./K,:CCOI:..~\[~; ::,~,.~ \[{ \] :~.!.b<)~\]\[!!¢~\]¢-'.¢~ '~ .... ' ~' % A)~:{}/,_I I;:IW@ : e- . ........ ........ h )) e :i.) e :rc_~ <_I e m <: 3/ 
!IA c;I\])A, I(6iti 
I:~stitutc Io:c New General;ion Compul;er 'ihchnology (IC()"I;') 
Mil;a. Kokusai \]~ldg. 2:\[\])', \]-4..28 iVt;ii;~, 
Minat, o--ku, TSky6 \]08 JAPAN 
' i;k,J: 18\]-3--4563194 
\]!Y-mM\]: ha.st da%icoi;.jp(C~rela.y, c,w~te g 
A i';',5':~VIA C7' 
A eolnlmtai, ional approach it, etnl>\[oycd here to explicate hu 
man hulgmlge faculty. Some iA~enomena, hwolvin:s mlhom:ded 
depeudency a.rc tim:: i)rovided with cognitive accox:ut based (m 
tit<; processlug load hnl)O:~ed by re\], rant sy::ga,ci, ic operations. 
It: particular, a, considerm;iou on local ,A:ructura,l amlfiguity ac 
counts Ii)t' some island effcet (t;hc noun-coml~lemcnt ca',ms oi" 
Complex NP Consh':dnt) which is c.rre~:Lly unexplained by 
slatic approaches i:~ I, raditioua.\] theories o1 syntax. This ex- 
empllfies Ill: ~t some rule:, of syui:,x tmndlc local a.ml)iguil:y, ~mg 
gestlng insuHh:icncy of traditional al)proadws to to sy,:tax. 
"1- * :, . o LI. .L i~-~ FO OI'J_tC L \]1.0 K~ 
'\]'he purpose of the present article, is to aceounI; for sy:itac 
tic constrai,:ts on some aspects of unborn:deal dependency 
(UBD he~'eafter) phenomeua, by means el a computational 
model. A COml)ul, a.donal model is described in terms o!' for- 
real operations on founal structures. At: eXlAanatien based on 
such a. model is to ascribe the pheuomcnon in question to some 
compul;ational properties of the model; i.e., sud: properties as 
computational cornplexity with regard to both space and time, 
accessib:li~;y to some parts of memory, and so forth. A major 
advantage of this sort; of computational approach is that it can 
mention dymmlic aspects of phenomena, such as i;emporal or- 
der of processing and structural ambiguity a~ising dynamically 
in tit<', course of comprehending or producing utterances. 
Linguistics, by coutrast, has paid littk, a{,tention to Gosc 
aspects and has limited itself t;() investigation of characteristics 
of langttage whid~ could bc tMked about in static terms. "lb say 
that S-structure is derived h'om D-structm:e, for example, does 
not imply that the former chronologically precedes the latter. 
In fact, many lingnistic (especially syntactic) phenomena can 
be understood without referring to ambiguity, processing order, 
etc. This is partly why linguistics has seen its successes. 
Nevertheless, languages haw~ some 1)roper tit's essentially stem- 
mint out of dyn&mic features of la.ngtmge processing (or lu~vb<: 
of a, more gencra.{ cognitive procese;er, after Piaget). Our re&in 
concern hole is with such dynamic aspects o\[ language. In 
what follmw, we shall first touch Ul)OU some cognltive view- 
point appUcd to several types of islaud phenomena, showing 
that some significant part of liuguistic accotlnt is reducible to 
processing terms. Furl;her shall we go on to demonstrate that a 
dynamic approach can elucidate some phenomenon, the noun- 
complement case of the complex ~lou. phrase constraint, which 
is unlikely to be explicable in sta.tic terms of tradit;kmal synta.c- 
\[;ic theories. 
~ f;Lag{c Ao::o: !~?t 
()onsider ~,he (nou) .~;eHtcnces below, which i:tvolv(' UI~/) 
c,<m:~trucdons; the' subserilHs iaud j imlicate co:nd~ xai:io::s. 
(I) Who, \[sod<, \[Npyou\] \[Vl,\[vbelieve \[.s that \[:hshe loves q\]\]\]\]\]? 
(PO ~'Wha.l~{ dld yo. ,-,e<, :,,.~ :.,hl \[spvho \[vp~ate 'AI': 
(3) ':What/ do you wonder wit<) :~t;e ei? 
(4) *Whoi do you wonder whalj / gave Li to ci? 
(5) *Whoi did a story about ~i surprise you? 
The grammatical si.~Lttls of these strings is ut:derstood without 
referring to dynamic terms such as tempocal processing order, 
structural ambiguity, etc. Let us see how. 
2ol ConM;:raing;s aDou~ Dislocated l,;h:.~'mni;,<; 
The syntactic operations of l!;ng\]ish wc will pay attet:tio)l lo in 
the following (Uscusskm are what we migtlt call complenmn- 
ration, specification, adjunetion, binding, and passing, 
each of which Lakes place in a branching local tree. Comple 
mentatlon is to ~ssociate an object with its head. lu (1), l\)r 
instance, a complementa.tion takes place in the local tree comd;i 
tuting of VP, V, and $1; the mother, the head, and the object, 
respectively. Specification attaches a specifier to its head; c.!,;., 
the subject of a sentence 1,o VP (or to lP in the recent traus.. 
formational thcories /Chomsky 1986/). A coucret, e example of 
complemental,ion is the local iree expounding So in (1), where 
the specifier is dm sub.jeer NP. Adjmlcdon associates an a.d 
junct with its head; e.g., an adverh with VP, aud a relative 
clause to NP. BLnding is to bind a dislocated ehm.mn~; (sce 
next paragraph), associated with a syntactic gap, to its an 
tecedent (e.g., a WII-phrase such as ~who' and %1: which day'). 
\]'i)r instance, the dislocated element a.ssociatxxl with ~e,' .gets 
bound by 'Who/' in the top local tree el(It a.bove.. I'assiug;is 
to pass a dislocated clement bel;weeu the inother category all(\] 
some of the daughters. In the local greee×paudiug VP ii:(/), 
the sa.me dislocated elelY/Cil{; is passed bc{.weeu the umthe\[' (VP) 
aud the compleinent daughter (S~). 
'Phus, blinding and passiug are both operal.kms ell dislocated 
elements, l/y a. 'dislocated element', we refer lo a iok(u iH 
mental representation which syntactically con:esl)OUdS to sev 
era\[ posH;ions in a, senLence I 'l'ypica\]ty, there a,re two sucll posi 
tions, the filler and the gap, the' former being ()hen calhx/ the 
antecedent of the latter. For instance, there is a dislocated eke 
meut corresponding both to ~Who? and t.o ~e? in (\[). l)ifferent 
approaches to syntax assume a dislocated element to addition- 
ally corresl)ond to diffel:em, sets of positions ttegweetl the filler 
and the gap. In generM, tra:mlk>rmatioual grammars tend {;o 
postulate fewer such positions than do phrase.struchu:e gram.. 
2~\] 
nars (PSGs, hereafter, which include, among others, GPSG 
'GKPS 1985/and IIPSG/Pollard 1984, 1935/) or LFG/Bres- 
lau 1982/. Also, these theories assign different status to dislo- 
:ated elements; %'ansformational grammarians talk about them 
Ls if they 'move' through sentence structure, while the others 
Lssumes they are simply 'associated with' grammatical cate- 
;ories, via such means as the SLASH feature. 
Such differences among various approaches to syntax, how- 
~ver, is irrelevant to the discussion in the rest of the papaer. 
Ne shall exploit no hypothesis specific to any of these syntactic 
.heories, so that our discussion will be neutral across them, We 
rill borrow some useful terminology and metaphors from spe- 
:ific grammar theories, but that is only for explanatory ease, 
md should not be taken to be any commitment to any of such 
~pproaches. The above description of passing, for instance, 
:cads as if a dislocate element were part of grammatical cate- 
gories, and hence might well remind the reader of PSGs. This 
~y no means implies that we should abandon transformational 
~ccounts in favor of PSGs. Although PSGs are neutral with re- 
spect to the temporal order of processing, incidentally, we shall 
~ometimes talk about passing as the dislocated element going 
¥om the mother to a daughter, reflecting the temporal order of 
~ctual sentence processing; el. transformational gralranarians 
¢alk about movement as if a dislocated element goes into the 
~pposite direction. 
2.2 Explanation 
Now let us resume solving the problem. The distribution of 
grammaticality over (1) through (5) is accounted for by as- 
suming the h)llowing constraints on the syntactic structure of 
English. 
(6) Passing of a dislocated element is permitted only between 
the mother and the head daughter or between the mother 
and the complement daughter. ' 
(7) Passing of a dislocated element and binding of another 
cannot take place simultaneously in one local tree. 
In (1), every passing obeys this constraint. For instance, 
passing the dislocated element bound by 'Who/' into $1 and 
passing it into S, are both O.K., because S~ is the cmnplement 
of 'know' and S~ is the complement of 'that' (whether you might 
employ a transformational account or such theories as GPSG, 
HPSG, and LFG.). 
However, (2) violates (6) and (7) First, the dislocated ele- 
ment bound by 'What/' is passed into $2, which is not a com- 
plement but an adjunct of 'girl'; i.e., a violation of (6) Second, 
the same dislocated element is passed into VP2 where another 
dislocated element gets bound by 'who', ending up with a vi- 
olation of (7). Similarly in (3) and (4), passing and binding 
co-occur at the local trees introducing 'who' and 'whatj', re- 
spectively, violating (7). (5) is blocked by (7), because of the 
passinginto the subject'a story about es'; i.e., the specifier of 
INFL (in the transformational account) or of VP (in theories 
like GPSG, etc.). 
2.3 Cognitive Aspects of Constraints 
The explanations about such phenomena proposed in contem- 
porary linguistic inquiries, especialiy the accounts in terms of 
barriers /Chomsky 1986/, are roughly regarded as formaliza- 
tions of the idea sketched above. This line of reasoning is in turn 
attributed to processing terms, when viewed from the stand- 
point of cognitive science. The background intuition is that 
232 
the extent of processing load imposed by a syntactic operation 
varies from one type of operation to another, and that there 
is an upperbound on the total processing load for constituting 
one branching local tree. On account of this, the reason why 
passing tend to be blocked under the certain sorts of circum 
stances would be that it is an expensive operation a11d thus is 
hard to perform together with other expensive operations such 
as adjunction or binding.. 
The processing load for carrying out various syntactic op- 
erations could be further reduced to more fundamental aspects 
of information processing. Consider, for instance, why specifi- 
cation and adjnnction should be harder than cornplementat~on. 
q'he reason seems to be that complementation is lexical\]y li 
tensed and is head-initial (the current discussion is limited to 
English); i.e., the occurrence of a complement licensed by the 
lexical entry of its head and thus is predicted from the occur- 
rence of the preceding head. On the other hand, the occurrence 
of a specifier or of an adjunct is harder to predict, because the 
former (though lexically licensed) precedes its head and the 
latter is not lexically licensed by the head. Passing and bind- 
ing of dislocated elements are also considered to be non-lexical 
operations, though there are a few exceptions involving, for ex- 
ample, so-called tough adjectives, as indicated by the following 
sentence. 
(8) \[Which violin\]/is \[this sonata\]j \[Apeasy to play ej on q\]? 
The difference between (4) and (8) is that the binding of the 
dislocated element carrying index j is lexically licensed in the 
latter but not in the former. In (8), the binding in effect occurs 
at the local tree expanding AP, where this binding is sanctioned 
by the \]exical entry of 'easy'. This is why (8) is grammatical de 
spite (7); i.e., the lexically licensed binding is not an expensive 
operation, so that it does not play the same role as ordinary 
binding would play in regard of (7). 
Note that the processing load concerning the examples we 
have discussed so far is defined within a single sentence struc- 
ture rather than by taking structural ambiguity into account. 
This is why the traditional syntactic approaches are able to cap- 
ture some of such aspects of language; in principle, properties 
of a single structure can be characterized in static terms. 
3 Dynamic Account 
However, the above static approach cannot by itself explain 
some UBD phenomena, especially so-called the Complex NP 
Constraint (originally termed by Ross /Ross 1967/) observed 
in the examples that follow. 
(9) *Whoi do you believe \[NPl\[mp2the claim\] \[gsthat \[ssshe 
loves q\]\]\]? 
(10) *What/ did you propose a plan \[vP to buy el\]? 
In (9), Sa is the complement of 'claim'. Hence the dislocated 
element, bound by 'Who/' should be permitted to be passed into 
Sa without violating (6). Similarly, VP (or CP, in the traus- 
formational account) in (10) is regarded as the eomplelnent of 
'plan', so that the dislocated element bound by 'What/' should 
be able to pass through, (6) aud (7) being respected. Hence 
the static account in the previous section provides no reason 
why these examples should be ungrammatical. It is considered 
because of essentially the same sort of difficulty that Chomsky 
/Chomsky 1986/ leaves unexplained this type of island effect. 
3ol A Model of Language Processing 
Now then let us turn to dymunic aspects of language processing, 
and consider what kind of syntactic structures a human bearer 
should have tmilt and tentatiwdy maintains when 'that' in (9) 
is encountered. As a basis for this investigatiou, we adopt the 
following postulates about ha,nan language processing. 
(\[1) t*. When ;~ word is encountered, it is attached to stcuc.. 
tures previously built, giving rise to new structures. 
ii;ven when several possible wa¥~ of attachment are 
acknowledged, the processing is sot postponed, but 
a~ many new structures corresponding to those ways 
o:{! attachment are made in parallel. 
b. 'lJ'here is a limitation on the size of the memory for 
.~toring these structures, and thus it is impossible to 
retain all the structures potentially sanctioned. Only 
strtn;tures activated strongly enough can survive the 
competition lbr seats in the limited memory. 
From (ll) plus some minor hypotheses, a general processing 
model %11o'¢% which describes both sentence comprehension 
and generation. This model postulates that. just after any word 
a is encounLered, every maximal structure of the sentence cur, 
rently hypothesized in mind should look at)t)roximately like the 
part enclosed within the curve in (:12). 
(12) j;2 
itere ew;ry branching local tree is assumed to be binary, 
without los~ of generality. AI and ,5' may be identical, and 
the short-term memory contains the information about Ai, Bi 
(1 _< i < d), and A, pins the inEormatin about the configura.. 
lion of these categories relative to each other. Note that, as a 
whole, enough information is thus retained to control the gram-. 
maticality ot the way the foregoing context fits the rest of the 
sentmme; Those categories are the points on which the currently 
hypothesized structure has contacts with the still unknown part 
of the sente0ce. 
Strictly speaking~ the picture shown in (12) should he looked 
upon merely as a first order approximation of the reality. That 
ia, the part of the sentence structnre enclosed in tlte curve 
might contain some variable parts, rather than being totally 
definite. Suppose, \[o;r instance, that a sentence begins wit, h a 
*~oun phrase say "this man'. Tit(; entire tree structure of this 
NP should he completed as sool~ as 'm~,on' is encountered, but 
its grammatical case would not be uniquely determined yet, he- 
cause the senteuce as a whole might turn out to be soinething 
like "l'his n-tP, n, I don't know', rather than "Fhis man is crazy'; 
The initiM NPis accusative in the former sentence, and nonfina- 
tive in the latter. In the following discussion, however, we shall 
merely exploit very rough properties of the model, so that such 
an inaccuracy is considered irrelevant. Readers are referred to 
llasida/liasida 1985/for how this model is ohtained and what 
it predicts, wtfich the limited space of the current article fails 
to accommodate. 
3°2 Explarmtion 
1,et us tm'n t>ack to (9). According to this model, when %hat' is 
encountered while (9) is being comprehended, the right-branching 
structnre covering the string from 'Whoi' through 'claim' has 
been nearly completed ;rod the most active structure around 
'that' should look like (13). Ilere arise two pieces of indepen- 
(13) NP1 
NP2 $3 
the claim (,Oral 1 
that 
dent two way ambiguity, as listed below, concerning how this 
structure might potentially grow. 
(14) a. "i'hat' is a relative t}ronoun. 
h. 'That' is a conjunction. 
(115) a. Ss contains a gap bound by 'What/'. 
h. ~a does not contain a gap bound by 'What/'. 
The combination of (14a,b) and (\]5a,b) gives rise to local struc- 
i,m'al ambiguity across four hypotheses: (l 4a~ \[5a), (14 a&l 5b), 
(Hb&lSa), a,~d ({4b~lSb). 
Since (:16),a~instantiation of (14a&15b),is clearly O.K., 
what we have to show is that out of these ff~nr hypotheses just 
(lda&15b) and (14b&15b) enter t\],e grammar to be ~tcquired. 
(I6) Whoi did you tell ci the I'awt t.hat d he kH('.w q~? 
ltence now let us consider why (I d a&l 5a) a~arl (14 b&lSa) 
are rejected, We pay attention to the beh~tvior of dislocated 
elements, as we did in the stath: approach, q'wo dislocated 
dements ar0 relevant to the grammatical status of (9). The 
first one, which is bound by q;hat', corresponds to the possibility 
(14a). Let us refer to this as a from now on. The other, the one 
bound by 'Whati', is present iff (15a) obtains. We shall calI it 
The status of (14a&15a) is parallel to (1). (7) rules out this 
possibility immediately, because it postulates that the local tree 
expanding S-a accommodates both the binding of cv by 'that' and 
the passing of fl into Sa. As for (14b&15a), however, (7) as it 
is fails to work. 
We need some preparation before tackling why (14b&lSa) 
is rejected. (7) says that two distinct dislocated elements, one 
passed and the other bound, cannot simultaneously take part 
in one local tree. As mentioned above, the cognitive-scientific 
motivation for the constraints (6) and (7) is that the mental 
granmtar does not admit a rule whose execution accompanies 
too severe processing load. On account of this motivation, (7) 
is natnrally generalized simply by taking off the presupposition 
that only an operation in one local tree is talked about. That 
233 
is, we hold: 
(17) The grammar cannot accommodate any rule which ma- 
nipulates two distinct dislocated elements in two different 
manners, binding the one and passing the other, at the 
same time (irrespective of whether or not the two dislo- 
cated elements are processed in the same local tree of the 
same coherent structure of a sentence). 
The essential difference between (7) and (17) is that the latter 
can mention a rule which simultaneously handles several hy- 
pothetical structures of the same sentence; i.e., a rule which 
explicitly deals with a local structural ambiguity. That is, (17) 
does, but (7) does not, reject such a rule if it binds a dislo- 
cated element in a sentence structure and at the same time 
passes another dislocated element in another structure, the two 
structures corresponding to two different hypotheses. 
Now let us return to (14b~;15a). (14a&15a) having been 
ruled out, we are now left with three possibilities: (14a&lhb), 
(14b&15a) and (14b&lhb). The former two give rise to a and 
/5, respectively, and hence these dislocated dements show up 
in the multiple structure representing the disjunction of these 
possibilities; i.e., the structure subsuming in parallel the three 
structnres instantiating those possibilities. According to the 
model introduced above, these two dislocated elements are si- 
multaneously manipulated, one bound and the other passed. 
That is, tile local tree expanding Sa (i.e., the local tree in which 
the binding of c~ and passing of fl are both supposed to take 
place) is built at the same time in all the three possible lines 
of processing. \]'his is understood by comparing (13) with .the 
next state (18). When you go from (13) to (18), the local tree 
expanding Sa is completed. 
(18) NP1 
the claim Comp Ss / 
/ 
t / 
\] that / 
Det I 
the 
According to (17), therefore, the rule of syntax in charge of 
this case must reject the possibility of the existence of either 
or fl; otherwise these two dislocated elements would be ma- 
nipulated (bound and passed) simultaneously here. Now note 
that a is chronologically newer than ft. What psychologists 
call the recency effect, consequently, tells us that a (hence 
(14a&15b)) should sm'vive, defeating fl and thus rendering (9) 
ungrammatical. 
The account of the ungrammaticality of (10) is the same 
except that the potential binder, which is the counterpart of 
'that' in (9), is hidden in 'plan' and thus is not overt here. This 
time a is the dislocated element bound by this binder, and 
the one bound by 'Whati'. 
Note that this explanation concerns language acquisition by 
children, rather than language use by adults. It must concern 
the acquisition stage; otherwise what we have shown would not 
be the ungrammaticality of (9) and (10) but merely the diffi- 
234 
culty of processing them. In fact, the above account does apply 
to language acquisition, because the ambiguity pertaining to 
(14) and (15) occm's every time a structure like (13) is encoun- 
tered, so that its disambiguation can be fixated as a part of the 
grammar of English to be acquired. 
3.3 Rules Handling Local Ambiguity 
Further discussion are in order here about the generalized con- 
straint (17) and its role in the above explanation. First, the 
above discussion postulates that the grammar rules are sensi- 
tive to structural ambiguity such as (14) and (\].5) about (9), 
in the sense that some rules of syntax work on multiple struc- 
tures, and thus are in charge of disambiguation. Here one might 
worry which types of ambiguity are handled by the grammar, 
and which are handled metagrammatically. Not every sort: of 
ambiguity is visible to the grammar, as is demonstrated by the 
following example, which is grammatical. 
(19) Whoi did you tell the man \[~4that Is, she loves ¢/\]\]? 
The local ambiguity arising here appears sirrfilar to that of 
(9). More precisely, a four-way local ambiguity arises at ~that', 
as a combination of two pieces of independent two-way ambi- 
guity, one concerning whether 'that' is a relative pronoun or a 
conjunction, and the other whether or not $4 contains a gap 
bound by 'Who/', just as in (9).: An outstanding difference be- 
tween (9) and (19), however, is that the ambiguity in the latter 
case involves two different hypothetical constituent structures 
that follow. 
(20) \[vp\[vltell \[Nplthe man\]\] \[g that Ss\]\] 
(21) \[Gtell \[NP2\[Nplthe man\] \[g that Ss\]\]\] 
To children learning UBD constructions, both of these con- 
stituent structures should appear ambiguous about whether or 
not S~ contains a gap bound by 'Whoi'. 
This ambiguity, unlike the one in (9), is considered invisible 
to the grammar, presumably because of the above ditference. 
That is, if any single rule were sensitive to this ambiguity, (19) 
should be rendered ungrammatical for the same reason why (9) 
is so, because in (19) a dislocated element would be bound by 
'that' and another dislocated element would be passed into $4 
simultaneously. To make sure that the binding and the pass- 
ing should be simultaneous here, notice that the local tree ex- 
panding S~ is completed simultaneously in the two pairs of hy- 
potheses corresponding respectively to (20) and (21). Itence 
the binding by 'that' and the passing into Ss must take place 
simultaneously. 
Seemingly the reason why the grammar is not sensitive to 
this type of local ambiguity is that the four possibilities are not 
coherent enough, in the sense that they are distributed across 
the two distinct constituent structures as mentioned above. It 
appears that the disambiguation of a local structural ambiguity 
is acquired as a part of the grammar only if the structures (or 
hypotheses) constituting that ambiguity are coherent enough 
with each other. Comparing (19) with (9) and (10), one might 
thus posit the following constraint. 
(22) A local structural ambiguity is handled within the gram- 
mar only if the parallel structures involved therein share 
the same constituent structure. 
As for (9), for instance, besides (13) there could of course 
be several other structural possibilities, but they are simply it-. 
relevant to the acquMtion process discussed above, rather than 
:~ystemat, ica.lly abandoned like (14a&15a) and (14b&15a). We 
would then be able to disregard any interaction across distinct 
constituent structures when considering the competence gram- 
mar. 
There are ~,|; least two more supports to constraint (22). 
l"irst, (22) f)llows from the following more fundamentM con- 
straint. 
(23) There is a severe limit, ou the size of the structure which 
one rule of syntax can refer to at once. 
An arnbigui'i;y within one constituent structure tends to full 
within this limit, because the parallel structures involw'.d therein 
share most i;f the storage with each other. An ambiguity across 
several con:~(.ituenl, structures, howc'ver, wouht more often run 
out of this limit, since, the rate of the shared memory is smldler. 
Note that (2a) claims, after all, nothing more than the lira\]ta- 
t.ion on the complexity of mentally feasible rules. 
Another reason h)r holding (22) is based on how ~stuhborn' 
an ambiguity is. As menl;ioned earlier, patterns like (9) and 
(10) constantly ;~ccompany the local ambiguity like (14) plus 
(15). In contrast, patterns like (19) are often less ambiguous, 
as shown in tile example below. 
(24) Whoi did you tell him \[g that she loves ell? 
In this s.mtence, the possibilit;y of 'that' being a relatiw~ 
pronoun i,s very unplausible, the local ambiguity being greatly 
w.duced; we are left with the aml)iguity of whether or not tit(; 
dislocai, ed element bound by 'Whoi' is coutaiiled in $6. Ill sum 
mary, the ambiguity in (9) and (\]0) is robust, while that in (19) 
is fragile. An ambiguity within a single constituent structure 
tends I;o be robust, tn comparison, an ambiguity encompass.- 
ing several different constituent structnres tends to be fragile, 
because the relationship (as for which is more plausible than 
which, etc.) between those eonstituent structures varies from 
case to case, depending on tim internal details of the relevanl; 
constituents context, and so on. The corresponding relation in 
lhe former tgpc of aml)iguity, oil the ol, her hand, is more con- 
stant. Robu:;t ambiguity is visible to the grammar, while fragile 
one is not; Some rules of syntax handle tile former, while the 
latter is trea{,ed metagrammatically. 
304 Cerman Case 
'\]~he following pair of German examples might fall out of our 
dynamic account on (9) and (10). 
(25) Weni glauben Sic, dab er ei liebl;? 
(26) *Wen/glauben Sie die Behauptung, duff er ei liebt? 
(25) and (2fi) are (.lerman counterparts of (1) and (9), respec- 
tively. Note that the direct translation between English and 
Cerman pre::erves grammaticality across these pairs of exam 
ples. 
Since '\]~ehauptung' ix of the feminine geuder rather than 
neuter, ~daff should not be confused with a relative pronoun, 
namely 'die' or ~der', whose antecedent is 'Behanptuug'; The 
relative proi)oun pronounced the same as 'daft' is 'das', which 
is of I;he neul, er gender and the nominative or accusative case. 
So it appears that in the case of (26) children learning German 
should face no ambiguity like ( 14 ). O u r current ap p roach, which 
is essentially based on local ambiguity, hence seems unable {,o 
account for the grammatical status of (26). 
In reality, however, the above example does not contradicl; 
our approach; i.e., constraint (17). Ambiguity parallel to (14) 
arises in fact, and t, herelbre the ungranimaticality of (26) is 
predicted by (17). To children learning 1.JBI) construction of 
German, (26) accompanies the same sort of local ambiguity as 
(9) does, because they acquire the rough fl'amework of UBD 
construc(.iou before the gender system is properly installed into 
tile morphology of relative pronouns. According to Mills /Mills 
\]986/, when children begin to use rela.tive clauses around the 
age of 3, relative pronouns in the relative constructions they 
make a.w either simply ommited or le~cking iuforulatiou about 
the geunler (and the case, too). \])'or eXaml)le: 
(27) Das \]st {flu l)iiz *ram in Wahle. ist. 
That is a mushroonl REI,.I}I/ON in wood is 
"l'hat is a mushroom which D; tl~ wood.' 
This means that at; the beg\]truing chikh'en canaot distin 
guisha relative pronoun of a gender fl'onx ~mothcr relat,ive pro 
noun of a different gender; e.g, between 'die' and '(lair. Mills 
reports that the use of relative pronouns is equlppe.d with the 
proper system of inflection only a.ftcr the age of '1 is reached. 
The significance of (25) and (26) is not crystal clear, iuci- 
dentally~ if the uonsentence below is also taken into account. 
(28) *Wen gla.uben Sic, daft er schen wollte? 
Who think you that he see-IN\]" want-.PAST 
'Who do you think that he wanted to see?' 
Thus it might be that the ui~gra)runaticality of (26) is a, ttribnted 
to that of (28), without regard to (17). Otherwise the above 
pair of examples (25) and (26) should provide a. furl.her evidence 
supporting (1.7). 
4 f~inal R,emarks 
We have accounted for some island condi(ions by me~ns 
of computational evahtation of relevant syntactic operatk)ns; 
i.e., the e.valul,tion reflected in constraints such ~s (6), (7), 
etc. These constraints are regarded its captured by the e.x. 
\]sting linguistic theories. A generalized const.laint (\] 7) applica. 
ble to dynamic aspects of la.nguage processing, especiaJ\]y local 
structural ambiguity, has been demonstrated to account for the 
noun-complement case of the Complex NP Constrain(., which 
seems hard te elucidate in static terms of traditional apt)roaches 
to syntax. 
One important aspect of our approach enlpJoyed here is the 
hypothesis tha.t some sorl, of local ambiguity is visible to and 
thus handled by the grammar, lI' this hypothesis finally turus 
out true, which we have attempted to demoastrat(.', I.he static 
al)proach pursued so far in the linguistic inquiries must be re 
considered. '.l'hat is, an explanation on tim grammal, icality of 
sentences will have to sometimes take into account sere3 al pos 
sible structures in parallel. 
The explanation of the same sort of island loud\]lion by Mar 
cus/Marcus 1980/is comparable I,o ours in that it also exploits 
local ambiguity, postulating rules handling them. Since Marcus 
pays no attention to what kind of ambiguity is visible and what 
kind is not, howcver, his discussion has n<>thing to sa~y about 
235 
the contrast between (9) and (19). Besides, a radical difference 
between the two approaches is that Marcus exploits a stipula- 
tion called the determinism hypothesis, whereas we ernploy 
a more humble working hypothesis of parallel processing plus 
memory limitation. 
A caution would be worthy of noting here. A success of 
computational explanation does not necessarily support either 
innatis m (~ la N. Chomsky, J. A. Fodor, D. Marr, etc.) or 
eonstruetivism (~ la J. Piaget, etc.). If any part of hmnan in- 
teliigence could be understood to be the outcome of a simple 
optimization for some computation, it should subject to two 
different interpretations: that this part should be a domain- 
specific innate endowment because such a simple optimization 
may well be preprogrammed in the course of evolution, or, con- 
trariwise, that it should be generated after birth by the work 
of the domain-independent general intelligence because such a 
simple optimization could be carried out through maturation 
and internal experiences. Further scrutiny would thus be sim- 
ply needed in order to steer our way either to innatist or con- 
structionist dispositior/. 
Along the line of the present argument, perhaps the first 
point where we could face the choice between these two doc- 
trines is the problem of how much processing load we should 
ascribe to various syntactic operations. The evaluation of com- 
putational load as we have exploited here should vary across 
languages, depending on the relative statuses of syntactic op- 
erations. For instance, the situation must be drasticMly dif- 
ferent between dominantly head-initial languages like English 
and Spanish and head-final languages such as Japanes e and Ko- 
rean (and maybe German, too). Also open to further scrutiny 
is whether the variation is explained by the parameter setting 
approach of transformational theories, or by more general com- 
putational considerations. 

References 

/Bresnan 1982/ Bresnan, J.: (ed.) The Mental Representa- 
tion of Grammatical Relations, MIT Press, Cambridge, 
Massachusetts. 

/Chomsky 1986/ Chomsky, N.: Barriers, MIT Press, Cam- 
bridge, Massachusetts. 

/GKPS 1985/ Gazdar, G., Klein, K., Pullum, G., and Sag, 
I.: Generalized Phrase Structure Grammar, Basil Black- 
well, Oxford. 

/Hasida 1985/ Hasida, K.: Bounded Parallelism: A ~Tteory 
of Linguistic Performance, Doctoral Dissertation, Univer~ 
sity of TSky5. 

/Marcus 1980/ Marcus~ M.P.: A Theory of Syntactic Recog- 
nition for Natural Language, MIT Press, Cambridge, Mas- 
sachusetts. 

/Mills 1986/ Mills, A.E.: The Acquisition of Gender: A Study 
of English and German, in Springer Series in Language 
and Communication, Springer-Verlag, Berlin Heiderberg. 

/Pollard 1984/ Pollard, C.J.: Generalized Phrase-Structure 
Grammar, Head Grammars, and Natural Languages, Doc- 
toral Dissertation, Stanford University, Stanford, Califor- 
nia. 

/Pollard 1985/ Pollard, C.J.: Lecture Notes on tlead-Driven 
Phrase-Structure Grammar'; Center for the Study of Lan- 
guage and Information, Stanford University, Stanford, Cal- 
ifornia. 

/Ross 1967/ Ross, J.P~.: Constraints on Variables in Syntax, 
Doctoral Dissertation, Massachusetts Institute of Tech- 
nology, Cambridge, Massachusetts. 
