TRANSFER AND MT MODULARITY 
Pierre Isabeile & t"~J.l.iott Macklovltch 
Canadian Workplace Automation Research Centre 
1575 Chomedey Boulevard 
Laval, Quebec 
Canada H7V 2X2 
ABSTRACT 
The t cansfer components of typical second 
generation (G2) MT systems do not fully conform to 
the principles o~ G2 modularity, incorporating 
extensive target language information while failing 
to separate translation facts from linguistic theory. 
The exclusion from transfer of all non-contrastive 
information \[eads us to a system design in which the 
three major components operate in parallel, rather 
than in seqnence. We also propose that MT systems be 
designed to allow translators to express their 
knowledge in natural metalanguage statements. 
I. Modularity: a Basic Principle of G2 S~stems 
Modularity is a defining characteristic of 
second genera ti.on machine translation systems 
(hereafter C2 MT). G2 systems are claimed to be 
based on a mode\] in which linguistic descriptions are 
clearly separated from the algorithms and programs 
that actually produce translations. Moreover, in 
this mode\]., the linguistic facts that pertain solely 
to the source language (SL) are supposed to be 
clearly separated from the facts that pertain solely 
to the target language (TL), and from those facts 
that concern the lexlcai and structural contrasts 
between TL and S\],. Such, at Least, are the 
principles of G2 design, as set fortb for example in 
Vauquois \[i \] . This conception of MT gives rise to 
systems composed of three distinct and successive 
phases : a monolingual analysis component, which 
produces a SL-dependent structural description (SD) 
of the input text; a transfer component, which maps 
that SD onto a TL-dependent SD; and a monolingual 
synthesis component, which transforms that SD into a 
TL output text. 
As pointed out by Kay \[2\], this classical G2 design 
offers a number of advantages. Ideally, it should 
allow the formal\[ description of a given language to 
serve the needs of analysis and synthesis 
indifferently. As well, it should allow a given 
analysis or synthesis component to be coupled onto 
other MT systems of simi.lar design to produce 
translations for other language pairs. And finally, 
because the algorithms are independent of the 
particular linguistic descriptions, they too should 
be reusable in other MT applications. 
il. Transfer in T_yj~ical G2 Systems 
Transl!er in G2 systems like TAUM-AVIATION \[3\], 
ARIANE-78 \[4\] and METAL \[511 is essentially a tree- 
transformation system, relating the SDs of two 
complete translation unlts (generally sentences). 
Lexieal units (LUs) are not translated in isolation; 
rather, transfer rules typically test the structural 
environment of each SL LU and, after inserting the 
appropriate TL equivalent, may rearrange that struc- 
ture to accord with contextual\[ constraints imposed by 
the TL LU. Details of formalization aside, transfer 
rules ia these systems will encode facts like the 
following : 
(,.a>  fP\s =+ /v:\ 
,' V S l I 
know savoir 
V NP V NP 1 I 
know conna2tre 
Such transfer rules are usually deterministic: 
each input tree (or subtree) is mapped onto one and 
only one output tree. Consequently, corresponding to 
each SL SD produced by the analysis component, the 
synthesis component will receive one and only one TL 
SD. In order for a correct TL sentence to be 
produced, this unique structure has to be correct in 
all respects. 
This conception makes it necessary for the 
transfer component to encode a lot of knowledge about 
TL grammar. \]for example, if the input tree required 
by synthesis is an Aspects-type deep structure, 
transfer will need to include the equivalent of the 
complete base component of the TL grammar, including 
\].exicaL insertion mechanisms. 
A close examination of the transfer components 
of the three systems mentioned above shows that this 
is indeed the ease. Rules such as (la-b) clearly 
state facts that belong to a description of TL: 
namely, strict subcategorizatLon conditions on the 
insertion of savoir and connaltre. A sizable portion 
of the so-called transfer rules in these systems 
actually deals with TL strict subcategori zation 
requirements. 
In addition to incorporating TL linguistic 
descriptions within transfer, these systems also make 
use of impoverished TL grammars for synthesis. For 
example, none of the systems mentioned accesses a 
full-fledged TL dictionary during synthesis. It has 
long been a truism in MT circles that synthesis is 
much easier that analysis. This is hardly 
surprising, when so nmch of tile work of synthesis has 
been passed over to transfer. This move has at least 
115 
two unfortunate consequences: (a) the burden placed 
on the translator/lexicographer is greatly increased; 
and (b) the translation relation is implemented in a 
highly directional fashion. Analysis and synthesis 
become totally different tasks, each standing in a 
different relationship to transfer. 
In order to remedy these defects, we favour a 
more rigourous adherence to the principles of G2 
modularity: all information which is not strictly 
contrastive should be removed from the transfer 
component and returned to synthesis, where it 
belongs. This would allow G2 systems to fully 
benefit from the advantages of modularity that Kay 
has noted. It would also ease the burden on those 
who must write the transfer rules. 
On the other hand, such a move would make the 
transfer phase non-deterministic, and this may lead 
to severe efficiency problems. If, as in typical G2 
systems, transfer and synthesis are applied in 
sequence, communication between the two components 
will be effected through tree structures representing 
complete translation units. The use of non-determin- 
istic rules at transfer would mean that local trans- 
lation ambiguities would then generalize to these 
complete units. For exalaple, if tranfer no longer 
selects between savoir and connagtre for the transla- 
tion of know, or between cheveux and ~ for the 
translation of hair, then at least four different 
French SDs will be transmitted to the synthesis 
component for any SL SD containing both know and 
hair. 
One possible solution to this problem is to 
have analysis, transfer and synthesis operate in 
parallel rather than in sequence, while providing the 
three components with the means to exchange partially 
specified tree structures at each stage of the 
processing. This is probably what Arnold et al \[6\] 
have in mind when they suggest that transfer should 
be viewed as a relation between two generating 
devices. 
We have recently begun developing a system that 
seeks to implement such an approach. We have chosen 
PROLOG II (Universit~ d'Aix-Marseille) as the 
implementation language, in order to take advantage 
of the handling of partial information in a time- 
independent manner that is offered by unification- 
based formalisms (cf Kay \[2\]). PROLOG II also has 
built-in facilities for parallel processing. Though 
our system is still in the embryonic stage, it can 
already translate in both directions between English 
and French, or simply enumerate pairs of sentences 
that are translationally equivalent. 
III. Transfer Rules as Metalanguage Statements 
Assuming that the scheme outlined in the prece- 
ding paragraphs takes us somewhat closer to the 
ideals of G2 modularity, we can now ask what content 
is left for transfer rules to express. 
Since a good measure of compositionality is a 
prerequisite for transfer to be possible at all~ the 
tranfer relation (trf) has to be recurslvely defined 
in terms of relations between progressively smaller 
subparts of the SDs representing complete translation 
units, up to the level of basic LUs. Depending on 
the degree of compositionallty involved, three 
classes of rules can be distinguished. 
116 
First, there will be rules of purely 
compositional translation, ie rules which define the 
translation of some type of unit as a function of the 
translation of its parts. Rule (2), for example, 
states that the translation of an NP is a function of 
the translation of its components. 
(2) trf(np(Det, Adj, N), np(Det', N', Adj')) 
:- trf(Det, Det') 
trf(Adj, Adj') 
trf(N, N'). 
Second, there will be rules dealing with purely 
non-compositional cases, such as those of (3): 
(3) a) trf(n(chair), n(chaise)) 
b) trf(v(know), v(savoir)) 
trf(v(know), v(connaltre)) 
c) trf(nom(n(pressure), n(gauge)), 
nom(n(manom~tre))). 
Example (3a) states a simple correspondence between 
basic LUs. (3b) illustrates a case where transfer 
leaves a choice for synthesis, as discussed above. 
Example (3c) is meant to illustrate the possibility 
of purely non-compositional translation involving an 
SL phrase. From the translator's point of view, 
pressure gauge is a unit; SL-internally, however, we 
see no reason to consider it an idiom (cf the possi- 
bility of conjunction, as in pressure and temperatnre ~). 
We need a third type of rule to handle partly 
compositional translation, " ie cases in which the 
translation of some but not all of the sub-phrases 
are required for the translation of the whole 
expression. To express the restructuring that takes 
place when miss is translated as manquer ~, for 
example, we would need a rule like (4): 
(4) trf(s(Subj, vp(v(miss), Obj)), 
s(Subj', vp(v(manquer), pp(p(~), Obj')) 
:- trf(Subj, Obj'), 
trf(Obj, Subj'). 
Notice that in the transfer rules given in 
(2)-(4), just as in those of G2 systems in general, 
translation facts are cast in terms of the specific 
linguistic theories embodied in the analysis and 
synthesis modules. This has a number of serious 
disadvantages: (a) in order to be able to directly 
contribute to the development of the transfer 
component, the translator/lexicographer must 
assimilate all the details of the specific linguistic 
theories on which analysis and synthesis are based; 
(b) descriptions of translation facts have no 
independent status with respect to the analysis and 
synthesis modules: changes to the linguistic 
descriptions of SL and TL will likely force changes 
to the transfer rules, creating potential consistency 
problems; and (c) translation rules are unlikely to 
be portable to other systems. 
Natural bilingual dictionaries state 
translation facts in an altogether different manner. 
The information they contain is expressed in what 
essentially amounts to metalanguage statements, like 
the following: 
(5) a) "chair" translates as "chaise" in French 
b) "pressure gauge" translates as "manom~tre" 
c) "sb misses sth" translates as "qqch manque 
qqn" 
Translators obviously manage to express a good deal 
of translation knowledge in terms of such statements. 
One way to ease the burden of the translator involved 
in MT would be to build into the system a capability 
to automatically interpret statements like those of 
(5) in terms of rules like those of (3)-(4). In 
fact, G2 MT systems already possess much of the 
knowledge required to build SDs out of the quoted 
expressions in (5). For (5a), little more than SL 
and TL dictionary lookups are needed to produce (3a). 
To handle cases such as (5b) the system wouid only 
have to access its regular grammar rules. Obviously, 
other cases will arise where such a simple scheme 
would lead to the generation of spurious rules. For 
example~ if not supplemented by additional cons- 
traints, an equivalence might be created between the 
English adjective ~ and the French noun lumi~re. 
However, it does not seem difficult to imagine moti- 
vated principles that would automatically take care 
of such incorrect equivalences. 
Cases of partly compositional translation raise 
more interesting problems. In order to produce (4) 
from (5e), relations must be established between the 
sub-parts of the two quoted expressions. Sb, stl!, 
~ and j~l are clearly meant to be interpreted as 
meta-variables. Sl!) must be linked to ihg~ , so that if 
the first takes some expression as its value, then 
the second takes tile translation of that expression 
as its value. The clue to this link, of course, is 
the fact that ~ is itself the translation of sb. 
We are currently working on a scheme that will 
interpret translation rules expressed as simply as 
those in (5), and allow them to be fed into the 
transfer component of the MT system mentioned at the 
end of section II. Though many difficulties remain 
to be resolved, the research program we have 
described does have the merit of addressing the 
shortcomings that result from the insufficient 
modularity of G2 systems. At the very least, we hope 
to stimulate theoretical reflection upon the 
translation component of MT systems, which has not 
until recently received the attention it deserves. 

References 

Vauquois, B., La traduction automatic~__~ 
Grenoble, Documents de linguist:Lque quantitative 
29, Paris, Dunod, 1975. 

Kay M., Functional Unification Grammar: a Forma- 
lism for Machine Translation, Proceedi~ of 
COLING-84, 75-78, 1984. 

Isabelle P., Bourbeau L., TAUM-AVIATION: Its 
Technical Features and some Experimental Results, 
Computational Linguistics , ii:i, 18-27, 1985. 

Vauquois B., Boitet C., Automated Translation at 
Grenoble University, Computational\[ Linguistics, 
ii:I, 28-36, 1985. 

Bennett W., Sloeum J., The LRC Machine Transla- 
tion System, Computational Linguistics , 1\]:2-3, 
111-121, 1985. 

Arnold D., & a\]. A MU 1 View of the ~C,A>, T 
Framework in EUROTRA, Proceedings of the Con- 
ference on Theoretical Issues in Mac|line Transla- 
tion of Natural Lansuages, Colgate University, 
1-14, 1985. 
