Valency Theory in a Stratificational MT-System 
Paul Schmidt 
IAI Eurotra-D 
Martin-Luther-Str. 
6600 Saarbr~cken 
West Germany 
14 
Abstract: 
This paper tries to investigate valency theory as a 
linguistic tool in machine translation.There are 
three main areas in which major questions arise: 
(i) Valency theory itself. I sketch a valency 
theory in linguistic terms which includes the 
discussion of the nature of dependency 
representation as an interface for semantic 
description. 
(2) The dependency representation in the trans- 
lation process. I try to sketch the different roles 
of dependency representation in analysis and 
generation. 
3) The impl~m~_ntation of valency theory in an MT- 
system. I give a few examples for how a valency 
d~xzription could be impleme~ted in the EUROTRA- 
formalism. 
O. Introduct:\[on 
This paper ~ies to apply a piece of established 
linguistics, known as valency theory, to the 
problem of machine translation. As such, it is 
meant to fit into the forthcoming EUROTRA MT 
system, though it does not deal with EUROTRA 
problems sp~;ifically. 
There are two aspects which play a role in the 
building of ~Ln Ff£-syste~; 
(i) the development of linguistic specifications 
and 
(2) the development of a formalism which allows for 
the implemenlmtion of linguistic results. 
~lis paper aninly deals with the first aspect and 
~/~us relates not only to EUNOTRA but also at least 
to all stratificational systems, i.e. systems t2mt 
break up the translation process into a sequence of 
sJx~#ler translation processes. 
Furthermore it relates at least to any system wi\]ich 
uses dependency/valency inforalation as e.g. IAgG 
does with its functional structure. 
In EUROTRA, the level where information about 
dependency/valency is used is the hqIS (Eurotra- 
Relational-Structure) which lies between the 
constitueat stzncture (ECS) and the semantic 
intea-face stzucture (IS). 
So, in EOROZRA terms I try here to give a kind of 
ERS-definition in the language of empirical 
linguistics without touching the formalism itself. 
The investigation divides into three parts: 
(i) the sketch of a v~l.ency theory which comprises 
the following points: 
- the informal definition of the concepts valency, 
cc~plement,ar~ adjunct, thereby trying to give a 
definition which holds for verbs, adjectives, nouns, 
and prepositions, 
- the operationalizati~1 of the complement adjunct 
distinction derived frc~! this definition, 
- the classification of the complements, a 
subclassification of f~e complement classes (C- 
classes) according to their syntactic realization, 
and the determination of the relevant sentence 
patterns, 
- a short discussion of the relation of the 
dependency level to constituency level and from 
dependency level to semantic representation. 
(2) the application of the linguistic 
specifications to the problem of MT which llas to 
investigate the role of dependency representation 
(D-representation) (oi" rather the role of the 
transitions to D-representation, since levels do 
not "play a role") in an,91ysis and generation. 
(3) the way the gained linguistic information can 
be implemented according to the EOROTRA formalism. 
i. Outline of a Syntactic Val~ ~_~ 
I. 1 Definition of Concej3~s 
The syntactic vale~zy of an el~ent of a word class 
(a nonterminal category) is its property to bind a 
certain number and a certa~ kind of syntagma. 
Those valency--bound syn'l~o~,'e:-~ are the oon~icm~e~ts. 
'I~e s%a~tagnms which are ~t valency-bound are the 
adjuncts. All syntagmas which are obligatory in a 
syntactical sense are valency-bound, i.e. the 
sentence a~st contain tt~t item otherwise it would 
not be complete. 
Bedsides the obligatory syntagmas the wordL~ubclass 
specific elea~ts are valency-bound (Engel 1982). 
Wordsubclass specific e\]em~--nts are those syntagmas 
which can occur only with elemlents of wordsub- 
classes. 
i. 2 Qperationalization of valency 
'fhis point usually forms the major part of each 
paper on vale~cy theory and mostly consists of a 
discussion of various proposed tests in the 
literature on valency theory showing their 
deficiencies and ccmdng up with a new proposal. 
307 
~e discussion of proposed tests will Im left out 
.in t~is ~ and the proposal for an operationali- 
zation tb~t will be made consists of an adaption of 
a test which has already imen established m~ which 
~x3uld be called "the free addibility test"° This 
test will k~ l~filled here. and is ~lus a "n~x\]j.fie~ 
addibility test." It, will be. explained. ~min\].y by 
applying it to verb valenc~ lxl't is also meant to 
apply to word--,classe~ oth.ea~ 'tt~ ver~x~. 
In order 'to operationalize the ter~t "obligatory 
c~iDle-~nti' the eli~mlrk~tJ.em test :\[.s alE~lie<l. ~.the 
b~\]ta~nu in a given sentPnce is (_%~\].igai~ry j.f it 
cannot ~. e\]..i~tir~\]t~l wit/qout Z~ea~e.ring the serltel~oe 
ra~/\[:nmm~t, ical. I:n G'ez\]tmn, (rn\].y vel}3s, adjc<:;tives 
a~fl prelx)sitions c,m l~,ve Gt?ligatx):\['y c~2~aple:uexlts0 
\]:n ozder i~) operatio~m\].Jze ~tJ:aclass s\[~\]JfJ.oJ.%y, a 
:modified addibility test. Js al?P:\[ic~i. '.l.'bz'ovlghol_lt the 
his'tot!/ of 'tkm thcm)ry of valenoi/, l~/mly tests have 
i:bJ..~ ~e#Jst adeq0kt'tg; as t:t){,¢',:e j£; ;m (~)vJ,ous relatioll 
be.t~,qeen defJ.nJi:ioil, o:£ coi~cept an(\] ope:ua'tinal:b-. 
zat::\[o!i0 It a\].sc~ has the property of being 
l.~lt_'-tia:l.ly a\[~pli<:able J_i-,.l:~_9:\],Jrlc\]uLl'J.ly~ wh:\[dl is l\]et 
ll~:ltii\]:~£t-{~aTlt for 'i'~LI\]~<Xi!\[~.< 
'.lXbe ~f:cee add,ihd\] J.ty "test" {:onsis'ts Of chec\]cJng 
:,::~si:z'icLtJ.o~ts :ko'c a ~;)zntagma~ i,eo J.J: Ji; ea~l bu 
fr(;ely added, it :if., an a~.djunot ahd no'i:, va\].euey 
bO'tli%d 0 9J,llJlg t{JS\['. 7 hi IwexJ8:C ~ C\[!~lliOt h~l~ld\].e i::h(~ 
ftJ\]luwili U \[{)ro.b\].ea{~;o ',l+'h~ :<(:.sp~o'tJvo ~xp:lanat:i.oil:', 
lead to ~ z+efilic¢~teyii: ():if i:b~ test ~:~KI i:hu,'; i-o t'hc- 
%li<×li:iJ£7:l o.dd4bi\]J:L~" t'.e:;h~i+ 
<*.) '£'.b.eG ~, ar:,~ :gest:~J.oh:{.ons Jx~ {?olitbJ.ning tehkPOra:i 
adjnuc;ts wkth 't~<~s o.f t:h~ verb0 
(\]) ~'a\]!tr kbltl ltto~xJo:~lo 
(*He Czlhl\[; i:~7~Jgrovx0 ) 
b) 'i'here are resi:riciki.ons Ill coiabining (;e\]?ta:h\] 
elr>im,lits o.< adjuuot classes wJkh certaJrl ~m.J:Jjeuhso 
(2) *\[~;',_: l'~'u,,t st-eht a~l~sic._ht\]_:Lch am Ilang0 
(,VZhe l'.:t.'~e Js <{dl.:{bel:at-mly on the s\].up~;o ) 
\]21 (a) and (b), it can bR proved tllat the 
restz'iotions do not c, olile f\]2o\]il the ve~% bl/\[', a\].o 
caused by other reaso~'ts~ L~:i: is possible tu fo.~mt 
coL~e{:t sen\[~nces <x)n-taJnii~ both the verb ate! I:h¢; 
adjunct. 'l'herefore~ there is no principled in. 
possJbilJ:Ly of their ccm~binahiono 
(3) E.~r koiimit i's3~:~{(_'n (lie 
wil:L c~m~. t(xr~)rrtm~. ) 
Fritz steht absichtlich am At~mng. 
(1,'ritz starKis delJlx~.rate.ly on the slcrDe, ) 
c) Moi~ proble~intic is t/le case Miere elea~n'ts of 
subclasses of adj~xl~t c.lasses are in facR% not corn-, 
binable with a verb. In 
(4) *~citz atz~t ge~'lt\].ich0 
(*~%itz l)mmi91~s cozily.) 
the restriction (~L~s Jn facYc fr~li t.he vel%. ff1~e 
syntagi~. "gearfltlich" which is denoted as an adjtmot 
would fulfil tkle r~uirement of being s\[~m\].ass s~- 
cific, Therefore, the addibility test has to l~ 
restri cT~<l. 
(5) l'~J~tz ai-~t sd~we:r. 
(Fritz breath~ heavily° ) 
AS (5) is COr£'cf;h alld "schweaf" belo~3s to the sa~ae 
~\]" ~"~ o5 adjnncts~ i0e. adjunct.s of manner, hhe 
addibility test will have i:o be rest:td.c:ted to 
c\]asses of synt.agmas. :f.:t\] elex~lents fro~L adjuncL 
classes are :D:.(~-~I.~ c(mtblnable, the add:\[b:ility test 
is considered as ft:d=fil\].~ 
d) Howe.vet, 'there are c~_~t:ts_in problc~t~ltic c.lses0 
(6a) *_\]~ x'eiff\]o% \]<i.t, einoa,~ !I~_n~iuaro 
(*l:t is :~:'a:inJ.~g w;i.th a bamrmr0) 
(6b) ';d~:ts ALl:to :iSh :r:Ot b.J.~:J \]t.~oi~3er,~ 
(~',H\[IO Ca\] *~ JS \]?(Y! Lust \]. tohlOI:£'OW.. ) 
CZtn sca:r{~e:!y be. ~ desorJL~<d as \].x.~.j.-q.g :!-\]eel.y adtak.te 
as '\[.hey onJ:y (:om})Jl~e ~:LLb. actJ.o~\] verbs and su 
:{u\]_\]'i! the (;vi't*e3<J_o:U 0£ ,~;U*X:',IOSS ;;p cificJ g<, 9,1 
i:he Ot'h{¢<' .}\]aTg.\], /'bey (".;:}vl }AE~ Sl~b~l\]ttLoL{ 1\]\]~t~eK' ~llC'h ¢{ 
l\[-tmf(j~ ,~:~li\[)(TIoL:4/, eLL: V(!!LU'OS t~\]ai: ~qe \[~\]D ,l(} " o~i(:}(~< '\[w {.\]< 
y:{e{~s(t~lab\].y elf ::;b'bC\]a,qS s~.~:-'.<.d.:gJ.¢J:tyo \]\]l fac:~ tklese,. 
LWO ~2t~Se2E~ ~i'(-; i}!e i~l:);i;{X (\]Jff:\[cwtlt 'be ~})~\]~\]le a~m'{ \].~t{ve 
\]d;d i :Lngr :\[};t.H \['.{} dr~w dJ.fferen% t c, on(du~ ' one,, ~u~ 
~:!otT",, t\]ice ;4S,~I!~@)\[:.JA,x~. ,~~ di./:i'c:;17eK\[t (\]e9>'c2(_,.;: (1{: 
~xghesJ.o~l., '\]\]hJ_S¢ h(&~zc~v~Y/~ li\[,:-~al)~3 ~l\]Ett '\["\[lc~ £\]J.S\[LJAI(Z\[:i(.\]\]I 
o:\[: ~-;\]~n'\[.(tg*ttas ;.\]l/'(~ (x)llg)\].{~mel:F\[::4 atld adjvl'~l(5{~f?, jr: }~ 
~;eHt{~nce (Somers 1984) ha:~ :-~Jmp\]y beoai ak~r, done:10 
7t m~.rd: be. stat.ed he:r:; that. a) .- d) :i.;~ not. a 
c.omp\].eL.e l:i.si: o£ !)rob\]e\]tiatio' ¢:,as:~,~,, tt'hJ:, pe-i.>- 
iloeds Lo have more .i.tfv3sitf\[gatiol\] and tbu,~ \[:.h~ 
meu:lif:i.ed s \]d' bi\] i-?2 test nex2ds moge ~ft:kfJ.cat.ioi~,. 
gp to now '\[hr: "mc-di.fJed * (\]~ibil J,ty tt!st" Co~Is:i sts 
:h? the :;kit.'l.ow:h~\]:~' 
• :I:f th.e!:e JS :"reG addJbi:lJi:y, i:he ,'-~yntngm~.~ :is an 
adjunci < 
<' \]:f '|'h_t!:es ar:z:~ addJbi\].J.Q/ :<estrichJ.ol\] -i\]:<:~lilt et:lle.~: 
sou~'e~-~s than the veli~ (see a) ': c.) above) t'h.,e 
syntagt~, is au adjua-~ct. 
.. All o'hher l;yntag~_ims tG~ c.omple\]m~t~o 
:I03 Co~:{e\[mg!li:Z\].a<is,Ts 
703. :1 ?prl?s 
The above pz~oposed tests (elimination test, o~J. 
~-xlified addibilit!/ test) lend to ten compl<<um_ht 
classes for German verbs, which caYi be sub-, 
classified further according to their lttorpllo- 
syntactic realization, 
308 
'there are five case~determin~d C-classes, hlcluding 
the pre~x~sitior~l comple~to q_\]~e foll¢~{Jng table 
M~c~s the first 5 C,~lasses as an ex~?\].e: 
C--class synGci::ic; exar%ole 
:vealization 
C0 -. NP in no~:d-. 
~mt.ive 
-, that~clause 
-,o M~,-clattqe 
mis\[~ flee) 
- in:l:j ni'hiw..! 
.F£i~4 schl~ift 
( i,'~d.tz sleeps) 
('\]'ha:c, yOC'~, coz~ \]91ease,,3 l~l,q) 
:~(W_q,:'ht_L~a.cht (.;J2t¢,41 F~hl.er 
(Ei.'1')osf~ who :L(~v~, \]£a\]{o a 
.lhYIJy>" det ml wcl:(k~ULi st. (::J\]\[~ 
scihr:c~t',\],:l.iches f;d:t~c,'ks~'t ('L'o 
_b.:'~ \],;.\] :1.(.~:I :b; an u.wi:u\]. F~r\[:e) 
C:i. o' NP h~ accc(-' 
IDait \[:re 
' ' t:b.~ t'l:'-+::lm w.e 
-" %qh' < l{iILq¢~ 
,. :hi:~ is L-va\] 
(.tO\])'7~{ "t, tt( ; {::\[ C:~ 
v,_- ~;c<b:liigL I.L,U 
(H~-: beats :h.hl 0 
.~ : ;;iob\[::~ ~3~f:~ s:i( L kgn.:d:: 
(lie s<vx4 -Phat, ;41to :i ,q ~:(mlil:j) 
\]ii? ,~JeI'r\[, ~,',:~Ii CU (!:Lil\[JQ\]\[:c\],{)il 
1:~;I:<, (l{e ;;e~;!:\] W\]ucgll y:t~. 
J i ~v :t Led) 
V,<' (JiI~u S\[: Fff:II ~' :4t!. hpl\[)<::! 
)gi? ~.l;:l.gL~ do~:; :ist :i}'~!l,<<J} 
(i:£'. ,~ly~; "\[h\]t::; .i.~; W < :Lqf) 
( 1 2 
• Lhai: c<L{{use 
(iIe :i:,~m¢,_mb::'~:s ?he a sl ,d:-< 0 
(1\]:¢! 4K\]CX!;X'~ :~, }'. \] (~: i~;U1\[l:E{ 
(: (~(\[\](l'i:.'\['~:{'<\[)f\[l| \[<{~ ,:i °) 
'l,',t .......... l\] ~ :;(',}.)U' ((i :,i . :\]:\[)\[~¢ .W -'f Gell 
wh :i.(!h "~ ~ :lit ~\]i\[\[ K~@:k\]) 
\]~:H b'?gC\]:ll.\]\].(\] i.gL J\]~n, %i U::' 
(\]\[\[F~ L!C<;t\]G~I~ 1,, !L i,ff I) LV tz{} 
(:ommLiLD_zd :,,i ~ :fie, ) 
(!3 , N\]' :;1:, 0at.EVe, 
- Wbo (;:l.at~c.<~ 
:!:oh he:l.fe .lb,! (:!: h(<l'p h'im) 
:cb l)e:ld:e,.~:m :iC:h w.i:i::! 
(:!: D(~\]. U Wh(.~m \]: wa.?. "ko) 
(:4 ~. \]):> 
-. %h~t'\[:,- {\]\].auee 
-, inf,,{ x:,nst~'~ 
¢#x ~,m~' i;,: _kqf/. Jh ~ (:\]: wait 
re:,: h:h, 0 
ko:~m~t:. (:!: wai\[ j:o:_r him i.o 
:i?(:):r:' (g~tt:J ~KJ {{ ! thauc~'e0 
'. 'he:r~:; :i.s as t;:i xth class the sJl uativ: ~ (:xJmO\] e.m{n\[ 
l:t can b~ x~a:Lizc~l a.,:; :l;tP as :h~ (7) 
(7) 1,\]~: w~mt ;\[:u Saat'D~.L{!q\]~j3o 
or as adveLg)ial ph\]tase c:)r as Mx'~clause0 
El:he r~.\]ve1"~th Co~class Js the temporal ooltip\].eltmllt as 
in (8), 
(8) Igor1' Vortmag da.u~rt ~:L~K~e.. 
It (~.m be realized as PP~ advP ~x1 as 9~h~<l.auseo 
(tile di:\['(~stJ.onal cQmple~m!nt as fin (9) c~-m ~1.~¢.~ 
:IL~J i:ea:\]LJ.zc~\]\[ as \[~P, advP ar~! as wh~c\]ause0 
(9) i~ r :lf~\]hoct hath_ tIa~Ak~'c!,. 
(lie alL%yes hoa~. ), 
'£he \].denti:iTyJ.n U (~o111i?\],e:i\[lerrt ,~i.,<~ J~i (10) ~ <.%iH b.': 
Lea:i..-iz:_f! as :N!., (ncmdnati:w~) ~:; :{qp (ac,.~\]sat v~9, .A!? 
(:i(~) I':~: + :ist dCX)~.o 
(H{; :i'~ stupid°) 
Id.zlaJ\].l:v 'hh(~ v(rr.%~!l <~.(:,IKD\].cmlt~nL ~t~; 
:<t'~\]izc.d ~-~¢~ :hT{JniLva!. (~en~'{;\[T~o_(:::\[':ion, 
lit~L'i n \[\].'.~Ul~:; 1. 
(:L!)),~r :L66i:. :ih~L k0s~#me,#o 
\[::\[~!u;::i t:,~ \]:,)~ I,I{:1:! j'):l:loS:yi:d:'tu ;\['..i(: Z'~' i '/;{:"Li.o~l¢'4 :. ~t'l I 
\[X~ ;{'i\] ~ '(;iLl i'~\]'~ :i:'O:~< ~ SO~ I\]\['3 (!<~-\]~f'!i'.Ci'tl'P'i(;ii t ic;:~ 
{:.l:!(!,qt~ \].SU ,G(~h\[t'~IX(!':~ ).){{L'\['.~:!<~'Lq~ .')\]0f{ ~il ;:t:!'\]'j'i~:r'~::y 
': l!il}) r" i~:i ~(\]'itiU\[;~'<~{ 
10 : o ;.t Nf )U'il: 
t!.9:~:i(~ :{.,~ ~: ~\[,jo:u d;:iH::i:::u\]:Ly wiib H~c~ ({~;.L~,:i-~u~qaL:L{~h 
i:.s\]?y ('.{fl:l)3:l.@E'~eh-t:\[~ ?\]/L~:~ 11(:)~1H,~{o h'~fil~;~, (\]\]!"ly' 'Lht~ 
<0\].ov/ \['¢j:v: glar;;',ify:iug -i:b ~ t~\],.:i:'~corU! ~:;~u'k..'~Um{t~; :i~ 
(:!3) ~: c<J.mgl.u,~mmrb; btd: tho:;;~ J(* (\].4) ~.~; ~{~i\]rm~_:t'~;o 
(:~3) \]!k':{nil~:i:lmguJl }~t -T,l-~:ido:!'b'-~Wl 0;te~:toJ:J.e ; t::i' 
~_\]:~ !j.de~g):~:<g ) 
npa:,.,i<<Lga~ 9 #~ucc:h d¢-'~t Wa:!.g (~v~Lk: {b:,:oug~ Ph~:~ 
:i~¢ ):r ~:H'L.) 
:=!;t:Wa'Li:L~vj, ~ff~5\]e.T:: ~;b!i:q.!~ k£gmml; (! !\]{,V~{;i:aL:i.~ ::u 
that wJ~rLoV w:i:! \]. G~,~) 
(\].4) 
~;JAI WJs,%':~lLqt~lGft:!.t{ ~q:\[{~ .q'<2 (Q sej~ltJst \].:il<c~ 
\]tin 'tbfD:\[ #,ttjs SJ:~d:t! (apoL li~ade of ,Dteel) 
The ~,1odified addibility test. rules ord: Lhe 
mo1~iplement.s :i.n (\]3) 0 There are restrictionF; ih 
309 
addibility as shown in (15). 
(15) *die M~tter auf Frieden (*the mother for peace) 
*der Tisch durch den Wald (*the table through 
the forest) 
*der Tisch, dab der Winter k~ (*the table 
that winter comes) 
The adjuncts in (13) are freely addable to all 
nouns. There are scme semantic restrictions at most 
holds for those with verbs. 
There are six ocmi01ement classes for German nouns: 
Two c~e-dete~ CO (nc~inative) and C1 (nps in 
genitve, that-clause), prepositional cc~plement, 
directional, situative and identifying ccaplsm~_nt. 
io3.3 ~_~_~ectives 
The determination of adjective cc~plements works 
according to our two tests. There are even some 
adjectives which have obligatory complements, e.g. 
situated. They are dete/mined by the elimination 
test. 
There are six complementclasses for German 
adjectives: four case determined ccaplement classes 
accusative, genitive and dative cc~plements and the 
prepositional complement, there are quantificatio- 
nal ccmlolements and the situational complement as 
show~ in (16) : 
(16) ziemlich qroB (CO) 
(quite l~e) 
die Arbeit gewohnt (C1) 
(accust~ to work) 
der Azbeit m~de (~) 
(fed uP with work) 
den Eltern fremd (C3) 
(alienated from the parents) 
interessiert an Linguistik (C4) 
(interested in linguistics) 
am Flu6 gelegen (C5) 
(situated at the river) 
1.4 The theoretical nature of dependency 
The description of the theoretical nature will 
consist of a short description of the relation to 
the "higher" level of constituency and the "lower" 
level of semantic representation. 
1.4.1 Constituenc~!_and~ de__~ 
The relation between constituency and dependency is 
usually regarded as complementary. Constituent- 
grammars (C-~) are based on the part/whole 
relation and define a hierarchical structure with 
the respective higher ranking category defined as 
being composed of the lower ranking ones. In 
contrast to this, the dependency-grammars (D~- 
mars) define relations between categories of the 
same rank, i.e. there is no hierarchical structure 
in this ranking ~. 
However, this is not sufficient for a descTiption 
of the relation between constituency and dependen- 
cy. The relation becomes problematic if the C- 
is a recursive subject/predicate 
generating a deep tree. In this case, a translation 
in the sense of a stratificational MT-system is 
very problematic without a tree-flattening 
procedure. This procedure could be justified 
linguistically because a C-grmmnar generating flat 
trees can generate the same sentences as a C- 
gr~ which generates deep trees. 
i. 4. i ~_~cy and semantic_ ~_presentation 
The complement/adjunct distinction which has been 
made on the syntactic level using purely syntactic 
tests is of sentence-se~mntic importance. Eac/~ 
semantic representation, be it based on symbolic 
logic e.g. Montague Grammar or on a case gr~, 
is usually implicitly based on the complement 
adjunct distinction. There are, however, 
exceptions. E.g. Fillmore's "instrumental" is an 
adjunct according to the above mentioned tests 
(hc~ever, marked as doubtful) since it has the very 
same status as the other roles in Fillmore's fra~- 
work. 
The differentiation between ccai01ement and adjunct 
is made in Systentic Functional Grammar (Halliday, 
Fawcett) by distinguishing between participant 
roles and circumstantial roles. The participant 
role is a semantic interpretation of the comple- 
ment-verb relation, whereas the circum-stantial 
role is the s~antic interpretation of the adjunct- 
verb or adjunct-clause relation. 
As has been shown (Projektc~pe Vel~lenz 1981) 
the complement verb relation can be interpreted 
semantically in a lambda categorial granuuar. A 
semantic description of an expression of natural 
lanc~ge in a lambda categorial grmmnar consisits 
of a translation into an expression of the 
artificial language lambda L and of a mode\] 
theoretic interpretation of this lambda L 
expression. 
Valency frames of verbs are represented in a lambda 
categorial grammar as the number of lambda prefixes 
the translated expression receives by the transla- 
tion into lambda L° 
The lambda operator can bind a variable in its 
and makes predicates out of sentence~. 
(17) lambda x \[schl~ft (x)\] (ein x sein, das 
lambda x \[sleeps (x)\] (to be an x that 
sleeps) 
Two-place predicates are represented like (18) o 
(18) lambda xl \[la~0da x2 \[betl-dChtet (xl,x2)\] 
lambda xl \[im~da x2 \[i~ at (xl,x2)\] 
I shall not try to show how such an expression is 
310 
interpreted in a model. The point that is made here 
is that a c~mantic description in the frame of a 
lambda categorial grammar uses the syntactic 
relation between c~plement and governir~ verb as 
t2m basis of its sentence - semantic description. 
2. The role of the syntactic dep~ 
re oresentation level in M~ 
The facts discussed purely linguistically in 
section 1 give clear guidelines for application in 
MT. 
2.1 ~_n_cy structure (D-struct~ire) in analysis 
The depende/~cy representation (D-representation) 
serves two ~rpeses: 
a) The tr~unslation into the D~representation 
relates the D-structure to the syntagn~s analyzed 
on the C-level and thus contributes to the dis- 
ambiguation of the C-structures which cannot be 
achieved on the C-level as these can only be 
readied by the valency statements. (This at \].east 
is the case if the two levels are strictly se- 
parated.) Usually, the constituent analysis 
delivers several readings for a longer sentence. 
Each NP in the genitive case for example, which is 
an attribute to a noun, has to be interpreted as a 
potential genitive valency of a verb. In this case, 
the transition from C-representation to D-re- 
presentation filters many au~iguous structures by 
assigning the appropriate d-relation. As there are 
only a few German verbs with ge/litive valency, these 
readings are filtered out in ~st cases. 
b) A second function of the dependency stl~/cture is 
the disambiguation of the verbs (and other elements 
of word classes which have a valency frame). 
Differe~nt verb readings often are discriminated by 
different w~lency frames. An arbitrary example 
proves this: 
(19) anziehe~, CO Die Mieten ziehen an. 
(Rents are rising. ) 
anziehen C0/Cl Fritz zieht die Bremse 
an. 
(Fritz pulls the 
brake. ) 
anziehen C0/CI/(C3) Die Fmtter zie/tt dem, 
Kind die Schuhe an. 
(The mother helps the 
child with the s/toes. ) 
In a sentence J/~ which the verb "anziehen" occurs 
with only one CO, the reading can be identified 
unambiguously and translated by t-rules into the IS 
(Interface) atom with the corresponding case role. 
(In EUROTRA~ the IS is designed as a semantic 
interpreted r~structure). 
2.2 D-structure in qer~cation 
The task of the D~level in generation is the 
g~%eration of the target language D-structure from 
IS by assingw/ng the al~propriate (correct) surface 
syntactic valency fram~!~: 
In the source language, e.g. German, the verb "sich 
erinnern" has the syntactic valency frame C0/CA, 
(which is the ccmi01ement in the nominative case and 
a prepositional complement). In the target language 
English the verb "remember" has t21e fra~le C0/CI 
(which is the subject and the direct object). In 
the target language the language-specific surface 
syntactic valency frame (the direct object) is 
generated frc~ the interlingual IS. The D-structure 
is thus a precondition for the generation of 
cerrect censtituent surfaee structures aeoording to 
their valencies. 
(20) is an example. (It uses the case roles ,,pro ~ 
cessor" and "phenomenon" according to systemic 
functional grammar. ) 
The translation under:foes the mentioned levels. 
Relevant for generation is IS(T) to ERS(T). 
(20) ERS (S) #sic/l erirmern an Co/C4 
IS (S) #sich erinnern an Processor/~n 
IS (T) #r~ber 
Processor/~n 
E~S (T) #r~,ber Co/Cl 
3. The /J~pleme~tation of D-Representation in 
the EUROTRA F1~mework 
3.1. some ~ks on the EUROTI~A - Formalism 
The most important assumption in ~ is "that 
translation between natural languages is a sequence 
of primitive translations between a number of 
levels of representation u' (Arnold et al. 1985b). 
Such a level of representation is a "language L 
generated by a gramnar G and an interpretation I" 
(Arnold et al.1985b). I specifies the syntacti- 
cally and se~antically well-formed expressions of L. 
G consists of a set of atoms and a set of construc-- 
tots. Basically Atoms a\]:~ the lexical entries, the 
constructors are the grammar rules on the different 
levels. Atcms have a r~e and a set of features. 
Constructors have a na~i~ and a set of features and 
a set of arguments which can either be atoms or 
constructors themk~ives. ~ley look like (21): 
(21) Atom: name {f~tl,°..,featn} 
Constructor: name {f~tl,...,featn} \[argl,... 
., argi\] 
A constructor is syntactically well-formed if its 
arity equals the number of its arguments and if the 
arguments are well-form~\]. It is semantically well- 
formed if its arguments unify with the argument 
311 
places of the constructor arguments. 
~e adjacent rep ~resentational levels are related by 
a translator whi(lh is a set of t-ruleso It has been 
said that trans\].ations between representational 
levels are primitive which means that they are (a) 
compositional and (b) "one shot" (Arnold et al. 
1985b) = 
(a) means that atom is translated to atom arm 
constructor 'to constructor, where at least the 
relaxati~% is all~ed that the number and order of 
construoC~ors differ° 
(b) means that the translator takes only well -~ 
formed e~<pressions of Gi and yields only well.-- 
for~t~xl exp~ssiorm~ of Gi+l (All, old et alo 198.510). 
This ~ans that there is no intex-fkll strategy in 
the t~mfLe o 
3.2. Some ideas for t/~e implementation of 13~ 
Struct~nxa in the Et~cotra formalism 
~\]%e imple~rltmtion of D-structure coruqists of what 
has been called patterns, in 1 and it has to be. 
done J/l the for~nlis~ sketched in 3.1. ".\[\]",at: ~-mans 
t2~at for each of the patterns there l~s to be a set 
of ~str~t~z~. 
As impl~ntation Ires jt~t begun the. ideas here are 
sc~.~2~t "vagina, and the proposals c~_rt~inly do not 
~u~e all the possibilities the fonnalisn~ provides. 
3.2.1 Sentence Patterru~ 
For s~ntences the g~neral pattexn ioo~ like (22) 
(22) S{ ~9 } \[gov{ I,t) }, o0~pll{ FO },..., 
c~p\]~{ FD ),adjt~ncffc{ FD }*\] 
%his is just what has been stated above. A sentence 
S with a featrLre description FD consists of the 
pattern with a governing verb with a feature 
de~ciption FD and a compl~ent eonfiguration each 
cc~lement with a featuz~ description FO plus an 
arbitrary number of adjtmct~s each adjunct wit/~ a 
feature description FD~ 
~he above proposed subclassification by syntactic 
realization can be. handled by a cat feature. For 
patte~ C 0/C 1 (sentences with transitive ve~os) 
(23) is a~ example. 
(23) S{ FD } \[gov {cat--v,....}, CO {cat~-1%p, 
case=nero, .... }, C1 {cat=np, 
case=aoc .... ),adjunct{ FD }*\] 
So, the implementation of the 150 subclassified 
sentence patterns consists in an entm~ration of the 
S - constructors acoording to (23). 
3.2.2 NP-Patterns and AP-Patterns 
For NPs and APs the general patterns look like 
(24) : 
(24) NP ( FD } \[gev ( FD }, cc~pll { FD ),..., 
cc~i01n { FD },adjunct{ FD )*\] 
AP ( FD } Igor ( FD }, c~lZ { FD } .... , 
ec~pln ( FO }, adjuncb{ FD }*\] 
The subclassification according to sy~tactic 
realization has "to uF~ 1~e cat feature as J J% 3.2.1. 
4. Conclusi_on 
The D-stzucture is of major importance for an M\]~ -,o 
system 8rm\] a ~mreful liz~gt~stic definiton of this 
level .':~hould be ~radeo It is i~x~rtant \[.mrticm\[arly 
in a n~itilir~a\]. M2-syste.~ like ~URO2RA as it is a 
precor~\].ition for IS atK1 thus for tz~nsfer. The way 
i/I %£%.i~Ji the l~s~/:uc~ure has been pltesentexl he.re, 
it represents an interface between \].~nguage ~4\[~.c:L- 
fic and intt~rlJ/~gua\] levels. It is inte~clingtml Jn 
t/le s~nse that 1~e rxx~ple/~mt/adjun<~ distinotion 
is regarded i\[~terlirKJually, and \].a~.~znge specific 
insofar as tl~e classification of complements is 
langu, age Slg~i f ic. 

References

(l) Arnold,D., de~ %bm~s,L., Jas~pae~t,L., : 
Eurotxa l,h~ist:ic ST~;cif.ications, 0~nnuscript)~ 
1985 t (=1985a) 

(2) At-hold,D., et al° ~ A Md-i View of t/%e <C,Am,JI' 
I<~a~wo:t'\],.: in F.,Uf~Yi'~, in: Pr(x:'~e~/ings of the Confe- 
r~oe on Theoretical and Methodological Issues in 
Machine %9:anslation of Natural Imngi~ges. Colgate 
University, Ha~itilt~n, N~ York~ August 1985~ pp 
1 o- 14. (=\]985b). 

(3) Az~old et al." qhe ~rot.t.oa Referomce Manr~l, 
(ll~n\[~sc.ript), FeJoz~ary_ 1986 

(4) Engel, Ulrich; Syntax dem deut~chen 'Gegenwarts~ 
sprache, Berlin 1982, (Grt:~lagen der Genn~nnistik 22 ) 

(5) Engel, Ulrich s Schumacher, Helmut~" Kleines 
Valenzlexikon deutscher Verben, El~3ingen 1978~ 
(Forschungsberichte des Instituts fiir deutsche 
Sprache 31) 

(6) Projektgruppe Verbvalenz: Konzeption eines 
W6rterbuchs deutscher Verben, q%~ingen 1981, 
(Forschtn%gsberichte des Instit~ts flit deutsche 
Sprache 45) 

(7) Scmers, Harold: On the Validity of 'the Cc~ple- 
ment-Adjunct Distinction in Valency~Gram/~ir, in: 
Linguistics 22, 1984, pp 507-530 
