Ineallhlg (if l lie 
wiiole sell{OllCe. 
\~,'hen a new Iingulslh" expression is given aud 
ret)re~etll0d hi a Sll'llClllre~ Rhea ('all accelerale 
the ~earch for the \[iher paired wil h il if tile filters 
(if its conslillleillS are klloV.,'li. 11 lirsl identifies 
Ihe slrllCtllre's class, and lheil lllakes till(' rule 
tot each tilt,lilt)el' (if lhe <'lass lhat exlilains how 
the fiher elthe Illeull)er i~ broken down inlo tilt" 
ACRES DE COLING-92, NANTES, 23-28 AOt~" 1992 7 1 1 PRec. OF COLING-92, NANTES, AUG. 23-28, 1992 
Table h Possible Ioruls of inpul senton(:es 
<s> : := \[<~>\] \[<~>\] \[<V>\] 
<N> ::= <a><N> I <n> \[ <p> 
<V> ::= <v> \[ <a> 
<n> ::= "asi '+ \] "aZama" \[ "ahiru" I "okasi" 
\[ +'cup" l "kuti" \] "glass" I "co~fse" 
\]"sara" \] "spoon" \] "tabemono" 
I "tukue" \[ "re" I "ikimono" \[ "neko" 
\] "pamcake" \[ "milk" \[ "me" 
<p>: := "Kitty" i "Sacthan" \[ "Hney" 
\[ "Dowsy" \] "Louie" \] 
<v> ::= "aru" I "ugoku" \["sawaru" 
I "taberu" I "nai" 
<a> ::= "kiiroi" I "amai" I "kuroi" 
\] "maxui" 
fibers of' its eonsliluents, fl then generalizes all 
these ('tiles and expects that a st)e('iaJizaticln of 
the ge.neralize(I rule applies to tim strll('tllre ill 
question. Therefore, lilies l>uts the filler of its 
COtlStitllenl.s lille the general rule and eOlllposes 
a candidate tor the filler of tile whole slrlletllre. 
Tile nio<lei can liniit the search space for tile fil- 
ter to specializations of the ('an(li(late. 
6 Experiment: one-word sentence 
We test tim niodo\] to see whether it can aetluire 
the "setting" for now words given as otto-word 
so.n tel/Ces. 
An inpul scene is seleete(t fronl 48 possibili- 
ties that we trove prepared. The lexicon has 32 
words, but no( every word can descrilm a given 
S(;Olle, thlls for each S('(!lle we nia,d() a liSl Of wor(is 
that can lie use(I to (leserille it. |,ingulstlc expres 
siotls are randonfly eomi)osed using the words in 
tim llst and the gramiltar showil ill Tahle 1, 1 and 
are restricted to tie IIIOl'(! l}lall a lenglh of three 
words. These <n>, <p>, <v> and <a> roughly cot'- 
respf)lld iO ll()llns, pl'opei'-nOllllS, verbs alld all 
jeclives. 
After 4:12 pairs were input, Rhea divide(t :12 
words into tlnree, unconnected classes: ('lassl, 
('lass2 and Class3. hi the hiternal representa- 
tiOIIS of tWO or three-word SellteileeS, they were 
IEngiish tritnshti, ions of I.ernlinld symbols in 'l'id)h, I 
arc: 
<n> ::= "leg'+ i "head" I "duck" I "sveeis" 
I "cup" i "mo~Zh" \[ "glass" \[ "coffee" 
\] "pla±s" \[ "spoon" I "food" 
I "table" I "arm '+ I "living thing" I "cat" 
I "pancake" I "milk" I "eye" 
<v> ::= "to ex~s£" \[ "to move" into %ouch" 
I *'to eat" I "not to exist" 
<a> ::= "yello." \[ '*sweet" I "black" 
\] "round" 
((subseq 0 O) 
(snap-count all) 
(snap-sort all maxcount) 
(map snap-remove not-include *vaxiable*)) 
Figure 5: The general fiher of one-word sentences 
filrther (tivided into subclasses, \[:.tit here for sinl 
plicity, we ('Ol/('elltrato Oil the (lasses llia(h' If) 
express Olle-word s01tten('es. ('lassl e()lit.ainef\] 
one <v> word "aru" (to exist), ('lass2 (:ontained 
another <v> word "nat" (not to exist) and all 
ot}mr 30 words were classitied into the last class, 
Class3. 
Rhea learned that the word in ('lassl is ass()- 
ciated with a lih.er tilal extFacts assertions that 
Dec(isle I rile al the titlle of Ill tel'alice, slid t he fiI 
lers of the wor(I ill (71ass2 extracts Oll\]y assertions 
tiial beconw false at utterance. 
Fig.5 shows the generalized lilter of Class3. 
It (na.kes l)arameterized modificalio(ls to scenes. 
The first \[ilter-elenionl (subsoq 0 0) extracts 
changes ;it the time of utterance, (snap-count 
all) COtlnts how iilatly (isles each term all- 
pears ill the snapshol and (snap-sort all 
maxcount) changes order of assertiolls itl tile 
snapshot so thai mssertions that e(mtains the 
tel'll| thai appears lnOl'e frequelllly eollle eal' 
Iier. The lmsl filter+element (map snap-remove 
not-include *variable*) }tam a varial)le and 
Rites lta.s to select a torsi front the stiapshot lo 
sltbstil.ute for it. Tile s(lbstit(lto(l IiltorIelOn'lenl 
extracts assertions tiiat eontaill the terln. As 
lite tosser(ions in tlie snapshot are lhus sorted, 
the terlli that appears lnost frequf, nlly is seh,eted 
first, and the filter thai foeuses Oil tile terill iS 
tested for its va.lidily first. 
As for the reilttlonship i)etween a one-word 
selllence an(I its only eonsl;iluelil wor(I, Rhea 
corijeetiire(\] lhal tile filter of the senton('e is the 
s~lne as that of the word. 
hi short, llllea acquire(I the general filler tot a 
group (if one-word sentences and ii exlraels sucii 
assertions tliat deseril,o a tel'ill thai al)l)ears lilOS\[ 
frequently in the snapshot at the time of (liter- 
alice. As Rhea backtracks, asserlious with the 
ilOXt niosl fr0qlteul terln are ex/ra.cte(\]. 
S('elleS have inore labels for an object than 
labels for its attribules because eai'h assertioli 
expresses a re\]atioll I)etweelt two tertns and all 
object label appears in all the a.ssel'lions about 
ils aitribules. "\]'here|ore when the niodel is given 
a Olle-wor(\] Sellteliee wllose COllstii ilent word does 
IlOl belong to classes of words of exlstence/ncni 
ACRES DE COLING-92, NANTES, 23-28 AOLrr 1992 7 1 2 PREC. OF COLING-92, NANTES, AUG. 23-28, 1992 
existellce, it first assumes the Sellleltt'e I0 refer \[0 
the label ti:)r an object ill the scene. If the label 
is already known, tile model lhen ba(ktracks to 
refer to the label for its tnost salient attribute or 
a lahel for ;utother oh jet1. This is what ehihlren 
with the "selting f<>r new words" w(ndd do facing 
a llew Olle word Sellleltce. 
7 Discussion 
7.1 Semantic concepts and input 
Other acquisition models Ihal cover semantic 
acquisition are the syslem of Takagi el. al. 
\[4\], which accepts a sentence and visual input, 
Hill's language acquisition tnodel\[5\] and Self- 
ridge's CtlILI)\[@ llowever these models as- 
sunte semantic concepts front the slarl, and their 
task is to associate linguistic entities with thetrt. 
These systems, which receive a sentantic COtlCel)l 
to be associated with a linguistic expression as 
direct input, c81111ol 'ntisundersland I he tncaning 
of a linguistic expres:slon and cannot she(l light 
Oil tilt> difficulty of learnirtg the meaning of a cer 
tain expression. 
'~'Ve do llOt a~%Sl=lllle senlalltic £OIl(;epts in rep- 
resenting scenes given to Rhea. We formalize 
concepts as filnctions fi'onl the direct input to 
FOAs. They must |)e fi:)rnted and 1ested in ac 
corda.nee with expressions anti other concepts. 
We eqltiptled the model with filter=priutitives, 
which are means ofeslal>lishing the concepts. We 
have designed filler-i>ritnitives to I)ecome equiv- 
alents of human abilities of recognitlotl, l"ilter- 
printitives are given fi'onl the beginning I)ecause 
human beings have the abilily to focus 1heir at- 
tention Oil objects, attributes or changes when 
they begin language acquisilion. Rhea can se+ 
loot a \[)arameler frolll scelles alld make coilcrete 
lilter-elentents just like any child coming 1o dis- 
tinguish imtmrtanl features in its world. There 
ft)re, our formalization of con(:et)ts and its acqui- 
silioll process is a more realistic Olle. 
7.2 Acquisition of a constraint 
The principle of contrasl is deriw,d from the gen 
eral constraint on how a (:lass shouhl I)e fortned 
to make useful l)rediclions, and as shown in sec 
tion 6, Rhea has no language-specifh: constraints 
but yel can acquire tile "setting for new words", 
because its \[ilterq)rimitives and classification cri- 
teria can tel)reduce the tendency thai was con- 
tained in the input pairs. 
Ill Ollt' experilllellt, the one-word SelltellCeS 
given to llhea were often laxononli(' terms or at- 
tributes of any oltjecls in the scene and Rhea 
learned thai the best conjecture is that the one- 
word sentence presented with unknown objects 
refers lo ;i taxonomJ(; ieri'll of the lltOSt frequently 
describe<l objecl. If we give a label for the biggest 
oil jeer ill the SCell(! whelleyer llhea llleel~s a scefle 
with muhiple objects lhal are not yel laheled, 
Rhea will make a tiller of a cal.egory that sorts 
el)jeers I)y size and exlracls l.he \[irsl one. Our 
claim is thai chihh'en can also acquire the "sel- 
ting for the new words" fi'om a few inputs of 
olle-wor(t selllellces, ~lll(I thai it lleed llOl to I)e 
set a priori. 
8 Conclusion 
This paper ha+'+ described Rhea, the model of lan- 
guage aequisillou, which uses wwy general aC<lui- 
sition procedure. We assume neither semantic 
concept..,+ nor syntacli(' rules a priori, lnsiead, 
we have equipped the model with the general 
franlework to create the rules thai delimit tilt, 
possible conlhinatious of the input. We applied 
the model to the (Ic,1llaillS of Otlter-wt:q'l(Is alld 
linguistic descriplions of thent. The svsleIll Silk'- 
cessfully made concepts that are consistent with 
giwm inputs. "\['he experinlenl showed that il 
reproduced the "setting for lhe new words," a 
human lendency in language acquisition, with- 
out language-specific constraints or inforntation 
aS(mr hmv ouler worhls are orgattized. 
References 

\[1\] I{. C. Berwick ,k' A. Weinberg (1.9S3): "7'/u 
flrammatieal ba.si.~ of liufluistie perform~mc~ ". 
MIT Press. 

\[2\] E.V. ('lark (1986): "The i)rinciple of c(m- 
lrast: a constraitd on language acquisition" 
in B. MacWhilmey(Ed.), ":U+chani.sma of 
I,azzf'uage AcquisilioW', F, rlbaum. 

\[3\] E.M+Markman (1987): "llow children con 
strain the imssi|)h , meanings of words" in 
U.Neisser(l:;d.), "(/mze~pls o~td conc~pl'ual 
&velopme+nl ", (!anlbridge University Press. 

\[4\] A. 'l'akagi and Y. \]Ill 11987): "'Natural hm- 
yUa\[le prc+:e.ssing" (in Japanese), Maruzen. 

\[5\] .l.C. Hi1111983): "A model of language acqui- 
sition in the two-year-old", Cog. lh'ain The- 
ory, wfl.6, no.:l, pp.287-317. 

\[6\] M. Selfi'idge (19N6):" A computer model of 
chihl language learning", Artificial lnlelli- 
gence w)l.2.q, pp.171-216. 
