Why implemento~-~,~ ~f P~actic:fl NLP Systems Can net Wail 
fe~' Li~guis~ie Th~;or~cs 
Remnrh.s and 'lrhescs 
~)ie~mar Recs~mr 
c/o Project, GENES~S 
GW|I) 0'4, ~)(~livostro :~5 
~) 6100 D~vmstadt 
W(;,u$ (\]ermany 
A) Ther~ is no doubt~ the idea of taking an 
acc<,pted ling~istic ghee.<ff (lt) a}~ the basi~ 
of ~he :L~%~lca~entation of ~ practical natara\] 
\],angu~gc~ pz'ecessing (n\].p) system is ve)'f 
atix:active ,, 
But t In reality there are no .I.iztguistic 
theo:<J,es available that may be ,',~.imp\].y "taken 
f:com the shelf" and applied in such a woy0 
Some ~:easons for \[:.him 
There is often a mismatch between the 
\]?roblems dealt with extensively in a given 
\],inguJ stic theory and their frequency of 
cccr~rence i\[~ real data° 
On the other hand the chosen linguistic 
theory OK gran~aar may tell you only little or 
nothing abotrt phenomena that are pm"edomina.nt 
in your application (e.g. large nominal 9roups 
in technical texts), 
-~ Practical systems need a substantial if :c, ot 
co~ap\]eyte coverage of the syntax of the nl they 
aye bui \]. P for° ~\[~. general the language 
f<'agm<ants covered by linguistic theories a.re 
not ~u~_f \[\[icJ.ent ° 
-. Even if you a~:e able to sta~:t with 
.~tructures .i n accordance with a \],inguistic 
theory when building a practical n\].p system 
you will. have to augment the forma\].ism~ e,g0 
as soon as you have to prec@ss real data and 
not only text book cases a lot of conventional 
n\]aterial has to be taken into account as we\].l: 
format of dates~ measurements~ etc. 
.- ),inguistic theories very often have been 
eztensively applied to problems 'taken from a 
specific nat~a\] language ( in most cases .~ 
F.nglish)~ To transfer those theories to other 
\].anguages with different problems may be hard 
(if not impossible). 
B) Another difference between theoretical and 
ongineering problems lies in the fact 'that 
building a practical nlp system always forces 
you to take a lot of additional constraints 
into account that are negligeable from a 
theorist' s viewpoint : 
Fo~: the implementation, usage and 
maintenance of practical nlp systems dedicated 
software tools are needed e.g. fox editing, 
debugging and update of grammars and lexica. 
Practical nlp systems are often evaluated by 
users (and ~:efereem) that are non-linguists. 
,According to our experience those people tend 
to take "mupe~'fi.cial" aspects 'that have little 
OK no relevance for linguists as a basin for 
their judgements; eog° if oK if 'not a system 
fo~: German is able to handle umlauts correct\].y 
or how you may type in to a system for 
japanese, 
Practical nlp ~-}ym terns inevitably need a 
substantial \].exicon and - due to the 
generativeness of nl with respect to 
vocabulary - additional techniques for 
handling unknown lexical items. 
C) Some ef the abeve helds especially fo~: wo~k 
in nl gene~oation: 
- Linguistic theories are primarily concerned 
with aspects of (syntactic) analysis° 
\[~ome ef the hard problems for analysis (e0g, 
~?\[I? attac|ument) are not equally problemat:i,c in 
generation. 
In general nl generation is an 
interdisciplinary task involving a lot of non,° 
linguistic decisions ~ 
D) Theorists tend to ~'est~cict theist app~'oaches 
to the very techniques available withi~ their 
theories. In practical nlp systems it may be 
fruitful to f:ceely combine elea~ent~ f~om 
distinct "Zinguistic schools" ~ 'the morpho- 
syntactic front end generator FREGE \[Emele 87\] 
is a case in point: )?REGE takes functional 
grammatical structures comparable to those 
from LFG \[Kaplan~ B:cesnan 82\] ~ as input and 
working structures but provides means to 
specify constituent ordering as linear 
precedence relations similar to GPSG \[Gazdar 
eta\]., 82\] o 
E ) What about the following analogy ; ~#he 
relation between theoretical, linguistics and 
"language engineering" should parallel that 
between mathematics and (civil) engineering° 
F) Linguistic theory might profit from 
problems attacked in practical systems as well 
because (as with mathematics and problems in 
engineering ) these problems may gi~e 
motivation for future research. Practical wo~:k 
in mu\],tilingual generation .~, i. e o attempts to 
generate different languages from the same 
semantic structures - may for example give a 
new perspective fox' work in contrastive 
linguistics ° 
To sum up; One should not wait fbr theoretical 
solutions since theory might not attack some 
problems until they are of practical 
relevance. 
References 
\[Emele 87\] ~nele, M° (1987): "FREGE .. Ein 
ob J ektorientierter FRont-End-GEnerator" ~ in 
K0 Morik (Ed.) "GWAI-87 o- llth German Workshop 
on A~ctif icial Intelligence" ~ Proceedings 
Springext-Verlag 

\[Gazdar et al. 85\] Gazdar, G., ~:. Klein~ G. 
Pullum & I. Sag (1985)~ "Generalized Phrase 
Structure Grau~ar"~ Can~ridge~ Ha~v'ard 
University Press 

\[Kaplan~ Bresnan 82\] Kaplan, R.& J, Bresnan 
(1982) : "Lexical Functional Granuna~'~ A Formal 
System for Grammatical Representation" ~ in 
B:cesnan ~ J. (ed°) ( 1982 ) .~ "The Mental 
Representation of Gra~m~atical Relations" r 
Cambridge (Mass.): MIT Press 
