An Alternative to l)eep Case for Representing R,elatiomd information 
Nigel WARI) l 
University of Tokyo 
Abstract 
No one has come up with a eompletely satisfactory set 
of deep ca.ses relations (or thematir rehu.ions). The 
underlying reason is that any finite set of case re 
lations C&ll capture only some of the gellera\]izatkms 
desired, l prol)ose taste;u\] a feature Sl)aC¢ rel)resen 
tation of relational inh)rmation, where tim axes are 
such things as degree of responsibility, degree of nc 
tivity, and degree of alreetxuhless The role of a par 
ticipant in au ewmt (:all then be described as a pOiltl 
in tiffs space, allowing more accurate representation 
of relational information. The don,ahL of validity of 
each relevant linguist~ic generalizations corresponds to 
a prototype-centered region in the space. This pro 
posal is easy to implement. 
1. Background 
'Fhere are sew~ral things thai a representation of 
relational inforniation should do, and ease does them 
all, but not w',ry well. 
1.1 Tit(,. Problo.ms with Case 
The continued failure to come up with a satisfac- 
tory set of cases is a symptom of the impossibility 
of fixing a single set of eases that ham all the desired 
properties. 
A system of eases should work for the description 
of more thau a Dw syntactic generalizatious. Yet, 
h)r exaruple, the set of things which can be passive 
subjects is sot the same rus the set of things that can 
be direct objects, and so any definition of patient can 
ae(:oullt for at most oIle of |)nese two. 
A systmn of eases should reflect similarities of 
form. Yet there are many dimensions of similarity, 
and any set of cases will account for only some. This 
call be seen by considering the fact that many prepo 
sitions, for examl)le "with", hal/e meanings which 
span several eases (Tsujii £'. Yamanashi 1985). 
Moreover, similarities of meaning do noi always 
line tip with similarities of meaning. For example, 
whm~ a.ssigning a ease to "w~nd" in "the wind closed 
the door", syllta× suggests agent, as does the seman 
tic feature 'no-covert controller', but the the selnal\]tic 
feature 'liot-allilllate' suggests that wind is an mstru- 
illellt. 
For two languages the problems gel, worse; choos- 
ing a set of ea.ses to capture the generalizations of 
one language tends to obscure the generalizations of 
another. To use another old exanq)le, a definition of 
agent that works well for English will not suffice to 
rule out inanimate subjects in Japanese. 
A representation for relational information should 
be good not only for capturing similarities (general 
izations) but also for precise representation, llere too, 
l thanks Dan aurafsky, atom Edwards, Toshiaki llisada, m~d 
Mitaul)iahi lleavy lnduntries. 
ctLse often comes/i l) short. For example, in both "load 
the wagou wzth hay" and "load hay onto the wagon", 
"wagon" is traditioaally *u~signed the same ease, but 
this obscures the difference that in the first sentence 
the wagon is more a\[fi)cted it; is more likely to be 
fully loaded. In general, the goal of precise repre 
,';entatioit suggests lltally specific cases with llarrow 
ill(!anillgs, I)lll t}le goal (if eapt/lrhtg generalizations 
SllggeSIs broader c~kse,~. 
1.2 Tim Stat(~ (ff the Art 
'\[b smnmm'ize, the i)roblem with case is not that 
'we havell'l l()ltlltl tile right set of ca.se8 yet' bill thai 
it is impossibh~ to find a set of ca-ses which does ev 
erything. The goals of representing wu'ious types of 
similarity conllict with each other, and these goals 
conflir.l, with the goal of being able to precisely rep 
reseilt relat, iona\] intbrlnat.ion. While ther(! ill'(! refille 
meats which help somewhat (sill) cases ~dlow more 
precision, and multiple inherit~mce from mq)ercases 
increases the Utlluber of generalizations capturable) 
the problems remain. (l"or further discussion of past 
work on i:ase see Ward (to ~q)pear) mid the references 
cited therein.) 
Of course it is always possible to Col)e to uud¢c 
do with a set of cases which satislles only sonic of the 
desiderata. For one thing, it is possible make do with 
limited expressiw! power, i,'or example, iii;I,Ity tn~> 
chine translation researchers appear s~ttisfied if their 
c~se systeln is just detailed enough to account for 
choices among target language prepc~sitions. It is also 
possible to make do with ~. system of (:ames that misses 
geners./izatioils, l)esigl|ers of Iilaellil|e translatiolt sys 
Leuls, agaiu, lU'eStllllably luake rough trade offs as to 
the relative wdue of sinq)iifying the parser (by choos 
illg a set of c;mcs eo\[iv(}lliellt for tile source I~mguage) 
or simplifying the generator (by choosing the eases to 
suit the ~argel. hmguage). 
Tile probhmis with a system of cases are not al 
ways identified as such. No one ham ever written a 
paper saying 'I can't make case work for my applies 
Lion' shortcoutings cau always be COUlpetlsated Rir 
by t:onq)lie~ting the rules that refl~r to cases. 'i'hat is, 
ally I)roposal for a set, of c~mes is un\[~dsifiabl(! . .. but 
it is possible to do betler. 
2. Proposal 
2.1 Participatory Protiles 
i propose to represent in detail the 'participatory 
properties' of objects. For ex~Lrnl)h b in the scene in 
volving Jn(l~Ls, Jesus, and a kiting, Judas can be 
described as actiwb volitional, very responsible, ba 
sically uaafl'ected, a direct-cause, and so ()It. I will 
refer to the set of these properties ~m the 'participa-. 
tory profile' of that object. A participatory profile is 
il|lplelnetlted its it vector O\]' rallies over '(:;~,~e t(-:attlres'. 
ACRES DE COTING-92, NANTES, 23-28 Aot~Jr 1992 1 1 3 7 Pitoc. OF COLING-92, NANTES, AUG. 23-28, 1992 
kiss 
Judas: agent 
Jesus: patient 
Figure 1 : 
A traditional representation 
kiss 
Judas: 
active .7 
volitional .7 
responsible .7 
affected -.2 
direct-cause .7 
Jesus: 
active -.7 
volitional -.7 
responsible -.2 
affected .5 
direct-cause .2 
Figure 2: A 'participatory profile' representation 
>, aE 
2~ 
) " Jotm ! 
hn kissed ,' ,Iohn • stared at Mary. killed 
;' himself Mar;. / 
/ /' 
/./ John e 
• ,; got hinlself 
John " killed 
accidentally ,.'"'" • John kicked " seduced .." .....'Y'" 
Mur~,. ..'"'"" eaail3 . 
............. ..... .°'"" John • 
. "'" gOl 
J~)hn "'"-, %'. .,/ killed. 
noticed ', 
Mary "", ;: l 
", •John ~htt 
", '; sneezed , was made i John • 
'. to kiss was ' i killed. ~: 
the u on 
i 
Affectedness 
Figure 3: A slice of case space, chosen to focus (in causal/inchoative erects 
For example, Judas a.s a kisser may be (-.2 affected), 
(+.7 responsible), and so or,, as shown in Figure 2 and 
a.s contra.sted to the traditional representation shown 
in Figure 1. 
A participatory profile is a precise description. 
To illustrate this with a spatial metaphor, a par,it 
ipatory profile can be identified with a point in an 
n-dimensional space, which I will call 'c~me space', 
where the axes are the ea.se features. Figure 3 shows 
an impressionistic projection of this space onto two 
dimensions, populated with sentences about John, 
positioned appropriately for his role in them. Super 
imposed on this with curved lines is a suggestion of 
the way that a traditional ease account might divide 
up this space. This illustrates how case allows only 
a relatively coarse description, providing only the op- 
portunil, y to describe a participant's role a.s being in 
a certain region of the space. 
This proposal Mso makes it easy to explain simi 
larities. For e×ample, comparing the roles of "yeasl" 
in "yeast makes bread r~se" and "spoon" ill "eat with 
a ,spoon", they are similar in that both are concrete 
and directly acting, but different it, that the yea.st 
is not manipulable, nor is it identifiable as a sep~- 
rate entity afterwards. Profile representations of the 
roles of yea.st and spoor, can show that they are sim- 
ilar on specific stlared dimensions, while not obscur 
in K the differences on other dimensions. Profile rep- 
resentations also make it easy to quantitatiwdy de- 
scribe similarity on a single dimension. For example, 
it is possible to describe John as active in both "John 
.~peculated t~i commodities" and in "John watched the 
ducks", but somewhato less active in the latter; there 
is no forced choice between assigning John to a case 
where he is active and one where im is not. 
(-'.~se is traditionally considered to be a cla.ssific~ 
Lion of tile semantic relations between predicates and 
their arguments, but the proposal replaces it with 
an account of the roles of participants in events. In 
some languages things like. individuation or definit, e 
ness, which would seem to have nothing to do with the 
verb, affect choice of ease markers and constructions 
(Fillmore 1968; flopper gz Thompson 1980). q'hus it 
seems that meaning relations should relate to the sit 
ua,lion, Ilot just to the predicate. (Here 'situation' 
is meant in a narrow sense (DeLancey 1991), where 
"John asked Mar# to leave" involw~s two situations.) 
2.2 Profiles and Language 
Iamguage refers to regions of case space. 'Fhis is 
true, in particular, of 'case markers', constructions, 
aiid grammatical roles. 
Consider for example the Nmily of uses of "of" 
exemplified in ".lohn died of cancer". "Of" is used 
for causes which are direct causes, invisible, iInrna- 
terial, of unknown origin, and at most only slightly 
con,toned (Del,ancey 1984). If direct-cause, visible, 
and so on are treated a.s ease features, this use of "of" 
can he described as appropriate for participants in a 
certain region of ease space. Ill generM, the meanings 
of 'ease markers', that is, words conveying relational 
information, caa, be identified wit}, regions of ease 
ACRES DE COLING-92, NANTES. 23-28 AO~q' 1992 1 1 3 8 P~.OC. OF COLING-92, N^NTES, AUG. 23-28. 1992 
space 
The tueallings of sonic constructions also can he 
identified with regions. ('Construction' here is meam 
in the sense of l"ilhnore, Kay, and (')'('.onnor (1988).) 
For a given participant, the extent to which its profih~ 
leads to selection of function words or to mobilization( 
of constructions (affecting word order), or to bulb, 
depends entirely on the language. 
Regions in ca.se space can also he used to describe 
grammatical roles. For examl)le, consider lhe set of 
things which can he subjects of passiw: seld;en(:es. 
Rather than saying tha.t this includes Iheme.% l)a 
tients, and recipients, provided they meel certahl con 
ditions, we can describe this am the set of things which 
are highly topicalized, not very active, alld Inure or 
less aft'coted; this of course describes a region of (m.se 
space. The set of things which (:~m he direct objects 
is another region, ow~rlapping that for t)assive st;b- 
jeets, but also including the region of highly all'cried 
things even if they ~u:e not at all topics, and exchnling 
all highly tol)icalized things, and also mildly topical- 
ized things unless they are highly atfected. The set 
of things thai can be tnmsive subjects in Japanese 
is yet. another region, again overlapping but slightly 
different. 
'lb summtwize the ways in which this propos;d 
solves the problems raised in Section 1: it allows pre 
else representation heca/lse instances are rel)resented 
a.s points, ;rod this does not conflict with the need t(/ 
capture generalizations, because generMiz;ttlons art: 
represented as regions; and it can capture all gen 
eralizations because there is no assumption of corre- 
spondence bet, wren the regions required for different 
generalizations. 
2.3 Exalnl)les and Details 
'.Fo define the regions for various case markers by 
precisely specifying their boundaries would he oner 
ous at best. Instead we can define these inlplicit/y 
by reference to their prototypica\[ meanings. For ex 
ample, the prototypical use of "of" in "die of cati- 
cer" carl be described as ~ point in case space, lty 
computing the proxinfity of a participator's profiles to 
such prototypes for various case markers it is possible 
to determine the m(xst suitable case marker for tha~ 
participant. 
Similarly for constructions; they are used when 
a participant's profile is sufficiently close to tile con 
struction's prototype. (Polysemous constructions can 
probably be amdyzed as having several prototypes.) 
For example, one can analyze the Passive Construe 
tion a.s being relevant if a participant expressed in 
subject position has a profile is 'closer than 1.2' to 
the prototype (affected +1., volitiom:d -1., responsi 
hie -1.), as shown in Figure 4. 
Unlike prepositions, constructions' meanings do 
not form a partition of case space; thus a single point 
can fall into the regions of several constructions. It 
is son\]etlnles necessary to elllploy inure than one con- 
struction to adequately specify the profile of a partic- 
ipmlt. For example, to describe a participant who is 
active and possibly affected, but not responsible nor 
directly affected, the Passive and Causative Construe 
'lhe t'~asive Coils(ruction 
example: "Ma~y wa~, given a fork" 
('r)l;dd.ioIi (()r i'(deva\[;(:e: 
expression of a participant closer I.halL 1.2 
to tile l)rototype, using the we.ights below 
........ S°t~?t'Y P(L__weights 
affected -t I. 1 
volitional • I. .5 
respolLsihle - l. .5 
'lhe I)eriphr~usti(: Causative (kmstruction 
example: "John nlade Mary go lo C'hzcayo" 
condition for relewtnce: expression of a participant 
closer than 3.5 to the prototype below 
protolype weights 
volitional FI 1. 
resp(msibh~ -- 1 1. 
~tctive + I 1. 
affected -- \[ 1. 
direct i:;tuse - I \[. 
The State Ch~mge Construction 
ex~mqde: ",lohn died" 
comment: riwd lo the Passive (k)nstruction; 
prevellts ".\]ohr~ irla.s died" 
coIIdil, ion for relevaucv: expressloa of a \[)articil)alll 
closer than 2.5 to the prototype below; 
also the availability of a st~tte change-verb 
prototyI)e weights 
~fl'e(:ted \] 1 l. 
voliti(mal - 1 l. 
responsible -- 1 t. 
object-of force 1 1. 
Figure el: ~olne ronstrllclir)ns whose relevi£nce 
depends on profiles 
tiolls lntlst be rise(1 together, a. ~, in a John rllas made 
lo k~.~s lhe .stalur "; each constructio, expressing SOil|{! 
dimensions of the participant's profile. The ide~t of 
;tdditive colltribntiolts {ronl several constructions call 
also be applied Io, for example, "John was k~ssed", 
where "doh¢~" is a perfl~ctly good subject, and also a 
perfectly good passiw~ s~duect. This style of aualy 
sis means factoring out infbrmation, which of course 
makes \['or siIHI)~e (OllSIrllctio;ls. 
\[ise of COllSl.rtlctions provides a way to acc(Hln\[ 
for the 'subcaLegorization' properties of w~rbs. To 
explain why "John broke the dish" is English but 
"*lhe magwm~ vanished lhe rabbit" is not, one can 
say that the verb %real" can participate m the Lex 
ical Causative ('onstructkm but "vanish" can only 
participate in the Periphra.stic Causative Construe 
tion. Thus it is not necessary to directly describe the 
allowable cases of a verb and their mappings to prep() 
sitions ~md grammatical roles; that information can 
he factored Oil( into constr:lctions. That is~ the case 
frame (valence) of a verb ca1 be explained in terms 
of the constructions the verb can participate in. 
(}ralnmatical roles cart ~lso be analyzed in ternLs 
AC*rES DI.: COLING-92, NANTES, 23-28 Aotrr 1992 1 1 3 9 PROC. OF COLING-92, NANTES, AU(}. 23-28, 1992 
feature prototype it)cation weights 
topic +1 .6 
volitional +1 .4 
active +1 .4 
responsible +1 .2 
individuated +1 .2 
par tial-cause +1 .l 
affected - 1 .2 
Figure 5: The first constituent of the 
Subject-Predicate Construction 
of prototypes - for example it haw long been said 
that the prot.otypical direct object is probably that 
of "kill" -- and these prototypes can be mapped into 
case space. Proximity to prototypes can then be com- 
puted. This allows, for example, the simple rule: 'for 
subject, select tire partieipam which is closest in case 
space to the prototypical subject' (to slightly modify 
a proposal by 1)nwty (1991)). As sonic fitctors are 
more important than others, it is appropriate to as- 
sign weights to the various case features, to bias the 
computation of proximity. For example, the weights 
for subject shown in Figure 5 account for subject se- 
lection (in the context of the system described in Sec- 
tion 4), explaining: 
la) John kissed Mary 
lb) Mary made the boy eat a peach 
lc) Mary was kissed by John (if she is the topic in the 
larger context) 
ld) the wind broke a dtsh 
le) Mary was killed and Mary died 
This account of subject is more parsimonious than 
a subject hierarchy, that is, a list of cases in order 
of preference for which can become the subject (Fill- 
more 1968), plus rules for overriding it for tire sake 
of topics. This description also ohviat, es the need for 
explicit statements that topicness is more important 
than agentivity or that volition is more important 
than activity; such facts are simply encoded in tire 
weights. 
In the current implementation of ease space, the 
range of values for each feature go front -1 to +1. 
Whereas participants can be located at any point in 
the space, it seems appropriate to site prototypes at 
the corners or edges of the space. A few more exan|- 
pies of profiles are shown in Figure 4, arrd many more 
in Ward (to appear). 
3. Related Work 
Although the synthesis is novel, many of the ma- 
jor components of the proposal have been previ- 
ously proposed, if in somewhat different guises and 
for different purposes For exaanple, Cruse (1973) 
and Delaneey (1984) studied the components of var- 
ious meeming relations, Labov (1973) and Miikku- 
tainen and Dyer (1991) pioneered the use of vector 
spaces for describing meaning, Hopper and Thomp 
son (1980) showed how to relate grammatical reflexes 
to lists of scalar-valued paxameters (features), Ilinton 
(1981) noted the possibility of using a 'distributed 
representation of roles', Tsujii and Yamanashi (1985) 
viewed cases in terms of prototypes and their exten 
sions, Fukuda et al (1986) and Pederson (1991) in- 
troduced the spatial metaphor for meaning relations, 
and Dowty (199t) explained how to relate grmnmati- 
ca\[ relations to prntotype structured clusters of mean- 
ing relations. 
4. Implementation 
1 have built a parser (Ward 1992) and a gener- 
alor (Ward to appear) which use participatory pro- 
files. This section discusses the generator, not as a 
presentation of the best or only way to use profiles, 
but merely as a demoastration that case profiles are 
workable. 
FIG, a 'Flexible Incremental Generator', produces 
English and Japanese sentences starting from a mean- 
ing representation, using spreading activation in a 
knowledge net, work. One task of a generator is, given 
an input including some items with case profiles, to 
build a sentence whose syntactic form and function 
words reflect those e~e profiles. 
In FIG case features are implemented ms nodes 
in the associative network. They are linked to con- 
structions and words, with appropriate weights. For 
example, the node responsible,, has a link to the 
node by,,, representing the word "by" , and this link 
fias weight +1. 
The participatory profiles of concepts in the inpnt 
are represented by links to nodes for case fi~atures, 
appropriately weighted. For example, the node for 
Mary may have a link with weight .5 to responsible, 
to represent a given mput. 
For such an input, when mary,, becomes acti- 
vated, case features will become activated to the de 
gree appropriate for her profile. In tnrn byw and 
other prepositions will receive actiw~tion from these 
case features. The net effect is that the profile for 
a participant activates prepositions proportionally to 
their proximity m case space to that profile. (The 
measure of proximity computed is, to be precise, the 
dot product of the vector for the participant and the 
vector for the prototype.) The preposition whose pro 
totype is closest will receive the most activatiou, mid 
hence appear in the output. Like ease markers, colt- 
structknls receive activation from the profiles of par- 
ticipants, via case features. They thus become mobi 
lized to the extent that there is a participant with a 
profile matdring that of the construction. (Some case 
markers appear before the word they flag, others af- 
ter, and so FIG has a distinction between activation 
fi'oni the profiles of concepts which remain to be ex- 
pressed and activation from the profile of the concept 
just expressed.) 
Constituents which involve profiles also are linked 
to nodes for case features; from these activation flows 
to concepts, and so the concept whose participatory 
profile is closest to that activated by a constituent will 
receive the most activation. (Actually the case fea- 
ture nodes used for activation flow from constituents 
to concepts are distinct from those used for activation 
flow fronl concelpts to ca.se markers and constructions. 
That is, each case feature is implemented a.s a pair of 
nodes; this is for technical reasons.) There are multi- 
pie profiles in any non-trivial conceptualization, and 
ACRES DE COL!NG-92, NANTES, 23-28 nOLq" 1992 1 1 4 0 PROC. OF COLING-92, NANTES, AUG. 23-28, 1992 
it wouht seem that crosstalk aright be a problem, but 
this has not been the case in F\[G, primarily hecause 
generally there is one eoltsLruction an(l one (:oncepi 
with enough actiwttion to dominate. 
FI('~ originally expected deep (:as(: relations in its 
int)uts , and its grmnrnar and lexicon referred to those 
cases. One problem was that, ~s I extended 1,'lG's 
coverage of the two langllages, the nnlnher O\[ cil~ses 
kept growing and the grammar got uglier and uglier. 
In t)articular, there were lengthening lists of possible 
cases for constituents, for example there was at llst of 
fonr possible cruses to nse for subject. Switching to 
proliles solved these problems. Conversion wa.s rela- 
tively easy; other than the new references to profiles, 
the grammar did not need to be changed. Tim addi- 
tional eomi)utation required is negligible. 
FIG currently uses 10 (:~use fi~atures: volitional, re- 
sponsible, active, aflheted, direet-cause, partial cause, 
individuated, topic, object of-force, and touched; 
these replace the e~Lses agent, instrument, patient, 
experieneer, cause and percept. At this point the 
meanings of the cause features derive less from their 
flames than from the way they are related to the con 
structions of Jat)anese and English. This is hecause 
tbe numeric values for the t)rofiles, although originally 
chosen according to comlrlon sense arid with reference 
to the literature, tlaw~ had to i)e tuned in the course 
of making FIG able to generate sentences in both Inn- 
gnages for a largish mmlber of inputs. I ascribe no 
special significance to the particular set of profiles 
enrrentty in nse: they are specific to FIG's current 
gramtnar and implementation details. 
5. Summary and Hopes 
To summarize the advantages of the prof)osal: 
Participatory profles are a representation mech~v- 
nism that allows both precision and generalization- 
capturing. Precision is important for being able to 
tel)resent accurately the information that people (:an 
get from langnage, and it will probably also he use 
ful for artificial intelligence systems in the near fu 
lure. Better generalization-capturing allows siml)ler 
amt better grammars. Tbis is important for linguis 
tics, and also for comtmtational linguistic, where the 
cash vahle will be inq)roved manageability and per- 
formance for natural language systems. One example 
is machine translation. If the parser/mMerstander 
arrives at a narrow enough eaqe profile for a partici- 
pant, then it is lmssible to directly find the relewmt 
target language constructions t)y cornputmg in which 
regions the point lies. It shouhl thus he posslhle to 
eliminate the need for contraative knowledge relating 
the regions of the various constrnctions and words of 
the two languages. 
Judging from my experience converting FIG to 
profiles, these advantages may be easy to achieve in 
practice. Of course, to conre u t) with a general theory 
of relational information will require a great deal more 
work, both on the mechanism and on the analysis of 
language. 
This f)ropc,sM is in some ways a logical contin- 
uation of Fillrrtore's (19681 research i)rogram. Fill- 
more wanted to capture linguistic generalizations in 
terms of meaning, not syntactic structures, hi Ward 
(forthcoming) I suggest that a processing model can 
dispense with surface syntax struetures also; doing 
without csse eliminates yet another type of interme 
diate structure typically interposed between thought 
aml hmguage, allowing an even more direct account 
of linguistic generalizations in terms of meaning. 
References 
Cruse, 11. A. (1973). Sore(: Thoughts on Agentivlty. Jour- 
nal o\[ Li):guiMics, 9: I 1 23. 
Del,ancey, Scott (1984). Notes ()It Agentivity and Causa- 
tion. Studies m Lauguagc, 8121:181 213. 
DeLaacey, Scott 119911. |",vent Construal and Case l/.oJe 
Assignment. In Berkeley Linguistics Society, Pro- 
ceediugs o\] the Seventeenth Annual Meeting. 
/)owty, David R. (1991). Thematic Proto-Roles and Ar- 
gument Selection. Lauguagc, 67:547 619. 
\["ilhnore, Charles J. (1968). The Case lot Case. In E. Bach 
& It. Harms, editors. Urlivcraals in L~nguistic The- 
ory, f)P. 1 88. Holt. Rinehart, New York. 
Fillmore, Charles J., Paul Kay, & M. C. O'Connor (1988). 
ilegularity and Miomaticity in Grammatical Con- 
structions: q?he Cruse of Let Aloim. Lan9ua9c, 64(3). 
Fukudlt, I(ttlute, .\[lllt Y~unltguchi, Jun ichi Tsu.iii, &~ 
M~Lsaaki Yamana.~hi 11986). Kakn Kaishaku to 
Ninchi Kikoo, Sono 2 (Case Interpretation and 
Cognitive Structure, Part 2). In Proceedings 3rd 
daparlcse Coguititw Science Society, p. 66. 
Hinton, Geoffrey E. (1981). \]lnplementing Semantic Net- 
works in Parallel Hardware. In Geoffrey. E, Hinton 
& J. A. Anderson, editors, l)a)nlle.l Models o\] A,q- 
~ociatioe Memory, pp. 161-188. Lawrence Erlbaum 
t\~sociate~. 
lh)l)per, Paul J. & Sandra A. Thompson 11980). Transi- 
tivity in Grammar and Discourse. Language, 56:251 
299. 
Lahov, William (197:11. The Boundaries of Words and 
their Meanings. In Joshua Fishman, editor, New 
Ways ol Auolyzin9 Variation in English, pp. 340 
373. Georgetown University Press. 
Miikkul~tinen, t/isto & Mich aeI Dyer (1991). Natural Lan- 
guage Processing with Modular PDP Networks and 
Distributed f,exicon. Cofnitive Science, 15:343 400. 
Pederson, Eric (19911. The Ecology of a Semantic Space. 
In Berkeley Linguistics Society, Proceedings o\] the: 
Seoenteeulh Annual Meeting. 
T~,ujii, Jun-ichi & Masaaki Yamanashi (19851. Kaku to 
sono Nintei Kijun (Ca~es and Criteria for their Iden- 
tification). 'I~chnical Report 52-3. Information Pro- 
cessing Society of Japan, Natural Language Working 
Group, Tokyo. 
Ward, Nigd (1992). Au Evidential Model of Syntax 
for Understanding. Technical Report 88-3, Informa- 
tion Processing Society of Japan, NaturM Language 
Working Group, Tokyo 
Ward, Nigel (forthcoming), A Parallel Approach to Syn- 
tax h)r Generation. Artificial h~telligencc. 
Ward, Nigel (to appear). A Connectionist Language Gen- 
erator. Ablex. revised and extended version of A 
Flexil)le, Parallel Model of Natural Language Gen- 
eration, Ph D. thems and Technical Report UCB 
CSD 91/629, Contputer Science Division, University 
of California at Berkeley. 
ACIES DE COLING-92, NANTES, 23-28 Ao(zr 1992 1 1 4 1 PRO{:. OF COLING-92, NANTES, AUG. 23-28, 1992 
