Reasons Why' We Use Dependency Grammar 
Eva Ha J J. ~, ovh. 
Charles I\[niver sit y 
P:eague ) Czechoslovakia 
I. First, I would like to say why I do care 
E.~mmmr formalisms. The point is not only 
tha%-I was %ra.ined as a student of linguist-- 
ics and %hat I always ha.re been interested in 
theor~l Lcal ling'uJ, stics, but in the present 
context a.\]SO ) and main\].~ ~ that natu:e&:\[ \].ang- 
uage processing systems mostly a~'e too 
complex to be bui\].t, modified, comp\].emented, 
enriched.,., o , without a solid theoretical 
baekground~ As Prof.Nagao puts it, theory Js 
important and valuable for the explanaJ;ion 
~).nd undcrsta.nding~ a. \]angu&ge processing me.-- 
de\]. ~heu\].d be understandable on the bach- 
ground (,f a powerful \].inguistic theory. On 
the other hand~ :\[ would lille I;o stress that 
if lin~gtlJllties wants -to be nsefu\], a~l{\] to lllaqke 
safe its own perspeet:ives, then it has to be 
useful for :l.l.ng!\igi~J_~ q en__q6~ineeri_q C. Thie means 
for me that; the theory has 9o bc not o.nly 
adequate, hut also economical and merlu\]ar. 
The t},s\]~ of the theory is to offer a re\].ati- 
vel:7 complete framework, which never captures 
all th.e detai\].s in their specific and often 
exeept:i, ona\]. eharac%er , bu~ whic}! , am \](aren 
Jensen notes in her point (7) ~ o\['f(.'rs a 
maximal coverage , i.e. which ceYit~,irt.q means 
necessary and suff:':i.eien% for }iandlint< a\]\]. 
:\]uch detai\]s am :Far as they ~w'e re\]ev,qx~t for 
'the /~iven app,\]ication fie\]d, In "tl~:i.s respect, 
%he thec>re-tica\] framework e~m be compared to 
a fi \[;herman" S net ~ which nee(I n()'~ be llsed 
'whole, if this is no't; necessary for ~be given 
pool; some of the meshes may be left unused 
:il~ the \]~eag or ashore, 'but in a larger pcol 
they may be useful. The most important peint 
is that the meshes are there, and we know 
wbere \[;hey are and for what purpose "they 
might be 'useful° 
2o The formalism is /tot the only import~ln~ 
ingredieni; of a.n NLP system, and it Is not 
interesting here for its own ~akeo :\[% is true 
that %he bottleneck of an NLP system Is in 
hanjling-the "dirty" exceptional cases, 
ra'ther "i;han the cases dlreet\].y fitting into 
the main body of this or tha,% theory. As a 
ma-l;%er of fact, using any theory, we have. %o 
.face such intricate but common examples as 
Kirschner's cases of target language amblgui- 
ty (or vagueness) corresponding e.g. to that 
of F, nglish ,i.n_~\[-forms , or the long but 
lexiea\]ly bound sequences of nouns in 
termino\].ogioal noun groups, or a procedure 
translating \].exieal items by modifying the 
productive affixes of int.ernational terms of 
(\]reek and \],atin origin and other "emergency 
rules" ensuring that 6t leas% an approximate 
(at leatvt partially readable) output will be 
ae h i e w~, (',. o 
3o A\],\] this heavily sUppo:P't;s the a:egumen% 
that t})e theoretical framework sheu\].d be 
relatively eeonomi.ta!o If two theories~ which 
in the given conditions cannot be 
systematica\].ly compared with respect to their 
oover~ge ~ seem %o exhibit a more or less 
equal degree of ad.equaoy ~ 1;hen %be simpler 
one will be preferred. With this framework it 
is more probable that %here would be space 
enough for cap%firing all the necessany 
"dirty" del;ails illustrated in Section 2o 
Dep_ende__nn£;f. grammar meets %his condition, 
since its trees (or even the more comp\]ex 
representations required for a treatment of 
coordinated structures) are much s:lmplcr than 
any kind of phrase--structure baecd represen-. 
ta%ions: it is pom~ible to use here comp\].ex 
(although strictly limited) node labc\]s and 
i;hus dJstlnguish between the syntactically 
free lexiea.\] oectn, renees and function mor- 
phemes (the latter used not have nodes of 
%heir own); moreover, the representations can 
be conceived of without nonterminal symbols 
(if the kinds of dependency re\];v~Jon, similar 
e.g. to %hera roles , cases , an@ vs\]ency, a~.-~'e 
denoted by the labels of the edges)° 
II. \]~e.'~ J (\] e :3 the ad~, anta/<e of an ee onomJ (:;~\] 
description , a dependency based proee3ure m'~y 
derive from %he \]exical data ,'-t\].\] the Jnfo:pn,-- 
,a, tion necessary fo:P %be ~>~signmen{; of t, heLa.- 
-roles (va\].ency slots) and ()the> com\])Jeri!erlt-- 
&tions hy a. head° This tn'eperty is c~,:\[;--emsly 
valuable for %he formu\].~.Ytion of a ~ }'ser; 
once. 'the verb is identified ;rod 71ooked-.up for 
in the dictionary, it predJ C'I;~ lhrtny pieces 
of informat.ion necessary for the :\[denLJ.:l':i.c-- 
a%ion of the functions of nomina.\] (Jne:luding, 
preposition&l) cemp\].exes (a,'~ comn?ementa~ions 
of the given verb). A,~ fa~' as our own 
experience wi I;}1 the bu:l. Jd--up efl ~ `lJ 71",\] ~ % -,\[ %~; ~ '\] ; 
goes, th~::~ ho:l.d:~ true not; on:J5, wher! working 
w~th highly inflectional \]an~:uages (1,he word 
ord.er oJ" wbJcb J.~ ~ conseqllen~,ly ~ nst {~::l'~igl.- 
matlca\]\]v...,, determined) , b'ut ~-<Tso e.go .f'er' ;~ 
par,qcr of \]'\]nglish. 
'Phe information on "the,' " thet;a.--:r'o\] e s" 
belongs to olle dimension of \[.he l:-ree , ~\]ame\].y 
the vertical one. The seeorl(\] (\]ifne~l~\]ioH ~ }ler:i. 
zenta\] , :is \].eft free 1;o denote J\}i!{ :\[;qJ,}j\[q - 
-focus articulation and {;he deep word orc\]e:u 
which eontributes to the deh(Ymn:inal;~on oi' 
operator scopes J.n serllall~,:\] c in-~e:~'py'etat\] o~ o 
(,oar llnation (at we\].\], as ;q~po S j I; J on) 
con'zd;ih\]tes a third dimension , ,~{iir<'t\]}lr\].v ;~.s 
with other approaches ; rules b~,.,e \].een 
formu\].ated (by P\]htek, Sga\]I, PetPev'i~!) w~ic'.}l 
handle hierarchical as well a,~ sequer<i;:ia\] co.- 
ord inat;ed strlletllre s in ~}le \] J ne~rJ zed sent.- 
ence representations. 
5. The arguments brought forward agair~sL (le-. 
pend.ency as the ImsJ. n of an integrated ling.- 
uistic description arc not so weighty as they 
may seem at first blush: anaphoric relations 
and the relative closeness of ad junets er 
other modifications can best be so\]red on %he 
basis of an account of -topic and focu:~ , and 
instead of "double dependency" with TlredJc- 
afire complements we use a broader rto'l;J.on of 
manner adverbial° Our framework differs from 
(l.aifman's in imposing no specific limits on 
the sequences of non-terminals used Jn the 
derivations of representations (Dorota) o 
The framework has been used wit}~ ~(h, ant,. 
age ~xs the basis of the natural \] anguage 
processing systems built in our group~ }~e J% 
parsers for machine trans\]ation systems 
(Engl.l.sh-bo-6zeeh, Czech-to-Russian) or for a 
system modelling natural language comprehens- 
ion (TIBAQ) . 
4 51 
