4. Machine Tr,ansUation 
Martin Kay, Chairperson 
Xerox Corporation 
Palo Alto, CA 94304 
Panelists 
Margaret King, ISSCO 
Jack Lehrberger, Universit6 de Montreal 
Alan Melby, Brigham Young University 
Jonathan Slocum, University of Texas 
A Manhattan project could produce an atomic 
bomb, and the heroic efforts of the sixties could put a 
man on the moon, but even an all-out effort on the 
scale of these would probably not solve the translation 
problem. In one sense, this almost goes without say- 
ing. The first ninety percent of the work that was 
invested in reaching the moon did not get the astro- 
nauts nine-tenths of the way there and the heat gener- 
ated by the bomb did not increase steadily as the date 
of the first explosion approached. The translation 
problem is one whose solution must be reached incre- 
mentally. There will be no dramatic event to signal 
the end of the search and there is no single break- 
through that would assure success. 
The translation problem is real and will in fact rap- 
idly reach crisis proportions unless some action is tak- 
en. The problem cannot be alleviated by better lan- 
guage teaching, greater incentives for translators, or 
improved administrative procedures, worthy though 
these goals undoubtedly are. The only hope for a 
thoroughgoing solution seems to lie with technology. 
But this is not to say that there is only one solution, 
namely machine translation, in the classical sense of a 
fully automatic procedure that carries a text from one 
language to another with human intervention only in 
the final revision. There is, in fact, a continuum of 
ways in which technology could be brought to bear, 
with fully automatic translation at one extreme, and 
word-processing equipment and dictating machines at 
the other. 
The productivity of the professional translator 
could almost certainly be greatly increased by techno- 
logical aids which, though straightforward, are not all 
obvious. Powerful machine aids to translators could 
quickly be available and may be the best way to allevi- 
ate the translation problem in the short run. They will 
not arise as a natural by-product of work on fully au- 
tomatic translation because, for the most part, they 
address such issues as communication among transla- 
tors, identification of relevant secondary material, and 
special editing devices, rather than issues of syntactie 
analysis, pronominal reference and quantifier scope. 
The most valuable resources that a translator has for 
solving difficult problems are the text he is working 
on, other texts like it in the target as well as the 
source language, and his colleagues. At present, ac- 
cess to these resources is haphazard at best. But, 
improving it immeasurably is well within the scope of 
existing technology. 
Another easily identifiable point on the continuum 
is occupied by human-aided machine translation. This 
could be a very different kind of enterprise both from 
fully automatic and machine-aided translation, By 
human-aided machine translation, we mean to refer to 
systems in which the machine, while retaining the initi- 
ative, works with a human consultant, who need not 
be a translator. Once again, the subtleties in the de- 
sign of the system would not reside so much in basic 
linguistic questions as in how to recognize reliably 
when a difficulty of a certain type had arisen and how 
to communicate the nature of the difficulty to the 
consultant in such a way as to elicit a quick and unam- 
biguous response. Especially in the early stages, a 
human-aided machine-translation system intended to 
produce output of high quality might well require at 
least as much work on the part of the consultant as a 
trained translator would take to do the job in the tra- 
ditional way. However, two facts can be set against 
this. First, the consultant would not have to be a 
translator and could quite possibly be drawn from a 
much larger segment of the labor pool. Secondly, 
while the labor involved in translating a text grows in 
direct proportion to the number of languages into 
which it must be rendered, the work required of the 
consultant in such a man-machine team would grow 
much more slowly. Indeed, if those languages were 
closely related, it could be expected to fall off sharply 
as soon as that number exceeded one. 
A substantial proportion of what follows will be 
devoted to upholding the panel's view that it is impor- 
tant for work to proceed in parallel on a number of 
different fronts. While fully automatic translation is 
the most adventurous, it is from this that we stand to 
learn most about language in general, and translation 
in particular. If fully automatic systems can be built 
whose performance exceeds that of present systems by 
even a modest amount, we should profit greatly as well 
74 American Journal of Computational Linguistics, Volume 8, Number 2, April-June 1982 
Martin Kay Machine Translation 
from their practical utility as from the theoretical les- 
sons enshrined in them. Machine-aided translation 
can enhance the translator's productivity, though we 
have yet to discover how much enhancement is possi- 
ble in this way. It could also be a source of invaluable 
information on how translators work. Human-aided 
machine translation can be expected to give better 
results than could be achieved with the fully automatic 
method, since the human consultant can be called 
upon to resolve otherwise unresolvable problems, but 
at an unknown cost. However, there are important 
applications, notably where one text must be translat- 
ed into several languages, where the gains may be 
substantial. 
We have been at pains to make it clear that the 
three methods of involving machines in language trans- 
lation are not essentially different in kind but lie on a 
continuum. Except in a few special applications, some 
of which we will mention shortly, we do not foresee a 
time when translations of any quality will be produced 
without any human intervention whatsoever. In the 
so-called fully automatic method, the human plays the 
role of an editor, or revisor. His involvement begins 
only after an initial draft in the target language exists 
and it is for this reason that we remain content with 
the term "fully automatic". The other two methods 
involve him earlier so that he influences even that first 
draft. 
The methods therefore differ as to how the person 
is involved. They also differ in the extent of his in- 
volvement. If the human partner can influence all the 
decisions that are made, he may be in a position to 
forestall sequences of errors, each resulting from the 
one before, thus reducing the total amount of his con- 
tribution. On the other hand, if a conservative system 
insists on having him confirm even those choices for 
which its own decision methods are substantially ade- 
quate, then the overall extent of his involvement may 
be increased. In any case, the utility of a given system 
in a particular situation cannot be assessed by a simple 
equation. The appropriate utility function involves at 
least the human cost, the machine cost, the quality of 
the result, and the nature of the consumer9s require- 
ments. 
The consumer's requirements are, of course, cru- 
cial. The various types and degrees of automation in 
translation are, as we have seen, positioned along one 
dimension in a space of possible approaches to the 
overall problem. The different types of text and their 
consumers are another dimension and, not surprisingly, 
the two dimensions are far from independent. The 
type of technology appropriate to a problem, and the 
benefits to be expected from it, differ greatly with the 
type of the text to be translated and the use to which 
the result will be put. In the intelligence services, a 
great deal of translation is done for purposes of cur- 
rent awareness. The first priority is to know the sub- 
ject matter of the document. It is also helpful to be 
able to discern the gist of the argument so as to dis- 
cover whether it touches on certain key questions. A 
rough and ready translation, especially if it can be 
done quickly and cheaply, may give an excellent basis 
on which to decide which parts of a document, if any, 
need to be translated more carefully. Fully automatic 
translation, even of quite inferior quality, has already 
proven very valuable in this role. 
Fully automatic translation, or some close relative 
of it, has also proved useful in recent years in situa- 
tions where a sublanguage has come to be used, or 
where one can be readily imposed. Canadian weather 
reports are routinely translated by such a system. The 
system itself determines whether each translation unit 
- approximately a sentence - is within its capabilities. 
If it is, then it produces a translation, which is the one 
that will be used without human revision. If not, it 
presents the translation unit to a human collaborator, 
who makes the translation. We prefer to classify this 
with fully automatic translation because, though the 
machine does not translate everything that is translat- 
ed, the translation it does is done entirely without 
human involvement even at a post-editing stage. The 
machine in fact translates eighty per cent of all trans- 
lation units and readers of the reports prove unable to 
discern which parts were translated by machine and 
which by a human. 
The success of this METEO system comes from the 
fact that meteorologists naturally write in a highly 
constrained subset of English. Fully automatic trans- 
lation has also been successfully applied to the task of 
translating maintenance manuals for machines. The 
success of this does not rest on the existence of a nat- 
urally occurring sublanguage. In this case, the techni- 
cal writers who prepare the manuals learn to follow a 
set of rules intended to ensure that their products will 
automatically translatable by simple means. The rules 
are straightforward and can be learnt in a two-week 
course. The machine translates the whole text without 
outside assistance and preliminary results encourage 
the belief that little or no editing will be required. 
The features of a sublanguage that make it suitable 
for fully automatic machine translation are (1) restrict- 
ed vocabulary, with consequent reduction in the num- 
ber of words with more than one grammatical catego- 
ry, (2) small number of senses for each word in a giv- 
en category due to the restricted semantic domain, and 
(3) restricted syntax resulting from the purpose of the 
text, e.g., instruction manuals may contain only imper- 
ative sentences and weather reports only declaratives. 
It should not be thought, however, that a sublanguage 
is simply a subset of the sentences of the standard 
language. The syntax of a sublanguage may differ 
radically from that of the standard language so that a 
American Journal of Computational Linguistics, Volume 8, Number 2, April-June 1982 75 
Martin Kay Machine Translation 
grammar of the latter would not cover the construc- 
tions of the former. Thus "Fair tommorrow" and 
"Winds from the northeast" are "sentences" in a 
weather bulletin. There are also closely related do- 
mains in which texts have a common syntax, and differ 
only in vocabulary. An extensive study of sublanguag- 
es, their restrictions and interrelations, will be impor- 
tant for determining the range of applications of fully 
automatic translation. The question of is a complex 
subject to which the report of another panel is devot- 
ed. 
The following table summarizes our view of the 
three most interesting points on the continuum. 
Information 
Acquisition 
Denotative 
Translation 
Connotative 
Translations 
Fully Automatic 
Machine Translation 
can be quite cheap 
(revision excluded) 
requires effort and 
experience to read 
technical material; 
possibly other 
material 
applications 
exist, greatly 
improvable 
faster and 
cheaper than 
human 1st pass 
human revision 
required 
technical 
material 
technology coming 
of age; applications 
exist (METEO); very 
large intermediate 
and long-term payoff 
N/A 
Human Assisted 
Machine Translation 
N/A 
very high-quality 
especially multi- 
lingual 
possibly high cost 
technical 
material 
few or no existing 
prototypes; FAMT 
spinoffs possible 
in near term with 
suitable funding 
Machine Assisted 
Human Translation 
increased human 
efficiency 
more expensive and 
slower then FAMT 
almost any 
material 
technology exists; 
might use FAMT 
to select candidates 
increased human 
efficiency 
high minimum 
costs 
almost any 
material 
commercial systems 
exist (e.g., ALPS); 
FAMT spinoffs 
could reduce costs 
in near term 
advantages 
disadvantages 
text types 
status and 
prospects 
advantages 
disadvantages 
text types 
status and 
prospects 
increased human advantages 
efficiency 
necessarily costly 
N/A legal and religious 
texts; literature? 
technology exists; 
greatly improvable 
disadvantages 
text types 
status and 
prospects 
76 American Journal of Computational Linguistics, Volume 8, Number 2, April-June 1982 
Martin Kay Machine Translation 
Legend 
Fully Automatic Machine Translation (FAMT) refers 
to translation wherein the programs run in "batch" 
mode (off-line) and produce translations without 
human intervention; afterwards, human revision 
(post-editing) may be performed with a text editing 
program or via other means, if desired. 
Machine-Assisted Human Translation (MAHT) refers 
to translation wherein the program is a fancy edit- 
ing and dictionary concordance tool which the hu- 
man translator uses to increase his efficiency by 
automating his access to word definitions and ter- 
minology correspondences. All initiative resides 
with the human, unlike FAMT and HAMT. 
Information Acquistion refers to a situation in which 
translation is being performed for "current 
awareness" or "screening" purposes where a quick- 
and-dirty approach may be sufficient, at least to 
determine if a more careful translation is justified. 
No human revision (post-editing) is assumed. 
Denotative Translation refers to an information disse- 
mination situation in which the everyday and tech- 
nical definitions of the words are meant, and where 
subtle nuances of a word choice are unjustified or 
even undesirable. This is typical of technical texts. 
Connotative Translation refers to an information dis- 
semination situation in which subtle nuances of 
word choice are very important, if not critical, in 
conveying the intended meaning of the text. This 
is typical of, for example, religious, and literary 
texts. 
At the opposite end of the spectrum from current- 
awareness services are such delicate enterprises as the 
translation of legal statutes and political speeches. A 
lawyer in Finland can base his arguments either on the 
Finnish or the Swedish version of the applicable law, 
according to which he considers will most favor his 
client's case. All statutes must be translated and the 
translators must be at great pains to ensure that there 
is no construction, however perverse, that can be put 
on one version but not on the other. We do not fore- 
see a time when any part of this job could be usefully 
consigned to a fully automatic or even a human-aided, 
system. On the other hand, it would be a prime candi- 
date for machine-aided methods. There is constant 
need to compare one part of the text with others, and 
with other legal texts, to ensure consistency, and this 
is where these methods come into their own. 
If our optimism about the future of mechanical 
methods in translation has increased during the twenty 
years during which it has been seriously pursued, it 
must be largely because of important advances that we 
perceive in theoretical and computational linguistics as 
well as computer science. Advances in computer sci- 
ence are the least contentious of these. The construc- 
tion of most large internal memories was not available 
and external memory could be accessed only in a serial 
manner. The consequent inefficiency in the programs 
that were written is less important than the undue 
amount of effort that was required to make them work 
at all. A machine-translation program was large, even 
by today's standards, and each one produced in the 
sixties was a programming tour de force. The achieve- 
ments are even more impressive for the fact that they 
were made without the aid of the compilers, editors 
and other paraphernalia that programmers now take 
for granted, and before the great value of certain pro- 
gramming practices and disciplines had been recogniz- 
ed. 
Many of the important advances made in computa- 
tional linguistics during the same period also tend to- 
wards the easier construction of more robust systems 
that can be more readily maintained. The most obvi- 
ous examples come from the domain of syntactic anal- 
ysis which is now universally thought of as a job to be 
done by a fairly general parsing program, coupled with 
a grammar. The parser embodies the necessary strate- 
gies and techniques while all knowledge of the particu- 
lar language resides in a static data structure, namely 
the grammar. Associated with the grammar is a formal 
language in which a linguist writes rules from which 
the data structure is obtained automatically. This 
formal language is specially designed to facilitate the 
statement of linguistic facts and is largely decoupled 
from the grammar itself and from the methods that 
will be used to process it. This greatly increases the 
power that the linguist can bring to the job and his 
ability to modify the system in the light of experience. 
In the same period we have come to understand, 
not just how a general-purpose parser can be con- 
structed, but how to make these parsers more effective 
by the application of some very general principles. In 
particular, we have come to appreciate the value of the 
notions of complete parsing and of nondeterminism. 
By complete parsing, we mean simply the requirement 
that nothing shall count as part of the final result ex- 
cept inasmuch as it is part of an analysis of an entire 
sentence. The practical value of this apparently obvi- 
ous restriction would be difficult to overestimate. A 
parser that incorporates it largely releases the grammar 
writer from concern for when it would be incorrect to 
allow an analysis that would be correct in another 
environment, a concern which consumed much time on 
the part of the designers of early translation systems. 
General methods for implementing nondeterminism 
go together with this and have an equally liberating 
effect. These methods are useful in situations where, 
at any given stage of the process, a number of con- 
flicting possibilities are open, any number of which 
could lead to useful results. In particular, a complete 
parser, in the present sense, must pursue lines of at- 
tack that seem reasonable on the basis of local eviden- 
American Journal of Computational Linguistics, Volume 8, Number 2, April-June 1982 77 
Martin Kay Machine Translation 
ce, but which may or may not lead to a complete anal- 
ysis. A general method for handling nondeterminacy 
releases the programmer from all concern for how the 
machine will contrive to follow up on all possibilities; 
how and when it will return to the choice point and 
restore the exact state as of that moment, how it will 
follow all possibilities once, but none more than once, 
and so forth. To the extent that early translation sys- 
tems faced these problems at all, they did so on a case 
by case basis, and at great cost in human labor. 
The panel was also impressed by the advances that 
have been made in general linguistics and our overall 
understanding of the workings of human language in 
recent years. Various classical problems--noun-noun 
compounds in English, ambiguities of prepositional 
attachment, conjunction, pronominal reference, and 
many others--have been made the object of intensive 
research with results that are direct relevance in the 
construction of translation systems. 
The panel thought it had every reason to assume 
that progress on the relevant fronts would go forward 
at least as quickly as in the past. Most of the mem- 
bers confidently expect to see some major new fully 
automatic systems in use during that period. In partic- 
ular, it is hoped that EUROTRA, a very large-scale 
collaborative European effort, will result in a working 
prototype. In addition, there is work in progress in a 
number of places on computer-based work stations for 
use by translators. It is not clear what these will in- 
corporate, but it is likely that they will explore some 
new parts of the large space of possibilities that exist 
in machine-aided translation. 
The panel made no prediction about just which 
areas of research were likely to fall before the inexora- 
ble advance of theoretical linguistics but felt that the 
future of more technological areas was easier to fore- 
see. There will, in all probability be more flexible 
methods of syntactic analysis, capable of relaxing re- 
quirements in the face of constructions that would not 
meet with a pundit's approval. Either they will fall 
back on more permissive rules or they will modify the 
sentence to "correct" the apparent "error". It is also 
expected that wider use will be made of parsing de- 
vices which, while allowing for nondeterminism in a 
general way, will be able to make fairly accurate 
judgements about the paths that are most likely to lead 
to a successful solution, and so concentrate on these 
first. 
The panel was in agreement on the achievements of 
the past and on the desirability of following a number 
of parallel paths in the future. The disagreements 
concerned the extent of the optimism in future suc- 
ce,;ses that those past achievements warrant. Some 
me, mbers took the view that advances in computer 
science and computational linguistics, important 
though they are, do not go to the heart of the prob- 
lem; they make easier what once was hard but they 
make nothing possible that once was impossible. Lin- 
guistics has made advances of which it can feel justly 
proud but which, while they may indeed go to the 
heart of the matter, barely scratch the surface of what 
needs to be done. 
All agree that it would be unjustifiable to devote an 
excessive proportion of the available resources to fully 
automatic systems while neglecting cheaper and more 
modest approaches with more certain short-term pay- 
off. Unless caution is exercised, both in promises 
made and policies followed, there is a high risk that 
taxpayers and administrators will call all too soon for a 
second ALPAC report whose effect on the entire field 
of computational linguistics will be altogether more 
devastating than the first. 
Some members believe that sponsors have grown 
more realistic in their expectations, that so long as 
they are involved in a continual dialog about the prog- 
ress of the work they can be made to understand the 
problems, and that they have the fortitude to with- 
stand unreasonable pressure from their superiors and 
their electors. They no longer, for example, expect 
full translations of arbitrary texts, but are content with 
texts from suitably restricted domains. 
It is claimed earlier in this report, and agreed upon 
by all the panel members, that fully automatic transla- 
tion is the line of attack whose benefits, if realized, 
would be greatest. Futhermore, its success would 
contribute greatly to the successes of all other ap- 
proaches. The subscribers to this view are impressed 
by the extent to which the designers of early systems 
were overcome by the sheer complexity of the design 
and programming task that they had undertaken so 
that the systems they built cannot be taken as a meas- 
ure of the technology that linguistics, even the linguis- 
tics of that day, could support. 
