The need for MT-oriented versions of Case and Valency in MT 
Harold L. Somers 
Centre for Computational Linguistics 
University of Manchester Institute of Science & Technology 
ABSTRACT 
This paper looks at the use in machine Translation 
systems of the linguistic models of Case and Valency. 
It is argued that neither of these models was 
originally developed with this use in mind, and both 
must be adapted somewhat to meet this purpose. In 
particular, the traditional Valency distinction of 
complements and adjuncts leads to conflicts when 
valency frames in different languages are compared: a 
finer but more flexible distinction is required. 
Also, these concepts must be extended beyond the 
verb, to include the noun and adjective as valency 
bearers. As far as Case is concerned, too narrow an 
approach has traditionally been taken: work in this 
field has been too conerned only with cases for 
arguments in verb frames; case label systems for non- 
valency bound elements and also for elements Jn 
nominal groups must be elaborated. The paper 
suggests an integrated approach specifically oriented 
towards the particular problems found in MT. 
\]. Introduction 
Most (though not all) MT systems claim to incorporate 
versions of Valency grammar, and more recently have 
also looked to Case grammar. However, whatever 
theory they use is often imported more or less 
directly, without taking account of the fact that a 
model developed for one purpose may not be entirely 
appropriate for another. This is a less serious 
problem for Valency, though this was originally 
designed with a monolingual didactic purpose in mind 
(see Helbig & Schenkel, 1973:5) rather than the 
multiling~al needs of translation. With Case 
however, it is often the much-maligned and quickly 
supereeded Fillmore (1968) model which is adopted, or 
at best a 'consensus' model like that of Fillmore 
(1971) or Chafe (1970), loosely extended. What is 
not taken into account is the fact that these models 
typically concentrate on 'nuclear' arguments in verb- 
phrases, saying little about 'peripheral' roles, or 
about the structure of nominal or even adjectival 
groups. This paper will show need for a more task- 
specific model, combining Valency and Case into an 
integrated theory for the purposes of translation. 
More specifically, we will show (a) the need for a 
more flexible Valency system with six degrees of 
valency-binding instead of the usual two; (b) the 
need for a finely tuned version of Case to fit the 
new version of Valency proposed: in particular what 
depth of abstraction is appropriate; and (c) the need 
for this combined Case and Valency model to extend 
beyond verbs, especially to nominal groups. 
2. Valency in existing MT systems 
The essential notion borrowed from Valency theory 
found in MT is the distinction between 'complements' 
and 'adjuncts'. In several MT systems we find that 
the lexicon contains information equivalent to that 
given in a valency dictionary like that of Helbig & 
Schenkel (1973) listing the complements predicted by 
the verb together with associated syntactic and 
semantic features. This lexical information is used 
in a straightforward manner to assist in the 
computation of structured representations for the 
source text, with the complements and adjuncts 
labelled appropriately. In this way for example, the 
functions of the prepositions in sentences like (la) 
and (2a) can be differentiated, and the correct 
translations (ib-2b) arrived at. 
la) Er wartet auf seinem Freund. 
\]b) He is waiting for his friend. 
2a) Er wafter auf dem Bahnhof. 
2b) He is waiting at the station. 
The identification of complements is useful in the 
bilingual transfer stage (or equivalent) of the MT 
system, and it is appropriate at this point to 
provide a set of example translation pairs (3)-(5) 
that illustrate an important problem that Valency can 
help with. 
(3a) He likes the girl. 
(3b) La fille lui plait. 
4a) The farmer supplied the grocer with milk. 
4b) Le fermier fournit du lait a l'$picier. 
5a) CharIes entered the'restaurant. 
(5b) Charles entra darts le restaurant. 
Each of the sentence-pairs in (3)-(5) illustrates a 
change in complement structure between English and 
French. The example with like and plaire is 
something of a classic, involving a double argument 
change; but examples like (4) and (5), which, while 
less extreme, are nevertheless typical, suggest that 
the phenomenon is widespread. 
The Valency approach to this problem recognises each 
of the three verb pairs in (3)-(5) as having the same 
valency, but as having different valency patterns, 
which must be mapped onto each other. 
In the case of Rothkegel (1976), this mapping is 
achieved by simpIy juxtaposing the two entries, with 
equivalent arguments next to each other, roughly as 
in Figure i. 
supply Ns (animate) No (animate) Np (with,physobj) 
fournir Ns (animate) Ni (animate) No (physobj) 
Figure i. Valency patterns in Rothkegel (1976) 
Alternatively, in GETA's ARIANE-78 (Vauquois, 1978) 
and TAUM-Aviation (Lehrberger, 1981), the complements 
are assigned distinctive labels: in both systems the 
label ARG(ument) is used with a distinctive affixed 
numeral, roughly as in (6). 
(6a) The farmer supplied the grocer with milk. 
ARGI PRED ARG2 ARG3 
118 
(6h) Lc'fe:l'mier £ournig du \]el:it ~'c \] '~p:ic:ier. 
AllGJ PRF, I) A17C,2 AR(}3 
or ARG\] PlllT;I) AIIG3 AltG2 
Notice here tile \[;we possJhle, uses of ~\]lOS(! lllllllberod 
ARG \].abels: either riley are assigned mope or \].oss 
arbitrarily, and an ARG-for-ARG mapp:ing defined for 
tile particular verb pair; or the AIIG \] at)el s 
correspond to SOIlle aspect of the alla\]ysis (typically 
deep syntactic function). In this lattsr approach, 
there are several advantages. F:\[rst , J.t :i.s possih\]e 
tO assign AIKI numbering Ja a non-arhJtrary manner: 
tile ARG \].abels therefore take e~ SOIIIO sig:nif:icance 
which can 5.n fact be used for other tasks too. 
SOCOlld~ eric, may assume a default inapp:i.ng of like- 
numbered AIIG~I onto each other~ and this :indeed works 
par gia:l.\] y or comp\].et ely for a naniher of verbs. 
However, til(!re renlaJ.ns tile disadvantage when :it does 
nob, whore we need a specific mapp:i.nK a\].go:p:\[thm. We 
shall rcturri to i;hfis questfion in tile next section, 
because the introduction o£ Case notions gives us a 
thfird pass J hi:l.:l ty ~ with non-arbiLrary label s but 
withotlt special, mapping algorithms either. 
Let us consJ.dsr now how the dJ stinction of 
COlnplements and adjuncts is used ~n MT. Agafin we can 
look to ARIANE-'/8 and TAUM-.Av:Lation for 
exenlp\]:if:\[eatJoll, for eruc:i.al in theme systems J.S tl\]e 
ARG-C\]:RC (umsl:ant ia\]. ) dJ st J.l\]C t J.oi\] . NonlJ na:l s that 
cannot be matched agains~ the wtlency pattern of the 
verh are marked as CIRC, arid ttl:is distinction gives 
rise to dJfi'erent transfer strategies. The key area 
here Js the translation of preposit:1.ona\] phrases: the 
theory iS i;hat tl:le prepositfion ill a valency.-Imtmd 
element (ARG) is semantically void, and governed by 
the verh. P r'epo,~l J t :Leas Jn adjuncts (eli{C) (o.g, 
\]ljader t hc~alrle vs. _on Lhe .i;able.) on the other h}tnd 
are nlean:i.ngfl,\].~ and SO participate Jn the trails\]alien 
process. 
ThJ S neat cortes i)ondc~nce betweon AIIG and f'orllia\] 
i\]roi)esJ.tJ.on versus C:\[I{C and fallctJolla\] preDositJoil Js 
very conven:ient, theugh unfortunate:ly :Lt does nst 
fully nlatch thc facts. On the enc hand, we have 
valency-hound dJ reel:} onal elements with verbs of 
lnovement, where the preposJtien :i.s meaningful ($9_ 
}o/jn/(A~ 2 etc. ). :l:ntcresttng\] y, AI{\]:ANI,\]-'/8 treal;.~ 
these am 'vs\]ellcy hollnd cJ~ctlll\]stantJals ~ , perhat)s 
after the fashion of Tesn:i.~n'e (1959:128) and this 
makes sentence pairs \]:ike (5) proh\]eulatic, since Jri 
Eng\] J sh ~110 reptau~at~t Js All(}, but :in l"pench d a.l}s 
le restaurant CIRC. Ori I;he ol;hez" haIld we de ilot 
always have :\[soiilorphy hetwecn a ~rdveIl c:\[rcuElstaiiLi~l\] 
relationship and the proposit;Jon tllat expresses :it: 
for exalnp\]e, ':local:ten at' iS expressed by any of })lj., 
at, :iJ!, :i.n English (73), depending on tl~e nature oP 
the location, while ill \]~'rellch the dependence J.s gJvcl\] 
by qnite ddf£erent factors (7h): 
(7a) :in -. cities, countries, ei/c-\[osed spaces 
on -" J s:l.~Ine\[s i fflOllntaill\[\] > st:Peels 
at -, bui\]d:hip;s 
(7b) _~ i>a~:i.s / .en_AyJ~n~)n (phonet±c) 
pn 1,'12{l\[lce / }tit J~p()l{ (gender) 
pn. Co:\[.'st. / .17u\]\] .;\[LI\].c d'I\[,}.!~9. (po:l.J t:i ea\] status) 
\]o rfhe seed £or six deg:Peos of valency-d):ind:i.nl: 
:l:n Sonmrs (:1.984) :1: d:iscuss at length the question of 
whether the conlp\]ement-ad.iunct distJnct:ien is sJ K£1,v 
\]\]JlIary , or mhoa:l.d l)(; reD\].aced hy a sca\] (? cove1~J ii~; a 
greater range of values. The cent\]us Jell dra~$\[l there 
J.S bh&t Oils Call expal\]d the tradJ tiona\] two.wa\] ue 
system to a six-vat\[us scale of valency-hJnd:\[np~, wJth 
three d:iffercnt types oF complement, Slid so--called 
'mid(l\] es ' and ' extra-per J.phez, a\].s ' J n addJ tJ on to 
adjunct which :'LS kept (see Figure 2). 
integral comp\].ement ! 
obl :i L*atory eomplement 
optional conlplement 
Ii/iddle 
i 
adjunct 
extra-t)eriphera\] 
P'jgu_re .'2. IlJerarchy of valency binding 
(i) Integral compleinents are complements which ape so 
Ultlch i)art e:P the pred:i.cate that they cannot be 
ell\]riled andsT' !trig c\]rcllmstallCeS at a\]\] , net can they 
enter subst:\[tation i)aradiglllS (e .g . prononlJl\]ar:l sation , 
replacement by a relative c\]altse) . They are 
\]exica\]\]y degopm:h\]od by the verb, whereas other types 
Of' COl/lp\] elnent can be characterized by illore genera~ 
senlantic ai\]d lnorpho-syntactic resgi~ictJ.ons. F, xalllples 
arc the nomMm:l.s :in j)f~ve the way, take .SLF_~sk, keep 
pace, etc. The idea Js after WeissgerheP (1983). 
( JJ ) Ohligato:py complements are these comp\] ements 
which are 'oh\]:igatory' :in the traditional (Valency) 
sense, i.e. Jn acLJ.ve declapative sentences. Notice 
t;haL i t :ks not only dJ accurst phenomena \].ike 
'o\]il:ipsfi s' (IIerJ nger, J968:427) that can lead to 
those COlllp\]emslltS being ore:kited: Jn noll-l'illJto verb- 
phrases , and dn llOlllJ oa\]J.sationm ~ 8noh OlllissJonm &re 
ttllrCllla\],kklb\]c . Oh\] igatory celnp\] elllellt s need to be 
dJ.stint~lt J shed however sdncc , when om:i tied from 
surface structure , they must o:\['ten be 'recovered ' J n 
dcsper representations , e.g . for the Stir'poses o;f 
control (Bresnan, \]982 ). 
(JJJ) Options\]. COlllp\].elnentll are those slements whJcll 
arc prcdici;ed (or suhcategor:lzed) by tile verb~ hub 
which are not oh\]igatory as in (id): these correspond 
cdose\]y to tile tradJ.tiona-\[ Valency 'complement'. 
(Jr) The middies va\]ue is needed for e\]omenl;s that 
are not valency--bound as Jn (J)-'(J:i.J), but yet are 
stJ\]\] costly pred:Lcbod hy tile verb-type, and arc 
less per:il)hera\] than a(ljuncts (nee (v)). By way of 
exampic consider ( 8 ) , where B:i \]3 and wJ ndow are 
clear\] y colllp\].erliel\]ts ~ \[Jilt, where wJ th a stone seclns to 
be more elosc\]y bound to the VS:l'b than yesterday 
without acttlally atta.iaJ.llg complement stattls. 
(8) B:i\].\] sniashod the window w:ith a stone yesterday° 
(v) Adjuncts are tile remai.ning trad:i t;:i ona\] 
C J rculilSt~l.n IJia\] e\]\[enlel\]t f'. , Of'ton OX~\[)ressJ n~{ t :bile > 
\]ocat J.OD , lliar/ner ~ cause 1 collsoql/ellco anti so Oil ; 
cofltrary to the traditional view however, l;he,qe are 
not 'ljllSCP~al)\] O tO O:O {;\]lllJzlah\].e fPOlll filly uon\[;encs 
a:huost lit will" (est. IIe\]h:ig, 19'(:1:36)o Adjuncts al:>c ' 
StlbjocL /;o rather vague:i' ~-I elrlal\] t J c re st:e:i c t;:i ons 
del>:\[v J.n\[~ \['PO\]It tile; eetltra\]\[ pre(lJeaO:ioa (vei'h 
CO1\]lp\] elI1oats ) ;iS a WI\]o\] O 
(vJ) Extra .peripherals are on the other hand more or 
119 
less free in this sense. These are sentence 
modifiers, such as as mentioned above, indeed, inn 
contrast to this and so on. 
Notice that the scope of each type of element is the 
predicate plus those elements above it on the scale. 
Thus, middles pertain to the predicate plus 
complements; adjuncts modify this extended 
predication; extra-peripherals range over the entire 
sentence. 
4° How the extended system helps in MT 
Let us now consider the effect of this concept on 
mapping of arguments in transfer in MT. We assume 
that the general transfer algorithm has the 
possibility of correctly matching complements in 
corresponding valency patterns (whether by arbiitrary 
numbering, on a verb-by-verb basis, or using Case). 
We also assume that adjuncts and extra-peripherals 
can be dealt with satisfactorily. We must now 
consider what to do when, on a bilingual basis, 
there is some mismatch between the valency bindings 
of corresponding elements. Notice that this is a 
major problem in the old two-valued system. 
Tile most likely realisation of this problem seems to 
be as follows: assume that in a source language 
analysis of (8), with a stone has been assigned our 
'middle' label, whereas in the target language either 
(i) the equivalent argument is considered to be 
valency-bound for the predicate; or (ii) the 
equivalent argument would be regarded as wholly 
circumstantial. In the case of (i) we can treat the 
element as if it had been a complement all along: the 
valency pattern for the target verb will tell us what 
to do with it. Likewise in the case of (ii) we can 
treat it as if we always knew it was an adjunct: 
depending on the nature of the algorithm for treating 
adjuncts, we may be required to compute some more 
information about it (its case role for example), but 
this is no problem. 
It is easy to see that a generalisation of this 
solution is one in which each degree of valency- 
binding in Figure 2 is allowed to map onto either 
itself, or its nearest neighbour in either direction. 
Thus, integral complements and ordinary complements 
should be compatible, as should adjuncts and extra- 
peripherals. This blurring of the distinctiions up 
and down the hierarchy accords well with intuitions, 
which suggest that because the syntactic (and 
semantic) behaviour of elements at neighbourdng 
points on the hierarchy have aspects in common, 
partial compatibility should be achievable. If we 
add, at the very top of the hierarchy, a notional 
empty or 'zero' binding, we are even able to account 
for those cases where a verb plus integral complement 
in one language is expressed as a simple lexical verb 
in another, as for example with the English and 
French pairs in (9): 
(9) take part participer 
make a mistake se tromper 
catch cold s enrhumer 
welcome fairs accueil 
reverse marcher en arriSre 
stroke donner une carcase 
Problems would arise if analyses of two languages 
were so different that a complement in one was 
120 
regarded as an adjunct in another (though note that 
in the current two-value system, this is common), or 
an integral mapped onto a middle, and so on. Nothing 
much could be done in such a case, and it can only be 
said that the single condition upon which this system 
depends is that that situation simply does not arise. 
However, there is a second problem arising from the 
question of non-matching valency patterns, and 
indeed, this is a problem for which Case seems to 
offer a solution. The problem is this: in language 
A, verb V is considered to have two valency-bound 
arguments, but in language B, the corresponding verb 
takes three complements. When we come to transfer 
between A and B, how do we know which two of the 
three complements specified for V in language B map 
onto the two specified for V in language A? 
5. Case in MT 
Compared to the number of systems using Valency, 
there are relatively few systems that use Case in MT, 
even though there are a number of problems, as we 
have seen, with the bare Valency approach. Yet Case 
obviously has attractions for MT, especially due to 
its character as a bridge between syntax and 
semantics. There is some dispute as to what amount 
of 'semantics' is needed for successful translation, 
or indeed whether Case can properly be described as 
being part of this ~omain. 
Nevertheless, we have seen in the previous sections 
at least two problems which, we suggested, could be 
solved using Case. The first was the like/plaire 
example, wbere we had to have a special mapping 
algorithm, whether between arbitrary ARG labels, or 
between numbered ARG labels interpreted in terms of 
syntactic function. The point about case labels in 
this example is that they will provide a non- 
arbitrary transfer medium which is intuitive in both 
languages. If we extend the system to make it 
multilingual, the case for Case is even stronger, 
since it provides a single language-pair independent 
mapping in contrast to the necessary n(n-l) (for 
languges) bilingual mappings, which might, remember, 
need to be replicated for almost every verb. Even in 
the problem case above, where analyses required non- 
neighbour mapping, a solution in terms of case labels 
suggests that the valency-blnding labels can be 
ignored where they do not help us. And the Case 
solution to the final problem mentioned above - how 
to identify which arguments are considered in both 
languages to be valency-bound - is self-evident. 
The point about Case is that it exactly meets the 
need in MT for a level of representation that is both 
deep enough to serve in a manner relatively 
independent of surface form, while still being 
shallow enough to allow a fairly straightforward 
mapping from and onto these other levels of 
description (i.e. surface syntax, canonical form, 
etc.). This view is countered by Tsujii (1982:382), 
who claims that with only the shallow meaning 
representation afforded by Case, one does not avoid 
the necessity to look still at specific verbs in 
order to interpret the case labels and so to 
determine the appropriate strategy. But Tsujii 
wrongly attributes to advocates of Case for MT the 
idea that it should replace other forms of labelling 
in the representation of analysis. 
It should be made clear that indeed information about 
syntactic configuration goes hand in hand with case 
structure information. The introduction of case 
labels permits certain generalisations that can act 
as defaults to reduce the amount of explicit 
information associated with each verb, in the manner 
of lexical rules in LFG or metarules in GPSG. 
Although we may consider Fillmore's (1968) notions of 
a case-hierarchy for subject selection, or the 
association of prepositions with cases to be 
overstated, this is only because he made the mistake 
of positing these as generalisations rather than 
defaults. 
Of course, Case does introduce new problems, not 
least of which is the definition of a comfortable 
case set. But this is not a design problem, but a 
development problem, and one would like to believe 
that the appropriate case distinctions will become 
evident during a period of experimentation with a 
prototype system. Like in any engineering (as 
opposed to theoretical) undertaking, one must be 
prepared for a period of flux where the appropriate 
tools are developed and modified. 
It is appropriate at this point to look briefly at 
the extent to which Case is currently used in MT 
systems. Perhaps the most extensive use of Case is 
found in LRC's METAL system (Lehmann et al 1980). 
The system is basically transfer-based, with the 
results of a context-free parse of the source text 
passed to the 'case frame processor'. Case frames 
are also used Jn transfer. The cases in the system 
are divided into two types, 'central.' and 
'peripheral', corresponding roughly to the 
traditional Valency distinction between complements 
and adjuncts respectively. 
A key aspect of the METAL conception of central and 
peripheral cases is that these form two discrete sets 
of eases. This means that each of the cases that are 
recognized is regarded as either typically 
"conditioned by the particular verb" or not verb- 
specific. However, it is reeognised that "... a few 
classes of verbs may be closely associated with 
arguments which are usually considered peripheral for 
most verb classes" (Lehmann et al, 1980:1-24), the 
example given being verbs of motion which often 
specify a locative complement. 
Lexieal entries indicate the range of cases for each 
verb, for each of which three types of 
subcategorization information are given, as follows: 
(a) the '~emantic type' of the argument; (b) its 
canonical syntactic role (surface case); (c) the 
syntactic form of the argument, i.e. the kind of 
constituent expected for each argument, c.g. clause, 
noun-phrase, prepositional phrase, adverb. 
In addition, verbs are assigned to a 'transitivity 
type'. This feature gives the "potential 
configuration of arguments", that is, relating the 
possible ease roles to the canonical syntactic roles 
according to the mood and voice of the verb. This 
feature does add information not expressed elsewhere, 
even though at first sight it would appear only to 
corroborate tile correspondence between deep and 
surface caI~e: it is according to transitivity type 
that rules regarding sentence forms other than simple 
active indicative can be generalised. 
In contrast with traditional approaches to Case, note 
that requirements regarding semantic and syntactic 
restrictions are not necessarily directly associated 
with the cases, as was found in Fillmore's original 
(1968) proposal, but can be specified for individual 
verbs, though a system of defaults does apply. 
The case frame processor attempts to 'use up' the 
available arguments by matching them to the 
specifications given in the lexical entry for the 
verb, and then applies general 'case functions' 
associated with the peripheral cases to the remaining 
arguments. If all the arguments can be legally 
assigned case roles, then the clause Js considered 
well-formed. Otherwise, tile rule in deemed to have 
failed, and an alternative frame processor is tried. 
Case frames are also used in transfer, primarily to 
order the constituents and assign them appropriate 
syntactic function markers. Sometimes, if a case is 
'marked', there might also be a change in syntactic 
form, most often (presumably) from noun-phrase to 
prepositional phrase or vice versa. 
Lack of space prevents us from looking closely at the 
list of cases used. The list of central roles more 
or less reflects the consensus view. In addition, 
fully 25 major and additional peripheral ro\].es of a 
rather less abstract nature are proposed. 
Of interest in the documentation describing the case 
roles Js that some of them are exemplified in noun 
frames rather than verb frames (see below). More 
bizarre perhaps is that although these roles are 
allegedly typically peripheral, many of the examples 
given show them being used as 'obligatory' arguments, 
somewhat undermining the central-peripheral 
distinction. This leads us to our most important 
comment on the METAL system, with regard to the 
central-peripheral distinction. There is a serious 
problem, if we admit the possibility of using these 
cases also for 'obligatory' roles, of conflict 
between these more specific and the slightly more 
general central cases. For example, there is an 
Material case, as in (10a), which would however also 
seem a good alternative candidate to Target 
(=Patient) for the surface object of use in (iOb). 
(10a) John built the wall out of bricks. 
(10b) John used bricks to build the wall. 
My own proposals will show how this kind of problem 
might be addressed, in particular by making a 
stricter distinction between 'central' and 
'peripheral' cases, though allowing the latter as 
'secondary' labels attaching to the former where 
necessary. 
6. Valency and Case in MT 
In this section I wish to present a synthesised view 
of the use of Valency and Case in MT, taking into 
account the points raised above. In the following 
discussion, I assume a transfer- rather than 
interlingua-based approach, in keeping with current 
views on this dichotomy. 
Remembering the necessity to keep transfer as small 
as possible, we can see the motivation for 
introducing the degree of shallow semantics offered 
by Case. We saw in Section 2 the problems associated 
with transfer based solely on labels distinguishing 
only between complements and adjuncts. We have also 
seen in Section 5 how the additional, information 
12t 
carried by case labels simplifies the mapping of 
arguments in transfer. We should also take into 
consideration the fact (cf. Somers, in press) that 
all the traditionally proposed cases (with the 
exception of Patient) can occur both as complements 
and adjuncts, even simultaneously. It seems 
therefore that the key to transfer is a label 
indicating both a case relation and the valency- 
binding of that element, so that each constituent is 
uniquely identifiable. We also seek to incorporate 
the additional valency-binding values (integral, 
middle and extra-peripheral) introduced in Section 3. 
For guidance as regards the choice of a set of cases, 
I would like to propose an adaptation of the approach 
found in METAL (discussed above), where there is a 
strJ.ct distinction between the cases used for central 
and peripheral roles. We saw in our discussion above 
that some uneasiness results from the combination of 
general and specific cases, since often both a 
general (traditional) and a specific case assignment 
seems appropriate. The approach I wish to propose 
here involves the definition of a relatively small 
set of traditional cases which are associated 
primarily with complements. The notion 'relatively 
small' must of course be expanded. The cases in this 
set will be rather abstract in the manner of 
traditional cases. Because they are to be used 
essentially to distinguish valency-bound elements, 
and because the maximum possible valency for any verb 
is probably four, we could argue for having only that 
many of these 'inner cases'. However, we also want 
to take advantage of the mnemonic value of the case 
names, so as to make assignment of case labels (by 
humans, perhaps in relatively independent research 
groups) as easy as possible. The number and exact 
values of these cases is to be fixed by some 
preliminary research, and is not in my view an issue 
of import. The important point is that this set be 
fixed, i.e. forming a closed set, so that one of the 
cases can be used quite explicitly as a 'wastebasket' 
or neutral case (cf. Mellema, 1974) for those 
instances where there is no obvious case assignment. 
Judicious and motivated elaboration of the case-set 
will reduce such instances to a minimum. 
A further range of less abstract cases will serve for 
the remaining degrees of valency-binding - middles, 
adjuncts and extra-peripherals. These will typically 
be more specific, and their character determined by 
particular translation problems as they arise during 
the development stage of the system. Some 
suggestions of candidates for this list can be found 
in the METAL documentation. Given the existence of 
these specific cases, it would be quite acceptable to 
use them as secondary labels in conjunction with the 
traditional cases (particularly, perhaps, the neutral 
case), again as and where necessary or helpful, 
though their presence would not typically be a 
requirement for transfer. They might, for example, 
be found to be of assistance in choosing appropriate 
surface structures in generation for some target 
language, though not for another. 
There remains the problem of the roles of arguments 
in non-verbal constituents, since these too have a 
dependency structure. Furthermore, the recognition 
of the nature of these relatJ.onships is often 
essential in translation. Compare the noun groups in 
(11), all of which have a similar syntactic 
structure, but which represent different translation 
patterns in French, depending on the relationship 
122 
between the modifier and the head. 
(11a) a steam train un train ~ vapeur 
(111)) a football pitch un terrain de foot 
(llc) a cotton shirt une chemise en coton 
(lld) the baby linen le linge pour bSbK 
Although it could be argued that these noun-phrases 
could be treated as individual compound lexica\] 
items, the type of construction they exemplify is 
typical, widespread, and such 'compounds' occur 
freely in novel combinations, often involving a 
larger number of elements (12): 
(12a) the Geneva peace talks 
(12b) SDP election campaign promises 
(12c) North Sea oil. field offshore rig administration 
These pose considerable problems when they are 
translated into languages in which such opaque 
compounds cannot be formed and where the 
relationships between the elements are made more 
explicit. Therefore, these relationships must be 
represented at input to transfer. Indeed this has 
been recognised as a problem in translation (human or 
mechanical), and Case suggested as a descriptive 
mechanism for such structures (e.g. Bauer, 1978; 
Mackenzie, 1983). . My own approach would be to 
propose that the case set be supplemented by a number 
of relations specifically suited to noun groups: 
these would be compatible with the cases already 
established, with as large an overlap as possible. 
Clearly, in the case of complex noun groups whose 
head is derived from a verb, either as a direct 
nominalisatJon (13a), as a cognate form (14a) or even 
perhaps when the link is essentially semantic (15a), 
there is no reason why the range of cases (and 
valency bindings) that would apply in the 
corresponding verbal predications (13b, 14b, 15b) 
could not be employed within the noun group. 
(13a) the possible future dismissal of teachers for 
incompetence 
(13b) Teachers will possibly be dismissed in future 
for incompetence. 
(14a) my mother's temporary loss of memory 
(14b) My mother has temporarily lost her memory. 
(15a) Everton's recent 6-1 home victory over Arsenal 
(15b) Everton recently beat Arsenal by 6-1 at home. 
By the same token, nouns not so related to verbs must 
be seen as heads with arguments in some relation. 
Work on nominal valency (e.g. Sommerfeldt & 
Schreiber, 1977; Teubert, 1979) has recognised that 
many nouns can be attributed 'valency patterns', and 
Pi~ha (1980) and Mackenzie (1983) have taken a Case 
approach to the problem. Among their findings is the 
observation that while almost no nouns take 
obligatory complements, many noun modifiers can be 
regarded as valency-bound (e.g. the noun book has 
complements indicating author, subject and contents, 
signalled by by, about and of respectively). 
Furthermore, there are a number of case-like 
relations that are appropriate for noun modifiers, 
among them almost certainly an essentially nentra\] 
one, for use - in the context of MT - when there 
turns out to be no need to compute the reJationship: 
atLrihutive adjectives may be an example of this. 
It is not my intention here to establish, discuss and 
justify such a set of rela~.ions. In any case, only 
by extensJ ve research and experimentation with a 
working sys t, em could sne expect be be ab\] e to 
discover the range of relations needed. 
7. Conclusion 
It has been the :intent\]on of th:is paper to nlake a 
statement about tile proper approach to the use of 
case and Valency J.n MT. On tile one halld, I have 
presented a proposal to abandon the old two-~Jerm 
Valency system ill favour of a new six-term system, 
and have explored the consequences of such a change. 
In connection with this, I have tried to show where 
Case fits in. What this research perhaps lacks at 
this point in time is empirical evidence that such an 
approach is valid, and an accompany:ing set of case 
names for the various relations. It could be argued 
that tlle latter is a prerequisite for the former, 
though I would prefer to claim that the demonstration 
of the validity of the approach and the elaboration 
of a 'comfortable' set of cases form an essenl;ially 
unif:ied research task. This paper, ellen , has 
explained tile Lheoretica\] background to such a Lask. 

References 

Bauer, L. (\]978). The grammar of nominal compoundin~ 
with spec:ial reference to DanJ.sh ~ English and 
French, Odensc: University Press. 

Bresnan, J. (1982). Control and complcmentation. In 
J. Bresnan (ed.), The men~al r ei~resentation of 
grammatical relations, Cambridge, Mass.: MIT Press, 
282-39{). 

Chafe, W.I,. (1970). Meaning and the structure of 
lanEua@e. Chicago: University Press. 
Fillmore 
& R.T 
theory 

C.J. (1968). The case for Case. In E. Bach 
Harms (eds.), Universals Jn \]i1~uistic 
New York: IIolt, Rinehart & Winston, 1-88. 

Fillmore C.3. (1971). Some problems for Case 
grammar. In d. O'Brien (cd.), 22nd Annual Round 
Table. Linguistics: developments of the sixties -. 
viewpoints of bhe sevent:ies Washington, D.C,: 
Georgetown University Press, 35-56. 

He\]big, G. (1.971). Theoretische und praktische 
Aspekte sines Valenzmodells. In G. Helbig (Hrsg.), 
Beitr~$le - zur Valenztheorie The \]Iague: Mouton, 31- 
49. 

Helbig, G. & W. Schenke\] (1973). W6rterbuch zur 
Valcnz und Distribution deutscher Verben. Leipzig: 
VEB Verlag Enzyk\]op~die. 

Heringer, H.-J. (1968). Pr~tposJ tiona\] e F, rg#inzungs- 
best ilmnungen \]m Deutschen. ZeJ tschrJ ft flit 
Deutsche Philologie 87, 426-457. 

Lehmann, W.P. , W.S. Bennett, J. Slocum, H. Smith, 
S.M.V. Pflluger & S.A. EvcIand (1980). The METAl, 
system. RADC-TR-80-374 . Rome Air DeveIopment 
Center, Griffiss APB, N.Y. 

Lehrberger, J. (1981). The linguistic mode\]: general. 
aspects. In L° Bourbeau, Li~uistic documentat:ion 
of the computerized translation chain of the TAUM- 
AVIATION system, Montreal : TAUM, Universit$ de 
MontrSal. 

Mackenz J e, J.L. ( 1983 ). Nominal predicates in a 
PunctionaI Grammar of F, ngldah. In S.C. DJk (ed.), 
Advances in Functional Grammar, Dordrecht : Poris , 
31-51.. 

Mellema, P. (1974). A brief against Case grammar. 
Foundations of Lml_guage ii, 39-76. 

Pitha, P. (1980). Case frames of nouns, in D.J. van 
Alkemade, A. Feitsma, W.J. Meys, P. van Heenen & 
J.J. Spa (eds.), I,inguistJc studies offered to 
Berthe Siertsema, Amsterdmn: Rodopi, 91-99 ; 
reprinted Jn F, Sga\]l (ed.), Contributions to 
Functional Syntax± senlantics and language 
comprehension, Praha (\]984): Academia, ?25-238. 

RothkegeI, A. (1976) . Va\] enzgrammaLik I. 
LinguJstische ArbeJten 19. Sonderforschungsbereich 
Elektronische Sprachforschung, UnJversJ t~t des 
Saarlandes, SaarbrQcken. 

Somers, H.L. (1984) . On the validity of the 
complement-adjunct distinction J n Valency grammar. 
Linguistics 22, 507-530. 

Somers, II.L. (in press). Valency and Case Jn 
Computational I,insuistics, Edinburgh University 
Press. 

Sommerfeldt, K.-E. & H. Schreiber (1977). WOrterbuch 
zur Valenz und Distribution deutseher Substantive. 
Leipzig: VEB Verlag Enzyklop~die. 

TesniSre, L. (1959). ElSments de syntaxe structuraIe. 
Paris: Librairie C. KIfincksieek. 

Teuber~, W. (1979). Valenz des SubstanLivs. 
D~sseIdorf: P~dagogischer Verlag Schwann. 

Tsuj i\] , J-I. ( 1982 ) . The transfer phase in an 
English-Japanese translation system. In J. Horeck9 
(ed), COLING 82 (Proceedings of the Ninth Inter- 
national Conference on Computational Linguistics, 
Prague), Amsterdmn: North-llolland, 383-390. 

Vauquois, B. (1978). Description de la structure 
intermSdiaire (Communication pr~sentSe au Colloque 
Luxembourg, \].7 et 18 avril 1978). GETA, Grenoble. 

Weissgerber, M. (1983) . Valenz und Kongrtlenz- 
beziehungen : ein Model\] zur Vereindeutun~ yon 
Verhen Jn der maschineIlen Analyse und 0bersetzung. 
Frankfurt am Main: Lang. 
