Linguistic Bases For Machine Translation 1) 
Christian Rohrer 
Institut ff~r Linguistik 
Universittlt Stuttgart 
Keplerstraf3e 17 
7000 Stuttgart 1 
My aim in organizing this panel is to stimulate the discussion 
between researchers working on MT and linguists interested in 
formal syntax and semantics. I am convinced that a closer 
cooperation will be fruitful for both sides. I will be talking 
about experimental MT or MT as a research project and not as a 
development project.\[l \] 
A. The relation between MT and theoretical linguistics 
Researchers in MT do not work with linguistic theories which are 
'on vogue' today. The two special issues on MT of the journal 
Computational Linguistics (CL 1985) contain eight contributions 
of the leading teams. In the bibliography of these articles you 
don't find names like Chomsky, Montague, Bresnan, Gazdar, Kamp, 
Barwise, Perry etc.\[2\] Syntactic theories like GB, GPSG, LFG are 
not mentioned (with one exception: R. Johnson et al. (1985 0.165) 
praise I.FG for its 'perspicuous notation', but do not (or not 
yet) incorporate ideas from LFG into their theory of MT). There 
arc no references whatsoever to recent semantic theories. 
On the linguistic side one notices a similar lack of interest in 
MT or in the theory of translation in general. In the latest 
books on G/3, LFG or GPSG one will look in vain for any references 
to the theory of translation. Or, to give another example, look 
at tile research goals of the recently created Center for the 
Study of Language and Information (CSLI) at Stanford. The aim is 
to develop "integrated theories of language, information and 
computation". (CSLI Report No. 84-16, p.2). At CSLI philosophers, 
linguists, computer scientists and psychologists collaborate 
intensely on numerous projects, however MT is not among them. 
There is only one notable exception (M. Kay) which confirms the 
rule. None of the other 40 staff members has ever been involved 
in an MT project. 
This does not mean that there is no research in translation going 
on at all, quite on the contrary. A growing number of countries 
are creating chairs in the theory of translation, there are 
symposia on the theory and practice of translation (L. Grlthls et 
al. 1978), and there exist innumerable papers on practical 
problems of translation, which are written primarily for 
interpreters and teachers of a second language. The efforts in 
tiffs area can be characterized by key words like contrastive 
linguistics and literary theory. Within contrastive linguistics 
the emphasis lies on lexicology. (Typical examples which are 
discussed are pairs like Engl. to know, Fr. savoir, connattre, 
Ital. scala, Germ. Leiter, Treppe.) Yet, there is little detailed 
formal work on eontrastive syntax. (For an exception see Hawkins 
(1986)). The results in eontrastive linguistics ira general and in 
contrastive lexicology in particular will certainly be of great 
value for writing transfer dictionaries and specific transfer 
rules. The research undertaken in this area, however, has not 
produced sufficient results to build a coherent formal theory of 
translation on. Finally there is the philosophical tradition, 
best illustrated by Quine's famous 'Meaning and Translation' 
(1959) and Montague's 'Universal Grammar' (1970). In this context 
translation means translating from a natural language into a 
logical language. This sort of translation has had an enormous 
influence on semantic theories used in linguistics but again 
there are very few connections with MT (for an exception see 
Landsbergen (1982)). 
(l) 
(2) 
(3) 
(4) 
(5) 
(6) 
Why is there so little interaction between MT and theoretical 
linguistics? 
Most MT projects were set up as development projects. There 
was just not enough time or money to do basic research. 
MT projects were often directed by computer scientists who 
underestimated the complexities of natural language. 
Analysis grammars for MT were often written in a formalism 
which made communication with the non=initiated impossible. 
Theoretical linguists thought and still think that MT is too 
difficult, that high quality MT is impossible. Some argue on 
the basis of syntax: if we don't even have a satisfactory 
syntax for comparatives in English, how can we translate 
English comparatives into German. Some argue on the basis of 
semantics: satisfactory semantic analyses exist only for very 
small fragments of natural languages. How are we to translate 
German tense forms into French tense forms, if nobody has yet 
presented an adequate semantic description for French tenses. 
Linguists seem to Imve a one-sided (lop-sided) perspective, 
working nearly exclusively on analysis. In LFG one parses 
sentences and builds up functional structures, in GPSG 
sentences are parsed and translated into formulas of 
intensional logic, hardly anyone knows how to generate from 
f-structures or from logical formulas. 
If there is little research on generation, there is even less 
on transfer. Under the influence of Chomsky and (to a smaller 
extent) Moatague linguists started the search for Universal 
Grammar. Transfer, however, presupposes the comparative study 
of two (or more) langua~les. Linguists working with recent 
syntactic theories don't seem to be interested in questions 
like:\[3\] Is language A syntactically more ambiguous than 
language B? If a string S is ambiguous in language A, is 
there a string S' in language B which presents the same 
number of ambiguities? Is it really easier to translate 
between two languages which are closely related (say Italian 
and French) than between distant languages (say Japanese and 
German)? Under what conditions is it possible to translate on 
the basis of syntactic structure, and under what conditions 
do we have to use semantically based transfer rules? 
C. What can we do to stimulate interaction between MT and 
theoretical linguistics? 
1 would like to argue that MT should be based on a linguistic 
theory which can be expressed in a formalism whose formal 
properties are well understood, in other words, a formalism for 
which there exist results on decidability, generative capacity 
and recognition complexity. The linguistic theory itself, 
independent of its formalisation, should be well accepted within 
the linguistic community. Possible candidates for such theories 
are GB, GPSG, LFG etc.\[4\] What are the advantages of working with 
such a theory? 
353 
(1) Advantages for the linguist 
The MT linguists can use directly analyses which have already 
been proposed for the phenomena he is trying to describe. If a 
linguist writes a GPSG grammar for a fragment of English and 
wants to include coordination, he can just use the analysis 
proposed by I. Sag et al. (1984). 
He can discuss the insights expressed in his own analyses with 
linguists outside the MT groups. 
Since the formalisms which are used for expressing linguistic 
theories are closely related (Shieber (1985) has even shown that 
the formalisms, in which the different UGs are expressed, are 
interredueible to a large extent), a particular analysis can be 
transferred from one theory into the other. The treatment of 
long-distance dependencies (wh - movement) in LFG for instance 
looks formally very similar to that in GPSG. 
Linguists working outside of MT could use the implementation of 
LFG, GPSG etc. as grammar testers. 
Theories like LFG, GPSG offer a well-defined interface to 
semantics. To name just one particular problem we are working on 
at the moment, it thus becomes possible to test the adequacy of 
an analysis of the French tenses and of the English tenses by 
writing transfer rules which map a semantic structure, which 
represents the temporal information contained in a small French 
text, into the corresponding semantic structure for English. 
In UGs the lexicon plays a central role. A lot of linguistic 
information is located in the dictionary entries. The rich 
lexical information collected by contrastive linguists could be 
incorporated into a transfer lexicon written according to the 
principles of UGs. 
(2) Advantages for the computer scientist and linguist 
If linguists can write their grammars in a formalism whose 
mathematical properties are well understood, then the programmer 
will have fewer problems implementing the grammar. Results on 
formal properties of grammar formalisms can guide the programmer 
in his search for appropriate algorithms. Furthermore, in the 
same way in which a linguist working within LFG can benefit from 
the intuitions expressed in linguistic analyses within GPSG or 
GB, a programmer implementing LFG can benefit from insights 
contained in implementations of related formalisms. It is 
therefore not surprising that F. Pereira (1981), W. Frey and 
U. Reyle (1983), L. Karttunen (1985) and J. Calder et al. (1986) 
all use "gap threading" for implementing long-distance 
dependencies. The authors are working with different theories 
(Extraposition Grammars, LFG, Helsinki Unification Grammar 
(HUG) and Unification Categorial Grammar (UCG)) and different 
programming languages (Prolog and Lisp) and nevertheless the 
algorithm they chose is the same. Let me give you a further 
example. In the Prolog implementation of LFG by D6rre and Eisele 
(this volume) there are some problems with negative and 
disjunctive constraints. Such constraints are linguistically well 
motivated but difficult to implement. Now if anybody within the 
UG community comes up with a good implementation of negative 
and/or disjunctive feature specification, then this new 
implementation can be incorporated. 
Results on recognition complexity may help to locate the source 
of the complexity and suggest ways to reduce complexity either by 
modifying the linguistic theory or by showing that the "worst 
case" is linguistically (empirically) irrelevant (in other words, 
the "worst case" will never arise in practice). 
A famous example, where the linguistic theory was changed after a 
proof of its generative power had been found, is Peters and 
Ritchie's work (1973) on the generative power of TG and the 
subsequent change of the theory (constraints on deletion rules). 
354 
For GPSG Peters and Uszkoreit (1983) have examined the role of 
metarules and essential variables (variables which range over all 
strings of terminal and nonterminal symbols). They proved that 
unconstrained Metarule Phrase Structure Grammars have Turing 
machine power. This result has led to intense research for 
constraints that are linguistically motivated and descriptively 
adequate. 
For LFG there is a conjectured result by Berwick: certain 
Lexical-Functional Grammars can generate languages whose 
recognition time is very likely computationally intractable, at 
least according to our current understanding of algorithmic 
complexity." (Berwick 1982, p.98) 
Conclusion: Basically any argument in favor of a declarative, 
simple, expressive, flexible linguistic formalism with a rigorous 
mathematical foundation can be adduced here. 
D. Possible objections against using Uniflcatian Grammars o1' 
other recent linguistic theories for MT 
Are there linguistic phenomena, which make it impossible to use 
UGs for parsing and generation? Someone might argue that in MT 
there exist specific problems, which have never been taken into 
consideration by theoretical linguists, and which cannot be 
handled with UGs. Personally I hope that whatever additional 
problems may arise can be handled in the transfer component. If 
UGs provide us with the tools to write adequate grammars for 
individual languages, why should it not be possible to use these 
grammars for MT? 
R.Johnson et el. consider the possibility of assuming a current 
linguistic theory and implementing it directly for EUROTRA. They 
reject this solution "because there is not sufficient practical 
evidence of a single theory that encompasses translations between 
all pairs of the Community languages" (1985, p.165). This 
rejection can be interpreted at least in two ways: (a) there is 
no linguistic theory in which all the languages of the Community 
have been described, (b) recent linguistic theories have not been 
used for (automatic) translation between all pairs of the 
Community languages. Of course, both interpretations are correct; 
however, on the basis of this argumentation one could reject any 
existing linguistic theory as insufficient for MT. This 
reasoning, however, would force us also to reject any new 
linguistic theory because there would be even less "practical 
evidence". 
So far only fairly small grammars have been written and 
implemented in the UG formalism. What will happen if these toy 
grammars develop into big grammars with very large dictionaries? 
Won't we need special control mechanisms? Since the UG formalism 
is well understood, I can't see why one should have more problems 
if the grammars get bigger than with any other formalism. We hope 
that the contrary will be the case, namely that there will be 
fewer problems. 
There is still another side to the problem of "toy grammars". 
Theoretical linguists have been working with a fairly narrow set 
of data. An ambitious young linguist today is more likely to 
propose the 127th version of the passive construction than to 
write a detailed grammar for complex NPs like "die Eiufuhr von 
Jute aus Ll~ndern der dritten Welt zum Ausgleich des Nord-Stid- 
Gefalles". Will theoretical linguists ever sit down and actually 
write big grammars? Why has there been no grammar writing project 
since the attempt by R. Stockwell et al. (1968) to collect all TG 
analysesintoonegrammar? 
UGs have been used for analysis, but there are practically no 
results for generation. This lack of experience in generation is 
a serious drawback for the whole field of computational 
linguistics. Fortunately the situation is gradually changing. The 
organizers of the International Workshop on Language Generation 
(Stanford 1984) write: "More papers in this subfield \[generation\] 
have been published in the last four years than in its entire 
previous history." 
hi MT generation has played a minor role. In a system like SUSY 
(Maas 1985) for instance there is no explicit grammar for the 
generation of German. The JapaBese Government Project for MT has 
a well developped generation component (Nagao M. et al. (1985)), 
but it is difficult for the linguist to find out how this 
generation grammar would look in a declarative format. So, even 
if there are no results on generation with UGs we are not much 
worse off than anybody else working with another linguistic 
theory. 
The most serious objection 1 see concerns the transfer component. 
I have been advocating the use of linguistic theories which can 
be expressed in well-defined formalisms, bnt so far there doesn't 
seem to exist a theory of transfer for UGs. Well this is not 
entirely correct. M.Kay claims that his Functional Unification 
Granlmar (FUG) "provides an opportunity to encompass within one 
formalisna and computational system the parts of machine 
translation systems that have usually been treated separately, 
notably analysis, transfer and synthesis." (Kay, 1984, p.75). 
E. Conclusloz~ 
MT as a research project cuts across traditional disciplinary and 
institutional boundaries. Theoretical linguists, lexicographers 
and computer scientists must cooperate in order to solve some of 
the problems of (machine) translation. GrammllrS and dictionaries 
used ill MT should be based on linguistic theories which are 
widely accepted in the linguistic community. The linguistic 
theories should be expressed in fornlalisms that fulfill the 
requirements of a good programming language. An open and 
transparent discussion between linguists and computer scientists 
will lead to a realistic assessment of the possibilities and 
prospects for MT. We should avoid a reinterpretation of the motto 
'traduttore traditorc'. 
Footnotes 
1 Support for work on this paper w,~ provided by a grant from the BMFT. 
2 The choice of names and of theories is for exempliflcatory purposes. We do not 
try to be exhaustive in any sense. No evaluation ia implied by the presence or 
absence of any particular name, theory or publication. 
S Fortunately the ~fituation is changing now. Since Chomeky's Piss Lectures there 
is a growing interest in comparative syntax. Maybe some day a theory of 
translation will emerge from the theoretical and practical insights from these 
comparative studies within the framework of GB. It is too early to construct 
today a theory of translation on the basis of GB, unless this theory made 
reference to GB only in a metaphorical senne like E. Nida did in his proposal 
to use the Aspects model for translation and to do transfer on "deep 
structures" (Nidr. (1969)). 
4 Ia the following \] will use the term Unification Grammar (UG) aa hyperonym for 
GPSG, LFG, FUG, IIPSG etc., i.e, for grammars which have a contextfree skeleton 
and an operation of unification. In grammars of this type~ syntactic rules and 
lexlcal entries can be expressed as sets of attribute-value pairs. The value 
of an attribute can again be a set of attributes and values. Later on I will 
argue for the use of UGa for MT. Thia does not mean that GB or Joshi's Tree 
Adjoining Grammars could or should not be used for experimental MT. I just 
happen to have a personal preference for UGs. 

References 

Berwick, R. C. (1982) Computational Complexity and Lexical Functional Grammar, 
American Journal Computational Linguistics 8, 3-4, pp.97-199. 

Calder__J. et aL (19861, Problems of Dialogue Parsing, MS., Center for 
Cognitive Science, Edinburgh. 

Chomsky (1981), Lectures on Government and Binding: tile Plea Lectures, 
Dordrecht. 

Frey W., Reyle U. (1983), 'A Prolog Implementation of Lexical Functional 
Grammar as a Base for a Natural Language Processing System', Proceedings of the 
First Meeting of the Association for Computational Linguistics, Pisa, pp.52-57. 

Gazdar J. et al: (1985), Generalized Phrase Stueture Grammar, Oxford. 

Griihs L. et aL ~ (19781, Theory and Practice of Translal;ion, 
Nobel Symposium 39, Stockholm. 

tIawkins J. (19861. A Comparative Typology of English and German Uzdfylng the 
Contrasts, London, Sidney. 

Johnson R. et al. (1985), 'EUROTRA: A Multilingual System under Development', 
Computational Linguistics, ll, pp.155-169. 

~A.K. Lev L.S. Takahashi M= (1975), 'Tree Adjunct Grammars', J. Getup. 
and Sys. Sc. 10, pp.136-163. 

~resnan J~ (19821, 'Lexical-Functional Grammar: A Formal System for 
Grammatical Representation', in: Breanan~ J. (ed.), The Mental Representation of Grammatical Relations, Cambridge Mass., pp.17S-281. 

Karttunen, L. (1985), HUG : a Development Environment for UGa, MS., Stanford 

Kay, M. (1984), Functional Unification Grammar: A Formalism for Machine 
Translation', Coling 84, pp.75-78. 

Landsbergen J. (1982), Machine Translation Based on Logically Isomorphic 
Montague Grammars, Coling 82, pp.175-181. 

Mass II. (19851, 'Zur Entwlcklung yon SUSY-11', in: U. Klenk (ed.) Kontextfreie 
Syntaxen und verwandte Systeme, Linguistische Arbeiten, Tabingen, pp.107-llT. 

Montague, R. (19701, 'Universal Grammar', Theorla 36, pp.373-398. 

Nagao, M. et. al. (1985), tThe Japanese Government Project for MT', CL 11, pp.91-110; 

Nida F,. (1969), 'Science of Translatlon', Language 45,3, pp.483-49S. 

Peters, P.S., Rit~ R.W. (1973), 'On the Generative Power of Transformational 
Gramraars ~, Inf. So. 6, pp.49-83. 

Pereira F.: (1981), 'Extraposition Grammar', Computational Linguistics, pp.243-256. 

Quinem W. (1959), 'Meaning and "translation', in Brewer, 11. (ed.), 
On Translation, Cambridge Mass., pp.148-1~'2. 

Shieber S. (19851, 'Separating Linguistic Analyses from Linguistic Theories', 
Ms., Stanford. 

Stockwell R, et al. (19681, Integration of Transformational Theories on English 
Syntax, Los Angeles. 

Uszkoreit H. Peter~ (19831, Essential Variables in Metarulesp SRI Tech. 
Note S05. 
