An Introduction

to Machine Translation

ÉMILE DELAVENAY

THAMES AND HUDSON

LONDON
ENGLISH VERSION BY KATHARINE M. DELAVENAY AND THE AUTHOR
© THAMES AND HUDSON LTD 1960
PRINTED IN GREAT BRITAIN BY THE CAMELOT PRESS LTD
SOUTHAMPTON
EVERYTHING that is said is said about some part of the universe of
experience. . . . The universe of experience and the universe of
discourse must in the final analysis, be one.

The preconceived assumption that linguistics, physics, physiology

and neurology, force and energy, are all completely independent
of one another is precisely what has hindered and still hinders
progress, most of all progress in linguistics.
Joshua Whatmough
 Acknowledgments
THE author wishes to express his thanks to the scholars and their
publishers who have kindly given permission to quote from or to
summarize their works. Particular thanks are due to W. N. Locke
and A. D. Booth, editors of Machine Translation of Languages
(Technology Press of the M.I.T. & John Wiley & Sons, Inc.,
1955); to A. D. Booth, L. Brandwood and J. P. Cleave, authors
of Mechanical Resolution of Linguistic Problems (Butterworths
Scientific Publications, 1958); to the Academy of Sciences of the
U.S.S.R. and to D. J. Panov and I. S. Muhin; and to Erwin
Reifler, for permitting the use of duplicated reports and studies.

Transliteration of Cyrillic Characters

Russian names are transliterated throughout in accordance with
the norm established by the International Standards Organization,
which aims at transcription of characters, not pronunciation.
 Contents
page
I. TRANSLATION IN THE ATOMIC AGE 1
New aspects of the translation problem 2
Some aspects of electronic processing of information 5
Where translation differs 6

II. COMPUTERS AND LANGUAGE 12

Possibilities and limitations of computers 12
Tabulators 13
How a tabulator might translate 16
Electronic computers 18
Central organs 19
The binary code 22
Human language and machine signals 23
The central unit of a computer 24

III. VARIATIONS IN APPROACH 27

Brief history of research: 27
From Trojanskij to 1952 27
1952-1955 29
The expansion of research 30
The evolution of ideas: 32
Automatic dictionary and signalization of meaning 32
The separation of affixes 34
Birth and death of the pre-editor 35
German compounds 36
Better than word-for-word translation 37
The Georgetown-IBM experiment 38
1955—The turning point 39
vii
 MACHINE TRANSLATION
The concrete analysis of linguistic data: 41
The basic principles of early Soviet research 42

IV. FROM SOURCE LANGUAGE TO TARGET LANGUAGE 45

Inventory of means of expression 45
Languages, interlanguage and metalanguage 46
The linguistic mould of representation 50
Hieroglyphic conversion 51
Linguistic analysis by machine 54
Some examples of sub-routines 58
Grammatical analysis 59
Analysis as pre-synthesis 62
Towards a multilateral programme? 63
Priority of bilateral programmes 65

V. SYNTAX AND MORPHOLOGY 67

Importance and limits of grammatical problems 67
Morphology and the machine 71
Structural analysis 75
Classification and comparison of structures 76
Structural memories 80

VI. LEXICAL PROBLEMS OF AUTOMATIC TRANSLATION 81

Memories: technical alternatives 81
Consulting the electronic dictionary 86
Code compression 86
Linguistic problems 87
A French-Russian dictionary 87
Idioms and homographs 88
Genuine polysemy 90
Microglossaries 91
Statistics of word meanings 93
The Thesaurus method 95
Scientific and technical dictionaries 96
A national terminological centre and translation laboratory 97
From metalanguage to the untranslatable 98
Styles and vocabularies 100
The semantic atlas 102
viii
 MACHINE TRANSLATION
VII. FUTURE PROSPECTS 103
Limitations of the machine 103
Role of the machine 103
Operating cost 105
Literary prose 107
Collective methods 109
Poetry 109
Studies in poetical semantics 111
Literary analysis 112
Speeding up cultural exchange 114
What remains to be done? 114

Postscript to the English Edition 119

Bibliographical Notes 125

Glossary 129

Appendix 135

ILLUSTRATIONS
Figure 1. Sample punched card 14
Figure 2. Block-diagram of automatic translation programme
from English to Russian 55
Figure 3. Routine for stripping English word-endings and
dictionary check of remainders 72-73
Figure 4. Block-diagram of workshop including terminology
centre and automatic translating machine 84
Figure 5. Specimen of machine translation of a Foreword
for this book 134-5

ix
 CHAPTER I

Translation in the Atomic Age

FROM 1954 onwards the Press has from time to time announced
the invention or completion of a “translating machine”. These
news items have been premature, and more likely to hinder than
to help research, since they tended to encourage a passive attitude
towards a problem which still requires much patient investigation
and the collaboration, in new fields, of specialists hitherto little
accustomed to work together—linguists and electronics engineers.
Now in an advanced stage of planning, and certain within a few
years to become an accepted tool, the translating machine, to
all intents and purposes, is already with us. We can therefore rely
on the inventiveness of homo faber and study here, without
entering the realm of science fiction, the origins, workings and
potentialities of this invention.
It would no doubt be useless to swim against the stream and
to call it by another name. The law of least effort will assure the
success of “translating machine” by analogy with sewing machine,
knitting machine, washing machine, etc., even if we were to
propose a formula such as “electronic translator” or “automatic
translator”. Yet we are concerned not so much with a new machine
as with a new analysis of linguistic phenomena, particularly of
discourse, with a technology of language, made possible by the
application of electronics to the signs in which thought materializes
in the form of language. If we adopt here the accepted terminology
and speak of the translating machine, of the automatic or electronic
translator, it will be well to remind ourselves frequently that we
are dealing not with a robot brain replacing the mind of man,
but with a tool at the service of the human intellect and that the
main effort of research, which must be primarily linguistic, will
have to be focused on the process of translation, and not on the
invention of a machine, i.e. an assembly of parts and electric
circuits. Such machines already exist. It remains only to learn to
2 MACHINE TRANSLATION
use them for the purposes of translation. For we must guard on
the one hand against the cult of cybernetics and of the electronic
brain, and on the other against the complacency shown by those
who, hearing of Russian or American advances in the field,
imagine that all will be well if they let the engineers of these great
powers finish the job and then make use of their machines.
The idea of applying the new potentialities of electronic com-
puters to translation from one language to another has been in the
air since 1946. The opportunity of subjecting the material forms
of language to the analytical methods of machines capable of
arithmetical and logical operations was too tempting to go long
neglected. Moreover, the requirements of men in this atomic age
are such that automatic translation corresponds to a real need of
our time.

NEW ASPECTS OF THE TRANSLATION PROBLEM

The problem of translation, which has faced modern man ever
since the Renaissance has, like many other problems, taken on new
aspects in the light of the geographical shifts of power apparent
at the outset of the atomic era. Vast potentials of industrial
power are available to serve the political ends of great empires.
The conservation and expansion of this scientific potential de-
pends upon rapid and accurate information being made available
to scientists. But, today more than ever before, scientific in-
telligence is impossible without translation, since the fragmenta-
tion of knowledge and the intense specialization of scientists make
it extremely rare to find men with minds capable of synthesis
fully cognizant with various scientific subjects and accurately
and widely trained in linguistics.
Ours is not an age for learned disquisitions on “Unfaithful
Beauties”, the name given in the seventeenth century to Nicolas
Perrot’s translations from Greek and Latin classics in which he
claimed that he had embellished and improved the originals.
His enemies, who supported the cause of the inimitable superiority
of ancient writers, reminded him that translations, like women,
were rarely both faithful and beautiful. If we now pay any attention
to this old controversy, it will be rather to try and place the
problem of translation in its historical perspective from the
Renaissance to our own time, to put the emphasis on the needs
 TRANSLATION IN THE ATOMIC AGE 3
of science and to solve in accordance with the spirit of the age this
particular aspect of the perennial dispute between Ancient and
Modern.
Whether the aim of those who direct the action of today’s
scientists is war or peace, the welfare or the destruction of man-
kind, science itself needs translations of the results of contem-
poraneous work available in real time and sufficiently correct to be
understood. And this is true not only of the gigantic industrial
enterprises of the United States and the Soviet Union, but
equally so of the less well-endowed research teams of the lesser
powers. It is particularly true of the older nations of Europe,
which attach particular value to their languages as living ex-
pressions of their personalities but are no longer able to develop
their national scientific research on the same scale as today’s
industrial giants. It is significant that scientists and specialists in
the documentation of science, fully aware of the new and ever-
increasing need, have been the first to show interest in the problem
of automatic translation. It is only fair to add that perhaps they
were not, like the linguists, held in the leading strings of a historical
and literary training which continues to direct the study of
language towards the traces of the past rather than towards the
possibilities of the future.
Our scientific age is also a nationalist age. The empires of the
nineteenth century, advocates of the assimilation of the native,
or at best professing the theory of political and cultural inde-
pendence at a very distant date, dispensed the crumbs of western
culture through the intermediary of vehicular languages intro-
duced mainly for commercial purposes. These empires have now
almost all given way to sovereign nations whose first concern is to
assert their independence in every respect before recognizing
their interdependence with their former masters. They want to
expand teaching in the vernacular language, but at the same time
lay loud claim to their share in the scientific and cultural heritage
of man and proclaim their right to accede to universal culture.
These young nations demand translations: school textbooks,
books for teaching science and training teachers, readers for
children and for newly literate adults. Already, too, they are
demanding translations of the great works of world literature.
Let us make no mistake: the impassioned speeches of the
4 MACHINE TRANSLATION
former Lebanese President, Camille Chamoun, before the General
Assembly of the United Nations in 1946 in favour of the transla-
tion of the great works of world culture into the languages of less
privileged peoples called attention to one of the great cultural
problems of our time and opened up new vistas. But though
greatly broadened in scope when seen in the fresh context of the
nationalism of newly independent states, this particular aspect
of the translation problem has not greatly changed in essence
since the time of the Renaissance. What is required is not new
discussion of the old theme “Is translation possible?”, but an
effort to make available more authors, both ancient and modern,
in an increasing number of languages. The problem of quantity
comes first: quality will come later, with increased leisure. Did
we not for many years devour translations of Russian novels
which were certainly more unfaithful than beautiful? And are we
to think that these translations were useless?
So, without wishing to offend the classicists, the problems re-
quiring solution today are those of quantity and of speed. The
hard fact is that neither in advanced countries nor in the so-called
“under-developed” countries are there enough translators to
satisfy the priority needs of science and to communicate to the
masses all the scientific and cultural forms of knowledge. Good
translators are rare, and their work is inevitably slow. As for the
others, the most useful—or least harmful—are those whose
translations are faithful, if pedestrian.
Why do we not train more translators? It is true that this is
an urgent task for our schools and universities. It is also a very
long-term and often thankless task, and one which has so far not
generally been well tackled, except in a handful of institutions
where the approach has been realistic, without premature seeking
after literary effects. The advent of machine translation is no
reason for the schools to relax their efforts. On the contrary the
machines will require the services of a great number of linguists
schooled in the best methods of human translation.
Whether by coincidence or by almost prophetic foresight, the
first scientists who, some twelve years ago, envisaged the possi-
bility of using electronic computers to solve linguistic problems,
were moving towards a solution of these problems. The new high-
speed digital computers were still in their infancy when, in 1946,
 TRANSLATION IN THE ATOMIC AGE 5
Andrew Booth suggested to Warren Weaver, Vice-President of
the Rockefeller Foundation, that such machines might facilitate
the work of translators. Booth himself has said that his suggestion
was simply an intellectual exercise directed at finding yet another
use for the new machines. [7]

SOME ASPECTS OF ELECTRONIC PROCESSING OF INFORMATION

Since that date electronic computers have become so much a
reality that it is unnecessary to consider here at any length their
scientific, managerial, arithmetical and logical applications. But
the investigations first prompted by Booth and Weaver are part
of a whole series of theoretical and practical research work which
it will be useful to consider briefly before proceeding to examine
in detail the methods by which written translation can be made
automatic. It is also necessary to define their relationship, more or
less direct as the case may be, to various scientific theories or
fields of research.
In the fields of technology, of natural and social science, an
effort is being made to facilitate and accelerate access to informa-
tion by the use of modern methods of recording, classification
and search. The first attempts were made with various types of
card index, then with punched cards, whence it was relatively
easy to proceed to coding methods used by computers. Three
different types of system are at present in use: some based on
punched cards sorted either mechanically or electronically;
others using photo-electric sorting; while in magnetic recording
systems, the sorting is accomplished by methods similar to those
of the big computers. If the methods of analysis and indexing of
scientific documents, of coding and classifying the information
they contain, have not evolved rapidly enough to allow a really
revolutionary and economic use of the new techniques, it is
because these methods are closely related to one of the most
important aspects of research on translating machines: the creation
of electronic dictionaries. Progress in information retrieval by
means of key words will only come with the solution of the lexical
problems of machine translation, since both are closely concerned
on the one hand, with the cataloguing and classification of concepts
and the words expressing them and, on the other, with the
technical improvement of “memories”.
6 MACHINE TRANSLATION
For the purposes of official records, verbatim or summary, an
attempt is being made to “receive sound waves in order to extract
from them alphabetic information”, according to Dreyfus Graf,
who is working in Geneva on a “phonetograph” by means of
which a typewriter will be able to record speech directly.
In combination with a translating machine, the phonetograph
or some such machine enables us to envisage the possibility that
one day interpreters will be replaced by an interpreting robot.
But there is a long way to go before machines will be able to
identify with certainty the meaning of a sentence containing
homophones, to translate it, and to pronounce a corresponding
sentence in another language synchronically with the reception
of speech by the phonetograph. To be able to do this would
suppose that a solution had been found to the problem of the
matching of meaning from language to language while taking into
account at least some of the shades of individual meaning intended
by the speaker. Automatic interpretation involves other additional
problems besides those of homophony. It can therefore be
achieved only after automatic written translation has been accom-
plished.

WHERE TRANSLATION DIFFERS

Unlike machines designed for the transmission of spoken or
written discourse—the telegraph or the cryptograph—machines
designed to search for or to translate information must be able to
choose from among the material at their disposal. And, at first
sight, this choice seems to go much further in translation than in a
simple search for information. Now no machine can exercise
choice, except in accordance with precise criteria which have
been determined in advance and written into its programme.
The musical box or pianola of our childhood days also offered a
choice, as does also that embryonic automatic translator—the
tourist's Conversation Guide. Punched card sorters can regroup
certain data according to pre-established programmes. They can,
for instance, make a list of all customers in a certain business
sharing certain characteristics, or of all a firm's employees under
25 years old, etc. It would be easy to make a machine which would
“translate”, that is to say which would print out or even read
aloud ready-made translations of pre-selected texts chosen at
 TRANSLATION IN THE ATOMIC AGE 7
random from a certain list. Such a machine may well impress
sightseeers at a fair, but would serve no useful purpose. What
science is concerned to achieve is a machine which, while remain-
ing an object devoid of intelligence and of judgment, and per-
forming a series of strictly predetermined operations, is capable
of respecting certain of the original and individual characteristics
of discourse and of reproducing them faithfully in another
language.
The very conception of such a machine implies a thorough
exploration of the relationship between thought and language.
By exploring language in order to arrive at automatic translation
of discourse from one language to another, we raise once more the
question of the degree of freedom enjoyed by human thought, and
we are forced to consider the constraints within which it operates.
This is part of the eternal debate between a strictly determinist
conception of fate and human nature and the belief in liberty.
Here are problems of far more fundamental interest than that of
asking whether or not the machine will be able to translate
poetry or connotative language.
Thus the attempt to automatize translation leads us to a new
conception of linguistic studies. It is no longer a question of
delving into the past history of our languages, but of studying
the actual behaviour of language in the expression of thought,
of examining the inner dynamics of sentence creation, of the
materialization of nascent thought and the different possibilities
of expression offered by different languages. If we can think of
language as a mould, or framework, predetermined to a very
large extent, within which thought can express itself, we shall be
in a better position to assign to that research with which we are
here concerned its proper relation to various allied techniques.
An early observation concerned a superficial analogy between
translation and cryptography. It is indeed tempting to use the
word “translation” loosely whenever it is a question of trans-
ferring a message from one system of symbols to another. In this
sense it is possible to say that a stenographer “translates” her
signs before transcribing them into longhand. The telegraphist
“translates” a telegram from latin script into morse, etc. We shall
here use the word “translate” only to describe the transposition
of discourse from one language to another.
8 MACHINE TRANSLATION
Warren Weaver had not considered all the implications of his
thought, and allowed himself momentarily to be seduced by a mere
analogy, when he wrote, in 1947: “One naturally wonders if the
problem of translation could conceivably be treated as a problem
in cryptography. When I look at an article in Russian, I say:
‘This is really written in English, but it has been coded in some
strange symbols. I will now proceed to decode’ ” [17]
For the linguist this is an over-simplification, especially if he is
bi-lingual and trained in observing the divergent ways followed
by his discursive faculty as each of his two languages offers
different alternatives, even for the expression of precise facts and
scientific data. Weaver’s idea, which made for optimism at a time
when everything was still to be done, corresponds, however, only
to a very elementary state of research in the field. The history of
work on machine translation confirms I. K. Bel’skaja’s observation
that anyone basing research on the idea that the problems of
translating language by machine might be similar to those of
cryptography would inevitably be disappointed. [5]
In reality, in cryptography, coding and decoding, however
complex, operate always within the framework of a given linguistic
structure, to which the coded message must also conform. The
semantic and syntactic conventions are necessarily common to
both author and reader of the message, whatever transformations
the message may undergo between point of departure and point
of arrival. Translation from one language to another requires
something else altogether, and it is precisely this "something
else" that has been the object of research during the last ten years.
We shall none the less frequently have recourse to the experience
of coding and decoding experts, experience which is indispensable
for the realization of mechanical translation. We should not,
however, mistake the part for the whole.
The same observation holds good, but requires much finer
discrimination, concerning those aspects of cybernetics grouped
under the head of “information theory”. It is clear that if, as
G. Th. Guilbaud asserts, “one of the most active branches of
cybernetics will . . . be the application of statistical methods
to phenomena of which one of the dimensions is time” [14],
discourse is bound to interest cybernetitians precisely because
it is such a phenomenon and is susceptible of macroscopic
 TRANSLATION IN THE ATOMIC AGE 9
analyses offering a striking analogy with the study of thermo-
dynamics.
Studying on the one hand the structure and the measurable
properties of information itself, information theory will analyse
the measurable properties of sound waves of language, of al-
phabets, which are codes, and of those other codes, the systems of
symbolic figures which transmit messages through the circuits of
electronic computers. Still other codes which it will study are the
words in a dictionary—those conventional signs representing
things or ideas; inflexions which add information to the message
conveyed by the uninflected forms; and syntactic rules which, in
turn, give information by pin-pointing the individual meaning of
words and their role in relation to other words. All the methods
of measurement and analysis already applied to messages, to
keyboards and signals by specialists in information theory,
can be extended to the constituent elements of language and of
discourse.
Information theory also studies the circuits through which
information is transmitted, and the conditions governing their
stability and efficiency. Translating machines, like electronic
computers, are precisely assemblies of such inter-dependent,
mutually controlled circuits. Whether it is a question of discourse
itself, the object of research, or of the methods and techniques to
be utilized in finding a solution to the problems involved, it is
certain that information theory will constantly be called upon to
contribute to research on mechanical translation.
It is, however, necessary to guard against the idea that the
mathematical theory of the techniques of transmission can alone
provide art easy solution to the problem of translation. Certainly,
important parts of our work as a whole are closely bound up with
the methods of mathematical analysis used by the statisticians of
alphabets and signals; but we must never lose sight of the original-
ity and individual nature of discourse that must be translated
from language A into language B. The macroscopic application
of the theory of probability helps us to distinguish some statistical
properties of language as a system of information. But the useful-
ness of these statistical laws remains very limited in face of the
expression of individual thought in discourse. As Panov has said:
“The very nature of the problem of translation is such that
10 MACHINE TRANSLATION
individual features of the translated text cannot quite be ignored.”
[25, 26]

To look to information theory for the key to automatic transla-

tion would be to make a mistake similar to forgetting, when
looking at Watt’s governor on a steam engine, that this apparatus,
which in its way “feeds back” information, would be nothing
without the source of energy the application of which it regulates
and controls. A translation machine will always have to deal with
a text—the raw material, and, as it were, the source of energy;
with the methods of transforming this text (in the elaboration of
which information theory will play an important but limited role);
and then with a second text, the result of the work of translation,
an important aspect of which will be quality; this will be pro-
portional to the degree of respect which it will be possible to
give to the individual characteristics of the first text while changing
systems of information, that is to say, passing from one language
to another.
Nor can linguistics alone, even in its most modern forms,
provide the required solutions. Irrespective of whether it studies
language in its historical or in its structural aspects, this science
has rarely attempted to make a systematic inventory of the com-
plex relationships between a series of ideas in one language and the
most faithful possible expression of the same series of ideas in
another language. This kind of preoccupation has generally been
left to the translator, a practitioner whose art rarely receives due
appreciation. The mechanization of this art will never be possible
until the inter-relationship in the expression of the same facts or
ideas in different languages have been inventoried and scienti-
fically analysed with the help of the most modern statistical
methods and of the whole apparatus of information theory. Only
such an analysis of linguistic data and of the behaviour of the
constituent elements of language will enable us to progress and
finally to achieve mechanical translation. And such analyses can
hardly be made without the assistance of electronic machines
capable of treating rapidly and correctly large quantities of de-
tailed data and discovering their common characteristics. To
comprehend both the potentialities and the slowness of such an
undertaking, it is necessary to grasp the essential relationship
 TRANSLATION IN THE ATOMIC AGE 11
between the achievement of mechanical translation and the
empiric use of the very type of machine which we are trying to
realize. Like other industrial revolutions, this revolution in the
technique of translation carries within itself its own seed: to
begin is everything.
Technological research on the transmission of information,
the theoretical study of the mathematical laws of signals and of
messages, studies in the psychology of language and comparative
structural analyses—all these different lines of research form part
of the combined operations which will result in automatic trans-
lation. To say that it is above all a question of technology, rather
than of fundamental research, is not in any way to belittle the
problem or the work required. It merely defines the limits of a
particular field within a much wider area. One of the peculiar-
ities of this field is that it lies somewhere between the macroscopic
analysis of language studied from the point of view of quanta
theory and the microscopic analysis of those individual states of
consciousness out of which discourse is born.
 CHAPTER II

Computers and Language

POSSIBILITIES AND LIMITATIONS OF COMPUTERS
To the arithmetical functions of mechanical calculating machines,
the prototype of which is Pascal’s adding machine, the new
electronic computers have recently added logical functions; that
is to say that they perform operations closely akin to certain
actions of the human mind which, at first sight, would appear
less readily mechanizable than the four arithmetical operations.
Moreover, technical improvements to adding, accounting and
statistical machines have resulted in a considerable increase in
the possibilities of expression of these labour-saving tools, con-
sidered from the point of view of the detailed and explicit informa-
tion communicable by their output organs. As these machines
can now imitate an increasing number of mental operations, it is
important to clarify their relationship to language, envisaged here
as the faculty of expression of thought through the spoken or
written word, and to discourse, the materialization of this faculty
in the form of auditive or visual signs.
A factor common to all the above-mentioned machines is their
ability to transform and reassemble data, whether consisting of
numbers only, as in the case of adding machines, or of both
numbers and alphabetical signs, as in the case of accounting or
statistical machines and computers. Data (for example two
numbers the sum of which is required) can be fed into the input
organ of all these machines; they all perform operations, the
ordered sequence of which, if several are required in succession,
is called a programme; the result of these operations appears in
the machine's output, in one form or another, whether it be on
Pascal’s adding machine, on the paper tape of the calculating
machine, or on the typed statement of the tabulator. We shall use
the term peripheral functions or organs for those which are
concerned with input and output, and central functions or organs
 COMPUTERS AND LANGUAGE 13
for those performing internal operations, the results of which are
usually visible only in the final output.
Statistical tabulators have familiarized us with the high-speed
output printer, in which the final results are typed out in im-
mediately usable form. Direct reading of data by the machine is
now rapidly becoming possible, thanks to the photoelectric cell.
As for the central operations, such as the four operations of
arithmetic, logical processing, or the matching of data received
in the input with data stored in the memory, these have become
fully automatic and extremely rapid in the electronic computer,
where the substitution of inertia-free circuits for the components
of the old mechanical calculating machines has made it possible
to execute programmes of ever-increasing complexity and variety.
But however flexible their programmes, however dizzying their
speed of operation, the potentialities of all these machines are
fundamentally limited by the very fact that they are machines,
responding to signals, but incapable of doing more than repro-
ducing, even while reassembling them in another form, the data
fed into their input. Whereas the machine obeys signals, language
is an exclusive attribute of its human operator, who interprets
the results provided by the machine.

TABULATORS
Punch-card machines can, by appropriate entries typed out on a
printed form, give expression to the relationship between certain
figures and certain words designating objects or people. From
input material they can prepare documents, invoices, statements,
etc., taking into account relatively numerous and complex factors.
Their work already goes a long way towards imitating certain
associative functions which might have been supposed to have
been a preserve of human intelligence.
Punched cards are used in these machines for three essential
purposes: input of numerical information; input of alphabetical
information; input of programme instructions. Punched cards
also constitute a “memory”, since they are used to store in
permanent form information usable at any time.
But whatever the purpose for which a punched card is used, the
movements which it produces within the machine are always of
the same type. The card is moved sideways, and its columns are
14 MACHINE TRANSLATION
sensed by brushes which activate electro-magnetic components
by the action of electric circuits, through a switchboard which can
be modified at will when it is desired to change the programme. A
card can be divided into zones in accordance with a pre-established
plan, the holes in each zone having a predetermined meaning
distinct from other areas, so that the pattern made by the holes in
the cards, controls the details of the execution of the programme.
A punched card (see Figure 1) is divided into a given number

Fig. I. Sample punched card.

of—generally 80—vertical columns with ten horizontal positions

numbered from 0 to 9. Other perforations can be made above these
ten horizontal lines, thus allowing an even greater number of
combinations. The perforations from 0 to 9 set in motion the cogs
of an adding machine and the keys of the printer; the perforations
at the top of the cards, in combination with positions 0 to 9,
represent the letters of the alphabet and other conventional signs,
and set in action either the electro-magnetic mechanisms of an
electric typewriter, or any other required movement of the
tabulator's mechanism.
The punched cards lend themselves to various operations such
as sorting, analysing and reassembling information, and can thus
be extremely useful, for instance in linguistic analysis and data
retrieval. The principles involved in their use are perhaps best
illustrated in the relatively simple operation of commercial in-
voicing. A pack of cards is fed into the machine, the first card
 COMPUTERS AND LANGUAGE 15
bearing the current date and all other indications of a general
nature. The second card bears the name and address of a particular
client, together with the number of his account, the basic rate of
discount to which he is entitled, and any other special indications
applicable to this client. On other cards holes have been punched
to correspond to the names of the articles ordered by this client,
the unit price of each article, etc. The number of units supplied,
or indications concerning units ordered but not supplied, may
appear on one or several following cards. And so on for all the
orders of all clients for one day. Set in motion by the cards, the
machine performs all the necessary numerical operations and
draws up an invoice bearing the name and address of each client,
printing out in full the usual information concerning the number
of articles supplied, unit price, discount, sub-totals and overall
total. The invoice will indicate where necessary that such and
such an article is temporarily out of stock, no longer available, etc.
The strictly automatic nature of these operations, the fact that
they proceed without risk of error once the cards are correctly
punched and assembled, as well as their rapidity, distinguish them
from the same work performed by man. The machine has trans-
ferred alphabetic data from the cards to the printed form, in an
order which is predetermined by the programme (switchboard,
punching, ordering of the cards); it has performed numerical
calculations, and it has transferred the results into the appropriate
columns and lines of a pre-printed form. The whole effort of
reflection, of intelligence, had been made during the conception
of the tabulator and the establishment of its programme. The
use of the machine in this particular field has thus had the effect
of drawing attention to the purely mechanical character of various
operations formerly performed by a human being and accepted as
mental operations. The invention of the tabulator has pushed
back the frontier between the mental and the mechanical.
Is it permissible then to speak of tabulator language? It is clear
that the output component of this machine remits language in a
form directly accessible to the human reader, but it does so as
strictly mechanically as of a clockface which “tells” us the time.
From one end to the other of the machine's operations, signals
are transmitted along wires, movements are set off by trigger
actions. At the outset we find information “objectified” or
l6 MACHINE TRANSLATION
“materialized” in the form of punched holes; at the end the
material presentation, in the form of typewritten letters and
figures, of a new combination of the data originally fed into the
machine. This re-arrangement has been performed by the central
organs. To speak of machine language would be almost as mis-
leading as to speak of finger language when the hand strikes a
typewriter key, on the pretext that the corresponding character,
in hitting the inked ribbon, inscribes on the paper a letter which
has meaning. Yet the mechanical movements have, on the one
hand, performed calculations, and, on the other hand, imitated
association of ideas, such as that which consists in associating the
name of Mr Smith with the order for twelve coffee grinders at
x cost with y discount.
Moreover, if the machine can thus imitate language and
reasoning without risk of error, it is because all the numerical and
alphabetical elements of its work have, like its programme, been
meticulously prepared. The reasoning was done by man prior to
the operation of the robot. In the punched card system and in the
electric connections of the tabulator, each signal is unequivocal;
choice is no longer involved. Once the holes have been punched
in the cards and the cards selected and placed in the required
order, we are in a world of strict conventions from which am-
biguity or possibility of interpretation are excluded. The com-
binations of alphabetic signs which the machine transmits or
reproduces have taken a unique, positive semantic value, as
exact as that of the figures in the arithmetical element of the
tabulator. Everything in this system is predetermined and
inhuman.

HOW A TABULATOR MIGHT TRANSLATE

If we call syntax the associative faculty of the machine and
vocabulary the words punched on to the cards in the form of
alphabetic signs, we shall see that this mechanism is, in a limited
way, able to make some sort of sentences. Evidently this same
machine, with the same figures but different combinations of
letters of the alphabet, can make sentences in French, in English
or in German, in accordance with “instructions” provided by the
alphabetic punched holes and with a programme which may,
if required, alter the order of the words.
 COMPUTERS AND LANGUAGE 17
With a suitable pre-selection of cards, it might even “translate”
an invoice from French into English. For this purpose a system
of matched meanings would have to be established between (1)
the French names of the objects designated; (2) the corresponding
French alphabetic punched holes; (3) the corresponding English
alphabetic punched holes; and (4) the English names of the same
objects. All that would be required for instance would be a pro-
gramme enabling the French alphabetic punched holes, instead
of setting in motion the typing of the French words, to call for
the corresponding English punched cards, which in turn would set
in motion the typing of the English words. This would be a
clumsy and scarcely useful process: nevertheless, it illustrates
how a tabulator might translate—within strictly limited and
entirely predetermined terms of reference. We may ask whether
it is really possible to give the name of translation to this search
for and typing out of signs which have a meaning in one language,
mechanically reproduced by association with signs having the
same meaning in another language, all within a single, pre-
determined syntactic mould. But does not in fact the translator
do exactly this when he renders, in a list of articles on order, one
gross fountain pens by 12 douzaines de stylos?
Thus we are now ready to accept the idea that translations,
even if only of a very elementary nature, can be mechanized with
the help of a relatively simple machine to which naturally we deny
any creative aptitude or faculty. The name of translation is as
appropriate to these operations as is the name of arithmetic applied
to the sums totalled on Pascal’s adding machine. Thus the
borderline between the mechanical and the mental recedes yet a
little further when we observe how a machine can shift and com-
bine signs in a manner which leads to the same results as a trans-
lation. Once we have accepted this starting point, a process of
irrefutable logic will enable us to push back this limit even
further as we pursue a parallel course, on the one hand, towards
the creation of more complex mechanisms and, on the other,
towards the analysis of discourse considered as a series of material
signals meaningful to the human mind. Let us however bear in
mind that we are speaking here only of thought which has already
been materialized in the form of signs.
l8 MACHINE TRANSLATION

ELECTRONIC COMPUTERS
The electronic computers of today possess mechanisms sufficiently
complex to permit a further and closer analysis of linguistic
phenomena, not statically or in the abstract, but in relation to
dynamic sequences of mechanical operations and switchings of
electric circuits, the final effect of which is an output which
imitates discourse. The really original character of the linguistic
studies born of the research which will culminate in automatic
translation is just this: discourse can now be studied in relation to
the functioning of mere unconscious mechanisms, by means of
the laboratory instrument provided by electric circuits.
Computers have been in existence for barely twenty years. Only
since 1950 has their use become extensive enough to play a
significant part in our economic and social life. While utilizing
most of the old methods of mechanical computing and of tabu-
lators, they have introduced three essentially new characteristics:
Stupendous speed of operation, resulting from the total or
partial replacement of electro-mechanical cog wheels by electronic
circuits so that signals travel at speeds bordering on that of light.
Increased flexibility and complexity of programmes, also made
possible by electronic switching of circuits, instead of the former
mechanical or electro-mechanical methods.
The extension of the central functions, logic being added to
arithmetic, a development also speeded up by the use of electronic
tubes, rectifier circuits and magnetic cores.
These three basic characteristics have made it possible for
computers to imitate certain operations of the mind, certain
mechanical aspects of which had not previously been emphasized.
Simultaneously with the evolution of the central organs, the very
rapid improvement of input and output media has also increased
certain resemblances to human mental functions.
The first revolutionary change in computing machines was the
introduction of the memory, that is the faculty of holding within
the machine the results of a calculation before proceeding to the
next one, without output of the first result and its re-input by
human intervention, before the next operation is started. With the
traditional adding machine it was necessary for anyone wanting
to effect a series of operations to transcribe the intermediary
 COMPUTERS AND LANGUAGE 19
results and then reintroduce them manually. Charles Babbage,
who worked out the design of his Analytical Engine as far back as
1833, was aiming at the automatic performance of successive
arithmetic operations. His machine included a memory or store,
consisting of a group of accumulators, into which the results of
operations made by the mill or arithmetic organ could be trans-
ferred. These partial results could also be put back into operation
in the mill as and when required. The programme, which included
calculations and transfers from memory to mill and vice versa, was
controlled by two bands of punched cardboard similar to those
used on a Jacquard loom. Babbage was never able to complete his
machine owing to the inadequate production and tooling facilities
of his time.
In 1944 the Mark I or Automatic sequence calculator of Pro-
fessor Aiken followed the main outlines of Babbage's analytic
machine. The inertia of its electro-magnetic relays limited both
its speed of calculation and its memory capacity. The use of
electronic tubes, particularly of the double triode or flip-flop, in
the E.N.I.A.C. (Electronic Numerical Integrator and Calculator)
of the University of Pennsylvania made it possible in 1946 to
perform in 2.8 thousandths of a second a multiplication of 10
figures by 10 figures, as opposed to 6 seconds with Mark I. With
E.D.S.A.C., constructed at Cambridge University, and Aiken's
Mark III, the superiority of electronics was fully established; at
the beginning of the fifties the great industrial and commercial
enterprises began to be interested in computers. The big I.B.M.
data-processers, Remington Rand’s Univacs, Leo of Maison
Lyons in London, Ferranti’s Pegasus and Mercury, Bull’s Gamma
60 in France, the B.E.S.M. in Moscow, are all endowed with high
operational speeds and with arithmetical and logical components
capable of extending their operations far beyond mere sequences
of computation. All these machines are about to be superseded
by much faster computers. Their essential organs are more or
less alike.

CENTRAL ORGANS
The store or memory, as conceived by Babbage, is the faculty of
holding data in reserve either permanently, e.g. a table of logar-
ithms, or momentarily, e.g. the partial results of a sequence of
20 MACHINE TRANSLATION
operations which can be brought back into play at the desired
moment in the execution of the programme. Both figures and
letters can be stored in a memory, where they are represented
either by holes punched on cards or on teleprinter tape, or in any
other material form corresponding to the input technique em-
ployed. All that it is necessary to know here is that modern com-
puters use different kinds of memories for different purposes
which may vary in the following respects: capacity; time of
access to stored information; whether data are accessible at random
or according to some predetermined sequence; whether the record
is permanent or not. The main types of memory now in use are
magnetic tapes and discs, drums and ferrites or magnetic cores.
A memory can contain a varying number of signs or characters.
On a magnetic tape made of plastic covered with magnetic oxide
only ⅜ of an inch of tape is necessary to record all the information
contained in the 80 columns of a punched card, so that 10,000
characters can be read in one second. The capacity of an I.B.M.
magnetic tape is 5,760,000 characters. Such a tape constitutes a
very high capacity memory, but with sequential, and therefore
relatively slow access. The same is true on the whole of magnetic
discs, which like the tapes can be arranged in batteries so that their
capacity is virtually unlimited. Access-time on discs can be
lowered by an increase in the number of reading heads or by
high-speed motion of a single reading head.
Magnetic drums are metal cylinders covered with a magnetic
product. They revolve continuously at high speed and reading and
recording heads are arranged around them so as to make it
possible at any given moment to record a message at a given
"address" or to extract the contents of any section of the drum.
These memories have a limited but considerable capacity (294.912
binary figures in the I.B.M. 704 now installed in Paris). Access to
data is practically random, and is very rapid, the average access
time being of the order of a few millionths of a second in the 704
and varying from 22 microseconds to under one millisecond in the
Gamma 60. As with the tapes, the recordings may be preserved
or erased at will.
Magnetic cores or ferrites are small rings of magnetic matter
mounted in large numbers on insulating frames. An electric
current passing along wire through a core creates, according to its
 COMPUTERS AND LANGUAGE 21
direction, either a positive or a negative magnetic field which lasts
until a new electric impulse comes to wipe out the figure so
recorded. The number of figures which can be registered is
limited to the number of cores, each with its pair of wires. On
the other hand access time is extremely short. In the Gamma 60
it is of the order of 11 microseconds; in the I.B.M. 704 a word
of 36 letters can be transferred from a ferrite memory into an
operating unit or vice versa in 12 microseconds.
In all these memories data are recorded by positive or negative
magnetization of surfaces or volumes. There are other possible
processes, including that of photography on film or transparent
disc sensed by photoelectric cells. The photoelectric disc is likely
to play an important role in the linguistic and philological field as
well as in information searching and abstracting of documents,
particularly in the form first developed by Gilbert King at the
International Telemeter Corporation. In this form of memory
millions of characters and figures can be stored on a very small
surface and read at extremely high speeds. Cryotrons also provide
immense storage facilities on a very limited volume of matter—so
that the size and speed of access of memories are rapidly ceasing to
be a major preoccupation of machine constructors and users.
Other elements of a modern computer are also included in the
category of memories—an excessively anthropomorphic term which
the English language is happily able to replace by the more
accurate name of stores. Those which we have already mentioned
are in effect nothing but stores in which it takes more or less time
to find what one wants, as in an index, a library or a warehouse.
Others, which we might call intermediate memories, are designed
not to store information permanently, but either to hold back its
transmission in order to introduce it again at the required moment
(delay lines), or to keep it during a certain stage of a sequence of
operations (registers).
A memory may be used to store data, or to store programme
instructions. Both take exactly the same form. The punched card
of the type used in statistical machines is in fact the first memory
for input data. Here the holes represent figures, alphabetical and
other conventional signs. The first “programme memory” was
the punched card of the Jacquard loom and the programme-
controlling memory in Babbage’s Folly was of similar type. We
22 MACHINE TRANSLATION
have seen that a programme can be controlled by the holes in
punched cards, so that the same medium—the punched hole—is
used both to record the data on which the machine operates, and
to give the machine its instructions and dictate the order of its
operations. This means that the perforation corresponding to the
figure “1” or to the letter “a” may either actually represent this
figure or this letter in a recorded piece of information, or it may be
the material symbol of an instruction such as “multiply x by y” or
“transfer the contents of the arithmetical operator into the mag-
netic drum”, etc. The same signal will thus mean one thing or
another—be a fact to be operated on or an instruction to operate
—according to its position in a sequence of signals.

THE BINARY CODE

The unification of data and programme does not end here. The
universal computer, which performs arithmetical and logical
operations, must be controlled by the most simple and universal
impulses if excessive specialization of its organs is to be avoided.
So that the input data and the programme instructions, both
numerical and alphabetical, are usually communicated to the
machine in a single form, that of the binary code.
We have seen figures and letters represented on cards by
different combinations of the same punched holes. These same
figures (from 0 to 9) and these same letters can be represented by
means of a code comprising only two signs, 0 or 1 or + and −.
This is the binary code, now used in most computers. A minimum
of 6.41 binary signals or bits are required for one character of the
typewriter keyboard (14); 4 bits are needed to represent the ten
figures from 0 to 9; to the letters of the alphabet must be added
conventional signs, punctuation, brackets, etc., so that in practice
characters are fed into the memories in the form of 4 bits for
figures and 8 bits for alphabetical and other conventional signs.
The method used to change from decimal figures and alphabetical
letters into binary digits varies according to the codes used.
Figure 1 gives a concrete example of the intermediary stage. The
conventional signals in this figure are transposed by the machine
into binary digits, that is to say into series of 0 and 1, of positive
and negative magnetizations.
There has been much talk of machine language. In reality the
 COMPUTERS AND LANGUAGE 23
“language” of the central units of computers is a series of electric
impulses of plus or minus value, which can effect positive or
negative magnetizations of memory surfaces, corresponding to the
signs 1 or 0 and susceptible of being transcribed in perforations on
punched cards, or in letters and decimal figures by a high speed
printer. As in the tabulator, everything which for the human
card puncher was figures, punctuation signs, letters, etc., has
become, in the arithmetical and logical machine, signals capable
of activating mechanisms or electric impulses, of varying potentials
or of magnetizing surfaces. The code has become independent of
variations of meaning in the messages which it transmits. Its
object is to switch circuits, to set in motion electronic or electro-
magnetic movements, and each series of combinations of plus or
minus, of 1 and 0 does its work in the computer with complete
indifference to the fact that 0001 in binary code means 1 in the
decimal system and that 00001011 may mean "a" or that this
letter forms part of such and such a word.

HUMAN LANGUAGE AND MACHINE SIGNALS

While the very great speed of the computer makes it possible to
utilize this universal instrument, the binary code, the principle
involved is no different from the methods described a propos of
the action of punched cards in the tabulator. Once again we are
dealing not with language but with a sequence of electronic
operations transformed at the output into signs which a human
reader can turn into language by ascribing to them a meaning,
that is to say by establishing a meaningful relationship between
these signs and an exterior reality. Without this reference to
objects, the signals are void and belong entirely to the domain of
information theory and not to that of the psychology of language.
It is important to remind ourselves of this fundamental difference
between human language and what has been called, by extension
and by analogy, machine language, since, in the first case we
are dealing with a conscious mind establishing a relationship
between the signal and the object it represents, and utilizing
mechanical methods of acoustic or visual transmission and
reception to represent it, whereas in the second case we are
dealing only with transmissions which in themselves are devoid
of meaning.
24 MACHINE TRANSLATION

THE CENTRAL UNIT OF A COMPUTER

The central unit of a computer does however contain processing
organs which combine input data with others stored in its
memory.
The actual name of these elements of the central unit varies
with the different makes of machine and with the differing con-
ceptions of their manufacturers. In order to simplify and to avoid
too many technical details we shall here restrict ourselves to those
organs which answer to the needs of translating machines.
The arithmetical unit performs the four ordinary arithmetical
operations. An addition or subtraction takes 180 microseconds
and a multiplication between 400 and 800 microseconds. For the
main purposes of translation the machine needs only to add and
to subtract.
The collator makes it possible to compare two alphabetical
words of varying length. This operation is performed by com-
paring (subtracting) the binary figures representing them, the
result being zero if they are identical. Anyone accustomed to
consulting a bilingual dictionary in translation work will at once
appreciate the value of this operation, which consists of matching
a word of the input text with a word recorded in the memory. This
matching operation, when successful, can command the output
of the equivalent word in another language.
The logical processer performs logical operations including
the determination of appropriate programme instructions. It is
essentially an organ capable of performing an instruction of the
following type: if the result of a certain operation is positive or
zero, execute instruction (a); if, on the contrary, the result of this
operation is negative, execute instruction (b). This instruction is
known as a “jump” or “conditional transfer order”. If, for
example, the word “works” is a noun, the machine must translate
it into French by “travaux”, but if it is a verb, by “il (elle) tra-
vaille”. The translation programme will in such a case be con-
trolled by a “jump” instruction—the memory containing the
notations “noun” and “verb” against the word “work”. The
collator will match the input word against a dictionary word, then
it will identify its grammatical role in the sentence and check that
against the two grammatical notations. Only then will it find the
 COMPUTERS AND LANGUAGE 25
French equivalent of the word in the sentence under considera-
tion. All the various instructions for the performance of this
sequence of operations are given by the logical processer until the
sub-routine for this word ends in the correct matching of the
English word. In the same way any choice between possible sub-
routines will be determined by the logical operations of this
organ, which will take charge of selecting the appropriate sequence
of programmes whenever particular circumstances require it, as
often happens in computation for business management, and as
will happen in translating.
A modern computer comprises various registers on which data
are recorded during analysis or operation; some for storing changes
in programme, some for storing addresses (an address being an
indication of the place where a certain piece of information or
instruction is to be found in a memory), some which are accumu-
lators for storing partial results of operations, etc.
Thanks to the binary code, the computer can receive numerical
and alphabetical data and instructions in a single form admirably
adapted to the switching of circuits. Thanks to its central units,
it can combine these data in numerical or logical operations in order
to produce the results required in the solution of complex prob-
lems demanding matching of multiple data and numerous
sequential calculations. It can provide the results of its work at its
output in readable form. It is capable of infinitely more complex
and sustained operations than those required for a simple transla-
tion from one language to another. It can perform operations of
matching, identification, analytical logic and arithmetic on any
kind of data, provided that these data can be conventionally
expressed in figures and letters or in any other system of agreed
signs.
For these machines (each word being treated as a group of
alphabetical symbols) it is child's play to search a memory for the
exact French equivalents of thousands of English words and to
write them down one after another. But when English words make
sentences, the corresponding French words rarely make French
sentences, unless the machine has been given the necessary
instructions for substituting French syntax for English syntax,
for changing the order of certain words and for conjugating the
verbs differently, for looking up signals other than those of the
26 MACHINE TRANSLATION
alphabet, etc. The most modern machines can do all this, pro-
vided they are given a programme which takes into account all
alphabetical and other signals included in a written sentence. The
work of the past ten years has been directed towards the achieve-
ment of just such a programme.
 CHAPTER III

Variations in Approach
THE idea of automatic translation has generally been greeted by
linguists and translators with a certain degree of scepticism, the
natural result of their inbred knowledge of the difficulties of
translation. Very few have studied the structure and content of
language with the strict discipline of the natural sciences, examin-
ing them with instruments or methods equivalent to the micro-
scope, the slow motion projector or mathematical analysis. It is
scarcely surprising therefore to find that the ideas resulting from
the early co-operation of linguists and electronics engineers appear
on some points very far removed from what are now accepted as
the main avenues of research in this field. We shall, however, be
able better to understand the present state of such research if we
first examine briefly the past history of these new studies, the
evolution of the conceptions which underlie them, as well as of
certain points of detail. Moreover, in many respects this evolution
has been, and still is, dependent upon the perfecting of computers
and on improvements in techniques of memory and of input.
Without the hesitations and false starts of the pioneers, today's
bold advances would have been impossible.

BRIEF HISTORY OF RESEARCH

From Trojanskij to 1952. It appears that the invention of the
Russian Smirnov-Trojanskij, patented in Moscow in 1933, made
it possible to translate Russian into several languages simultane-
ously over a telegraph. But Soviet linguists failed to respond when
he sought their support in 1939, and the Institute of Automation
and Telemechanics of the Academy of Sciences was equally
unforthcoming in 1944.
In 1946 a dual approach to the problem of mechanical transla-
tion was made by the Englishman A. D. Booth, and Warren
Weaver of the Rockefeller Foundation. In response to Weaver’s
28 MACHINE TRANSLATION
suggestion that wartime decoding methods might be applied to
language, Booth pointed out that an electronic computer is capable
of storing a sufficient quantity of data to make it possible to effect
a “word-for-word” translation of the type which might be made
by relying exclusively on a dictionary.
Up to this point there was no question of syntax, or of word
order, nor even of translating all the words of a text. The idea was
simply that, to help the scientist to understand a document in a
foreign language, one might usefully put before him a translated
list of keywords, relying on his intimate knowledge of his subject
to enable him to find a guiding thread of meaning through the
disconnected words.
At Princeton, Booth and Britten began to work out the in-
structions necessary to enable a computer to consult a dictionary
recorded in its memory and to provide a word-for-word translation
of sentences fed into it on punched tape. In 1948, Richens, another
Englishman, introduced the idea of automatic grammatical
analysis of word-endings. This can not only improve the transla-
tion by giving the reader information on the grammatical role of
words, but also speed up the looking-up of words in the electronic
dictionary, since in theory it reduces the total number of entries.
Word-for-word translation can now take the following form:
amat—to love 3 p.sg. works—(1) travail sb.pl. (2) travailler 3 p.sg.
The coded grammatical indications assist the reader of the trans-
lation in understanding the meaning of groups of words.
In 1949 Weaver pointed out that by penetrating beyond the
apparent divergencies from language to language, one discovers
statistical invariants, as found in cryptography and recognized by
information theory, semantic invariants, as observed by the
sinologist Erwin Reifler between languages having no historical
link, and logical invariants, as described by Reichenbach. These
invariants, Weaver thought, may correspond to certain basic
characteristics of the human brain and to the common psycho-
social origins of language. Referring to the work of Booth and
Richens, Weaver maintained that a purely word-for-word transla-
tion is capable of rendering great services to scientific and technical
research. He also went much further and raised the question of the
possible solution of semantic ambiguity by exploration of immedi-
ate context.
 VARIATIONS IN APPROACH 29
The logical elements of language, Weaver claimed, can be treated
by the logical circuits of the computer; Shannon's information
theory can throw statistical light on translation problems, par-
ticularly if studies of statistical semantics are undertaken in the
light of this theory. Finally he raised the vital question of research
into “the common base of human communication—the real but
as yet undiscovered universal language”.
As early as January 1950, Reifler circulated privately his first
study on machine translation, the first serious attempt by a
linguist to analyse the preparation of written texts for translation
by computer. He postulated the necessity both for pre-editing
texts before translation and for post-editing them when translated.
Research schemes began to multiply, with apparent lack of co-
ordination. Oswald and Fletcher studied the mechanical resolution
of German syntax patterns. At the Massachusetts Institute of
Technology Bar Hillel, the Israeli logician, was able to devote his
whole time to research on language with a view to mechanical
translation. Early in 1952 the Rockefeller Foundation made it
financially possible for M.I.T. to call the first conference of
linguists and electronics engineers devoted to mechanical transla-
tion. The eighteen participants agreed on the next two stages of
research. The first step was to undertake, in scientific texts,
studies of word frequency and language-to-language equivalence,
while analysing the methods for using electronic memories and
examining other technical aspects of the automatic dictionary.
Later would come syntactic analysis for the elaboration of machine
translation programmes. The study of the circuits necessary for
the resolution of grammatical and syntactic problems could be
left until a later stage. Work on a multi-lingual machine should
await the first results of one-way translation from language A to
language B. But the possibility of using an intermediary language
—machinese—capable of serving as a turntable between all
languages, was not excluded.
1952-1955. American research developed rapidly as a result of
this first exchange of views. Studies were made of the storage
capacity of memories, the usefulness of restricted vocabularies,
the mechanical identification of meaning and of word-endings,
etc. In 1954, Dostert and Garvin of Georgetown University and
Sheridan of I.B.M., successfully carried out on a 701 I.B.M.
30 MACHINE TRANSLATION
computer the first experiment in automatic translation from
Russian into English with a vocabulary of 250 words and six
syntactic rules. Also in 1954, M.I.T. published under the direc-
tion of William Locke and Victor Yngve, the first number of a
periodical entitled M. T. (Mechanical Translation). In the following
year the first published book on the subject appeared—Machine
Translation of Languages, edited by William Locke and A. D.
Booth.
The year 1955 also brought the first news of Russian activity in
the field. The big B.E.S.M. computer of the Institute of Precise
Mechanics and Computer Technology of the Academy of Sciences
of the U.S.S.R. was used for experiments shortly to lead the
Academy to the conclusion that automatic translation was possible.
The expansion of research. In October 1956, M.I.T. called the
first international conference on machine translation, at which
foregathered some thirty specialists from Great Britain, Canada
and the United States. Dr D. J. Panov, of the Institute of Precise
Mechanics of the Academy of Sciences of the U.S.S.R., sent an
important written contribution on Russian research. The three
main centres of activity were Great Britain, the United States and
the Soviet Union; smaller groups were at work in Italy and
Scandinavia. The problem was no longer to prove by preliminary
research that mechanical translation was a possibility, but to
organize this research in such a way that effort was concentrated
for maximum efficiency and that synthesis of scattered studies
should become possible.
By 1959 a dozen or more groups were working actively in the
United States. At Harvard, Oettinger continued and expanded his
work on the Russian-English automatic dictionary. At M.I.T.,
Locke, Yngve, Chomsky and others were studying syntactic
structures, German syntax, basic methodology and devising
methods enabling the linguist to programme for a computer. At
Georgetown, Garvin and Zarechnak concentrated their efforts on
Russian syntax, while A. F. R. Brown worked on translation from
French. At the University of Michigan, Koutsoudas and Korfhage
were at work on Russian, with particular emphasis on the poly-
semantic problem. In Seattle, at Washington State University,
Reifler, Micklesen, Wall and Hill were working mainly on Russian,
in collaboration with the International Telemeter Corporation of
 VARIATIONS IN APPROACH 31
Los Angeles, where Gilbert King had designed a photoscopic
memory, with high capacity and very rapid access, now being
further developed at the Rome (N.Y.) Air Force I.B.M. research
station. Everywhere translation from Russian into English had
high priority. Russian/English research groups were at work at
the University of California, at the California Institute of Tech-
nology, at the Ramo-Wooldridge Corporation of Los Angeles,
and at the Rand Corporation of Santa Monica, where Hays and
Harper had undertaken a most interesting series of studies in
methodology and were seeking to synthesize work already done.
The National Science Foundation of Washington was financing
much of this research and endeavouring to co-ordinate it.
In Great Britain, Booth, Brandwood and Cleave, at Birkbeck
College, with the financial help of the Nuffield Foundation, were
also studying the methodology of research and doing practical
work on Braille, French and German, much of which is described
in their book Mechanical Resolution of Linguistic Problems [7]—a
rich source of facts and ideas. The Cambridge Language Research
Unit, under the direction of Margaret Masterman, was engaged
in studies of lexicography and universal syntax and attempting to
apply to lexis and syntax the idea of a “mechanical thesaurus”,
while Richens explored the possibility of an algebraic type of
universal language.
In Moscow, centred round the Academy of Sciences of the
U.S.S.R. a more vigorous concentration of talent and effort
appears to have been achieved, although later evidence suggests
the development of conflicting schools of thought between the
various groups at work. In September 1956, the review Voprosy
Jazykoznanija (Linguistic Problems) began to devote regular space
to the problems of automatic translation—with particular emphasis
on the work of the Steklov Mathematical Institute and the
University of Leningrad. The more empirical school, that of
Panov and Bel’skaja, both working at the Institute of Precise
Mechanics and Computer Technology of the Academy of Sciences,
defined a general methodology and guiding principles for the
collaboration of linguists and electronics engineers. Korolev was
working with this group on the problems of code compression for
dictionary making, and Razumovskij on the automatic program-
ming of translation machines. Mel’čuk and O. S. Kulagina,
32 MACHINE TRANSLATION
collaborators of Ljapunov at the Steklov Institute, published a
most interesting study on translation from French into Russian.
K. T. Mološnaja, at the same Institute, demonstrated how the
work of Jespersen and Fries on structural linguistics can be used
for the resolution of syntactic differences between languages.
Fifty-six papers were presented to an important scientific and
technical conference held in Moscow in May 1957, among them
contributions on the automatic translation of German, English,
Hungarian and Chinese, as well as on experiments in translation
from French. The Conference noted the need for directing
linguistic studies so that linguistics might be treated as a
natural science, making extensive use of mathematical methods
of analysis.
A year later, another conference on automatic translation re-
sulted in the publication of abstracts of seventy-one contributions
on the subject [28]. Again news has been received of a Leningrad
conference held in April 1959 where 59 papers were read [28a].
It is clear that applied and mathematical linguistics are being
studied in the U.S.S.R. with new vigour and enthusiasm and with
noteworthy clashes of theoretical views.

THE EVOLUTION OF IDEAS

While the basic ideas leading to automatic translation have not
changed over the course of the years, they have greatly increased
in boldness of conception. The initial notion of a mere automatic
dictionary has given way to that of completely automatic, gram-
matically correct translation. This evolution is due mainly to the
rapid improvement in computer techniques, and to the systematic
analysis of language, which for the first time has been conducted
with completely objective methods and based on the potentialities
and also on the limitations of computer operation.
The need to consult an electronic dictionary being an essential
feature of all mechanical translation, it was natural that early
research should be concentrated on the principles and preparation
of such dictionaries, and on the automatic retrieval of a word in
language B which is equivalent to a given word in language A.
Automatic dictionary and signalization of meaning. Reduced to
its simplest expression, the automatic dictionary instantaneously
 VARIATIONS IN APPROACH 33
supplies, for a word in language A, one or more equivalents
in language B, by a simple operation of retrieval. The word in
language A is input and compared with a list of words stored
in the memory, i.e. the dictionary. When the signs representing one
of the dictionary words coincide with those of the input word
(i.e. in existing machines, when the result of the subtraction of the
two figures representing these two words is equal to zero) the
computer is instructed to print out the letters of the word in
language B which translates the word in language A.
If the word has only one meaning, or if all possible meanings
coincide with all the possible meanings of the word in language B,
the “match” is perfect, and the semantic unit of language A
representing the input word will immediately be represented at
output by the equivalent semantic unit in language B. But when-
ever a word has several meanings it is necessary either to output
several alternative translations, from among which the reader must
try to make his choice, or else some method must be devised of
discriminating from among the various meanings, in order to
make it possible for the computer to print at output the particular
meaning chosen. This necessitates a more complex programme for
the computer, which must be supplied with instructions based on
criteria for selection from among the several possible translations.
The machine can only embark on a sub-routine of this type if it
receives a signal which will start the execution of a new instruction.
Must we have recourse here to human intervention to provide this
signal (in which case the operation is no longer automatic) or can
some objective element contained either in the signs of the
written language or in the structure of the sentence call forth the
necessary order to start the sub-routine? At this point it became
essential to investigate the whole problem of signalling systems in
written language, which, while seemingly adequate for human
communication, at first appeared very incomplete by machine
standards.
The problem of the signalization of meaning is by no means
simple: not only is the word dog a multi-meaning semantic unit—
or to put it another way, several semantic units; it also contains
indications of grammatical value. If the machine can identify dogs
as being either the plural noun or the third person singular of the
present indicative of the verb, then the absence of the s in dog is
34 MACHINE TRANSLATION
also a signal, indicating either the singular noun, or the other
persons of the present indicative, or the infinitive, etc. Other
objective criteria, such as the presence or absence of to, of a
subject or article, enable us to determine whether we are dealing
with a noun or a verb, and if the word is identified as a noun, then
the absence of the s is a positive indication of the singular. Except
in cases of exceptional ambiguity, the human mind grasps the
meaning of a sentence by instantaneous interpretation of signals
of this kind, without consciously recognizing them as signals at all.
The machine, on the contrary, needs to identify them without
possibility of error.
Thus the question arose, how to complete the automatic bi-
lingual dictionary by incorporating into it, in such a way that they
could be recognized by the machine, all those signals which make
it possible for us subconsciously to identify with exactitude the
grammatical value of the word dog in a given English sentence?
The separation of affixes. Starting from the idea that syntax was
of minor importance in the understanding of a language, Booth
and Richens regarded the mechanical dictionary simply as a
catalogue of all the invariant semantic units—stems and affixes—
in language A accompanied by their equivalents in language B.
By a semantic unit must here be understood any linguistic element
—whether it be part of a word, a whole word or a group of words—
having a distinct meaning. Booth and Richens decided to separate
stems and endings in their dictionary entries.
The Latin am—thus represents the idea of loving, the conjuga-
tion being rendered by the addition of the verbal endings to this
stem. Rego will be represented by three stems, reg-, rex- and rect-;
the general rule being that when the derivatives of a word are
not formed by simply adding the affixes, the stem of each deriva-
tive must be entered separately in the dictionary [17]. In German,
for example, Bruder and Brüder will be entered separately.
If the equivalent of an input word does not figure in the
dictionary, the machine searches for the longest segment of this
word which corresponds exactly to an entry in the dictionary.
Entries are compared from left to right, and this comparison is
repeated after the identification of a first segment until all the
elements of a given word have been identified. The Spanish word
comprarlo would, for instance be decomposed as follows:
 VARIATIONS IN APPROACH 35

compr- buy
-ar- (infinitive)
-lo- the/it
This method does not solve all the problems of semantic units
composed of of several words, such as the French ne . . . pas, ne
. . . que or German disjunctive verbs. But the use of the magnetic
drum memory first made possible one solution put forward by
Booth and Richens which has now become standard practice: the
machine is instructed to translate the first part of a two-term
semantic unit only when it comes to the second part.
The results of the trial translations of Booth and Richens were
of a character likely to discourage the interest of linguists and to
suggest that mechanical translation experiments would lead only
to rudimentary and disappointing results. Nevertheless, Booth
and Richens, by separating stems and affixes, had laid the founda-
tions of a sound and sure method which will certainly be necessary
as long as the size and speed of memories play a preponderant role
in the economics of machine translation.
By so doing they established by implication a rule which has
proved of great importance in the study of language for automatic
translation: the practical needs of programme-making, rather than
scientific and historical norms, were their guide in separating
affixes from stems. In other words, the determination of the
dictionary stem, or base, was made without regard for historical
linguistics: it was a question not of pure, but of applied science.
Birth and death of the "pre-editor". Reifler, like all the pioneers,
was at first convinced that a human operator will have to partici-
pate actively in the work of the translating machine; human
intervention consisting in improving and supplementing the
signals contained in the alphabet and in written language. He
first defined with precision (though after a more profound analysis
of linguistic data, he later withdrew from this position) the idea
of pre-editing texts to facilitate the work of the machine and of
post-editing them after translation to facilitate reading.
We have just seen that the conventional signals of the alphabet
and punctuation do not explicitly represent all the linguistic
values of which the speaker or reader is nevertheless conscious.
For instance, the word enfant is recognized as singular only by the
36 MACHINE TRANSLATION
absence of any signal: its gender is not represented by any sign,
except, in certain cases, by an agreement of article, adjective or
participle: “l'enfant que j'ai rencontré" as opposed to "rencontré”.
Moreover the signalling system of language A does not correspond
to that of language B, nor do the omissions of the two systems
coincide. In order to translate, it is necessary to compensate for
the omission of signals, wherever this is required by the given pair
of languages A/B. Reifler suggested the arbitrary creation, for the
machine, of distinctive graphic signals supplementing those of
ordinary written language. These signals were to be inserted in the
text for translation by a pre-editor.
Thus the role of the pre-editor would be to provide the machine
with texts explicit from the graphic-semantic point of view.
Reifler even considered the idea of a supplementary spelling
system which would give to both machine and reader all the
signals necessary for the complete understanding of a translated
text.
The problem of complementary signalization arose at two levels:
grammatical—the signalization of the grammatical value of
polyvalent words—and non-grammatical—the signalization of the
semantic meaning of polysemantic words. It was also influenced
by the restricted possibilities offered by the electronic computers
existing in 1952. But the first M.I.T. conference was soon to
suggest that new machines would shortly make it possible to
extract from conventional writing, without complementary
signalization, all essential grammatical information. It remained
to be seen how they could determine the grammatical role of the
constituent parts of extemporized compounds in a language such
as German. If this problem could be solved, then it seemed
possible that translation could become completely automatic.
German Compounds. Reifler therefore set to work on German
substantive compounds, and in August 1952 he announced the
demise of the German pre-editor, who had now become super-
fluous.
The difficulties encountered were of several kinds. In the first
place the meaning of a compound word does not always depend on
its constituent parts. In such cases the solution must be to list
the compound separately in the dictionary together with its
translation. The second difficulty was christened “the x factor”.
 VARIATIONS IN APPROACH 37
In compound nouns, a letter or a group of letters may belong to
one or to the other constituent parts of the word. How can the
machine identify the constituent elements of Wachtraum in order
to decompose it either into wacht-raum, guardroom, or into
wach-traum, day-dreaming, the t being the x factor?
After classifying German substantives according to whether
they can form right- or left-bound compounds or both, by taking
as characteristic signals the space which separates the words, or
the absence of such space, as well as the initial capital letter of
nouns, Reifler observed that the formation of German substantive
compounds is governed by rules of such a nature that a maximum
of four checks with the dictionary is sufficient to identify with
certainty the grammatical role of the constituents of a compound,
in spite of the x factor.
Reifler’s work on compounds was completed by the elaboration
of a form-class filtering system of German words, based on some
of the ideas of the structural linguists and on the use of separate
memories for each form-class: four big magnetic drums contained
the four main word classes and ten less important classes were
placed on smaller drums. This was the first detailed classification
of the grammatical categories of a language made from the point
of view of the operation of a translating machine. New computer
techniques will modify the material basis of Reifler’s system, but
its linguistic basis is permanent and well adapted to computer
work.
Better than word-for-word translation. Reifler’s work brought
mechanical translation well beyond the point of simple word-for-
word dictionary translation; as soon as an attempt was made to
analyse the relationships between words in view of mechanization,
the translation of word groups became possible. The upholders
of word-for-word translation gave way to those advocating phrase-
by-phrase translation, that is, by units of meaning, and not by
dictionary words.
The analysis of words from the viewpoint of the operations of a
mindless machine had given rise to a clarification of the role of
words in the communication of the idea to be expressed: certain
words or forms have a clear and precise meaning quite independent
of any context—for instance wegen in wegen dieser Schüler (because
of these pupils). Other forms have multiple grammatical and
38 MACHINE TRANSLATION
non-grammatical meanings, and the reader must choose the mean-
ing appropriate to a given context on the basis of the information
provided by the form of neighbouring words. The meaning of one
form may be pin-pointed by another form, or there may be
mutual pin-pointing of two forms. In the example given above,
dieser may be a nominative masculine singular, a genitive or
dative feminine singular or a genitive plural; Schüler may be a
nominative, dative or accusative singular, or a nominative,
genitive or accusative plural. One has four possible functions, the
other six. Taken together they pin-point each other's meaning
since only two common functions remain: nominative singular or
genitive plural. Wegen, which governs the genitive, excludes any
possibility other than the genitive plural. This leads to the
formulation of a theory of the creation of sub-routines for the
exploration of immediate context, enabling the machine to find
the right translation for a word of multiple meaning or function.
In theory machine translation had by then passed the crucial
point where mere word-for-word equivalence gave way to a
partial rendering of the relationships between words, and to an
exploration of the modification of one word by another and the
mutual pin-pointing of meaning.
The Georgetown-I.B.M. experiment. Dostert and Garvin at
Georgetown University were working in the same direction when
they set up their 1954 experiment in collaboration with I.B.M.
The interest of this experiment, prepared “manually” first on
typewritten and subsequently on punched cards, is today mainly
historic.
The scope of the project was strictly limited: a vocabulary of
250 words selected in various fields—general, technical, scientific,
military and political. In the Russian/English dictionary a dis-
tinction was made between the indivisible Russian words and
those divisible into stem and ending, the stems and affixes being
stored in separate memories. The number of possible English
equivalents for each Russian element was limited to two, so that
when a choice was necessary for translation, that choice lay
between two possibilities only.
Certain diacritical signs were added to the alphabetic coding
of the letters representing the words. “Programme Initiating
Diacritics” (P.I.D.)—bringing into play one of the six rules of
 VARIATIONS IN APPROACH 39
syntax; “Choice Determining Diacritics” instructing the machine
to effect a reference forward or backwards (C.D.D. 1 and 2) from
the word under examination in order to search for the necessary
signal to determine the choice between two translations; a third
group “Address Diacritics” (A.D.D. 1 and 2) gave the dictionary
address of the English equivalents associated with P.I.D.’s and
C.D.D.’s
The six rules of syntax were briefly as follows:
Operation 0: Immediate translation of the given input word.
Operation 1: Reverse word order.
Operation 2: Choice dependent on a following word.
Operation 3: Choice dependent on a preceding word.
Operation 4: Omission of a word redundant in English.
Operation 5: Insertion of a word necessary in English.
The success of this experiment in automatic translation of
complete sentences without pre-editing or revision, compelled
attention and achieved widespread recognition of the progress
already accomplished in research on mechanical translation.
There was now no doubt that the automatic dictionary and the
stammering translations of the beginnings had been left far
behind. No doubt either but that the aim—the completely auto-
matic translation of any scientific or technical text—was still far
from being attained. Only at the cost of an enormous effort of
coding and programming had the I.B.M. 701 data-processing
machine been able to translate these 200 sentences at the rate of
six or seven seconds a sentence.
The experiment also drew attention to the importance of
linguistic problems in automatic translation. Up to this time the
general belief had been that only a few science-fiction enthusiasts
could be interested in a game of no possible value to linguists.
Proof was now forthcoming that, given a series of sentences in one
language, an electronic computer could print out a series of
sentences of equivalent meaning in another language. If the
machine was capable of doing this, then it was necessary to face
resolutely up to the basic problems of studying language with a
view to making use of the new potentialities of the machine.
1955—The turning point. Such study was almost at once taken
up by the mathematical specialists, electronicians and linguists of
40 MACHINE TRANSLATION
the Academy of Sciences of the U.S.S.R., while in England,
Booth was able to enlist the support of the Nuffield Foundation
for his research. It may be said that 1955 put the problem well
on the road to actual solution, with the accent firmly placed on the
study of language, and considerable progress made in the direction
of machine exploration of the constituent elements of the sentence.
Word-for-word translation is still considered useful, but it is
definitely out-of-date.
In the purely technical field, memories and reading units were
evolving rapidly. In an essay in Machine Translation of Languages
[17] Booth emphasized that reading speeds of 100,000 letters per
second made it possible to find a word in the permanent dictionary
in 1/50th of a second, and, with the introduction of ferrites,
offering new possibilities in the form of temporary memories,
only 1/100,000th to 1/1,000th of a second would be required.
Booth described the main characteristics of a translating
machine, as follows:
(a) An input, either in the form of a “reader” of original
typescript, or in the form of a magnetic tape reader.
(b) A rudimentary “computer”. This need only be capable of
subtracting, shifting letter patterns, recording results in storage,
and discriminating on the size of numbers.
(c) A small-capacity, reasonably high-speed computing storage.
This might be realized on a magnetic drum, but would more
probably take the form of a ferrite matrix. Experience suggested
that a capacity of 64 words, each of up to 12 letters, would be
adequate.
(d) The main dictionary and grammar storage. A large magnetic
or photographic drum would probably prove suitable if the
micro-glossary technique were used. The capacity of this organ
might be 10,000 words of 12 letters.
(e) From 4 to 28 tape feeds to input the various microglossaries
in the machine repertoire. Photographic film seemed the most
promising material for this.
(f) A single output magnetic tape. Only one is needed since this
medium can match the speed of the computer itself.
Booth estimated the cost of this machine, which would occupy
a floor space of 10 to 20 sq. ft., at roughly 100,000 dollars—
probably an under-estimate.
 VARIATIONS IN APPROACH 41

THE CONCRETE ANALYSIS OF LINGUISTIC DATA

The methods of Reifler, Booth and of the Russian school of
Panov and the Institute of Precise Mechanics are roughly con-
vergent: all of them base their work on the idea that the special
requirements of the treatment of linguistic data by computers
provide an instrument for linguistic analysis which enriches our
knowledge of language, an instrument capable of exploring the
differences existing between the systems of expression of two
languages. Theories are discarded in favour of an examination of
linguistic material in its complex relationship with the expression
of ideas, of an extension and development of the methods of
analysis used by information theory, involving penetration in
depth, below the level of the alphabet, down to the semantic and
syntactic elements of language, studied from the angle of the
reciprocal behaviour of two languages in process of translation.
If language A expresses an idea by certain means, and language
B by some other means, what are the rules determining the mutual
behaviour of these two systems? How can one formulate these
rules in such a way that a machine can apply them automatically to
convey in language B an idea expressed in language A? In order to
do this it will obviously be necessary first to draw up an inventory
of all the similarities and all the dissimilarities between the two
systems of expression, and then to submit this inventory to a
further analysis in order to enable the machine to use the results
of the inventory, for instance, in the form of binary numbers. It
is clear that this inventory and this methodology constitute the
point towards which the preliminary studies of the years up to
1955 converge. This does not mean that very considerable work of
mathematical analysis will not be necessary for translation pro-
gramming, but this work can only usefully be done after the vary-
ing methods used by two languages to give expression to the same
ideas have been fully inventoried and processed.
The Panov school in the U.S.S.R. sees in the problems of
mechanical translation one aspect of a group of questions to which
insufficient attention has been given by specialists of information
theory, who must learn to take into account the individual qualities
of the information communicated instead of being interested only
in its statistical characteristics [26]. Translation being an art in
42 MACHINE TRANSLATION
which perception of individual factors plays an essential role, a
whole vast new area lies open for investigation, beyond the
statistical analysis of the conventional elements of the alphabet and
other signals; it is necessary to explore the actual means of expres-
sion in all their aspects, alphabetic, morphological, syntactic and
semantic. Panov, recalling the highly promising work of Jespersen
and Fries [13], nevertheless emphasizes the limits of structural
linguistics as defined by these two authors; he observes that the
logical analysis of language cannot by itself provide a solution to
the problem of translation. As for the possibility of reducing the
structure of language to mathematical formulae—another very
tempting prospect—Panov writes: “Should it be achieved, the
problem of automatic translation would join as equal those pro-
found problems united under the name of theory of information.
Unfortunately, so it seems to me, we must refrain from this
tempting road. The very nature of the problem of translation is such
that individual features of the translated text cannot quite be ignored.”
These individual features are found in the lexical content of the
text. This fact, writes Panov, should govern our choice of methods:
“I believe here we are faced with a problem which, though
statistical in character, requires methods of analysis, similar to the
experimental methods used in the study of natural phenomena.”
The Soviet mathematician concludes that “both lexical meaning
and grammatical characteristics of the word can and should be
considered in translating languages” and that “it would be highly
impractical to decline the information which can be thus obtained.”
In the analysis of the text to be translated, he refuses to separate
completely lexis, morphology and syntax, since all three elements
contribute to determine the meaning of the text.
The basic principles of early Soviet research. Soviet research
started in the Institute of Precise Mechanics and Computer
Technology, on Academician Lebedev’s B.E.S.M. computer.
Almost simultaneously Ljapunov, Mel’čuk and Kulagina worked
at the Steklov Institute of Mathematics on a smaller computer,
the STRELA, and followed slightly different methods, more
directly inspired by the structuralists. The efforts of the Panov
group were directed less towards a theoretical comprehension of
the general problem of machine translation than towards a detailed
investigation of lexical material. The Steklov Institute group on
 VARIATIONS IN APPROACH 43
the other hand were concerned with profound theoretical research
in the area of mathematics and linguistics, and saw mechanical
translation as part of the larger problem of the automation of
thought processes. They endeavoured in particular to determine
the correspondence between the grammatical structures of two
given languages. The work of Ljapunov, aiming at gradual
automation of the whole process of machine translation, inspired
the research of several other groups: in the Institute of Linguistics
of the U.S.S.R., universal rules for the analysis and synthesis of
a text are being worked out by Mel’čuk, and at the Experimental
Laboratory of Machine Translation of Leningrad State Univer-
sity, N. D. Andreev is also working on some abstract logical
system capable of serving as an intermediary language, constructed
by averaging the phenomena of various languages.
Yet despite some marked and sometimes even vehement
expressions of divergence as to theoretical approach, the Soviet
scientists and linguists are in the main following certain practical
principles which were developed in 1956 by the Panov group—the
more empirically minded and also until now the only one which
has actually achieved limited but genuine translation by com-
puter techniques. Those principles are probably fundamental to
all practical achievement and may hold the key to future progress.
From the very beginning of their work on machine translation
the workers of the Institute of Precise Mechanics decided to
organize their research on bases quite different from those of the
Georgetown-I.B.M. experiment. “In our opinion,” writes
Panov, “excessive contact between the translation programme and
the ascription of the control codes directly to the words in the
dictionary cannot but limit the possibilities of translation, making
the solution of the problem extremely complicated. Therefore we
made it our point to work out basic principles of machine transla-
tion before starting. Our five basic principles are the following:

1. Maximum separation of the dictionary from the translation

programme. This enables us easily to enlarge the dictionary
without changing the programme.
2. Division of the translation programme into two independent
parts: analysis of the foreign language sentence, and synthesis
of the corresponding Russian sentence. This enables us to
44 MACHINE TRANSLATION
utilize the same Russian synthesis programme in translation
from any languages.
3. Storing all the words in the dictionary in their basic form. This
enables us to make use of the standard Russian grammar in the
synthesis of Russian words.
4. Storing in the dictionary a set of invariant grammatical charac-
teristics of a word.
5. Determination of multiple meaning of the words from the
context whereas their variant grammatical characteristics are
defined by analysing the grammatical structure of the sentence.

These principles have proved quite reliable in the practical

test they were put to, and hence they must be considered as basic
in the solution of the problem.” [26]
Dr Panov has not deviated from his opinion that these five
principles will lead to a complete solution of the problem. It does
indeed seem that they may be able to serve as the basis for the
greater part of research, whatever pair of languages are under
consideration and whatever the differences between them—
whether they be languages rich in inflexions such as Russian or
poor in inflexion and rich in structural meaning such as English.
Details of the programme will vary, but the application of these
principles will remain the basis of research of a very flexible
nature, and will avoid much exploration of blind alleys.
To Booth and Richens, to Reifler and the M.I.T. team, to
Panov and his colleagues, must go the main credit for clearing the
ground and laying the solid foundations upon which it is now
possible to build.
 CHAPTER IV

From Source Language to Target

Language
INVENTORY OF MEANS OF EXPRESSION
To translate from a given language A into a given language B
is to attempt to reconstitute with the system of expression of
language B, the meaning of a sentence or string of sentences,
expressed in language A by means of the system of expression
peculiar to that language. The meaning of a sentence is the repre-
sentation in the speaker's mind, materialized by means of phonetic
and visual symbols grouped into words. Each word possesses, or
may possess, several values, semantic or grammatical. Each word
may be syntactically associated with other words in a number of
ways. Perception of meaning is dependent on the determination
of these different values and associations. Translation becomes
possible only after an analysis of all the linguistic elements of
language A, or source language, constituting meaning, embodied
in the words and in the relations between words, i.e. semantic
values, grammatical values (whether expressed by inflexions or
otherwise) and syntactic values.
This analysis is followed by a synthesis of the linguistic ele-
ments of language B, or target language, selected because they
make it possible to render approximately the same meaning as the
original sentence in language A, and combined according to the
rules peculiar to language B.
The experienced human translator performs this operation with
a degree of difficulty which varies with the clarity of the text for
translation and also with the degree of similarity between the
structures and semantic content of languages A and B. Before the
translator can be replaced by a machine, it will be necessary to
prepare all the operations of analysis, of synthesis and of enumer-
ation required for the elaboration of programmes and sub-routines
46 MACHINE TRANSLATION
enabling the electronic machine to transpose the content of a text
in language A into a text in language B. The principal contri-
bution of the linguist to the preparation of programmes and sub-
routines will consist in making an inventory of all the means of
expression of languages A and B, and of the relationships
which can be established between their respective systems of
expression.
The linguist will endeavour to draw up this inventory in such a
way that it is easily reducible to numerical data and coded in-
structions for use by the machine. Clearly, the more closely the
means of expression of the two languages resemble one another,
the simpler the programme. The greater the difference between
the structures, the morphology and the semantics of the two
languages, the more numerous the ramifications of the programme
or sub-routines required, that is to say the more circuitous the
ways of finding exact equivalents between a word in language A
and a word in language B. It is therefore true to say that pro-
gramme economy will depend directly on the degree of structural
relationship between the two languages under consideration.

LANGUAGES, INTERLANGUAGE AND METALANGUAGE

Considerations of economy, as well as certain preconceived ideas
of a philosophical and logical nature, have given rise to a somewhat
premature discussion concerning an intermediary language, or
interlanguage, the use of which, it is alleged, might facilitate
automatic translation. The desire to find such a philosopher's
stone of machine translation appears to have been enhanced by
the early absence of any firmly established basis for more clearly
rewarding empirical research.
The issues of this debate have not been clarified by the inevit-
able talk of “machine language”. And confusion has only been
increased by the idea that the binary code controlling the actions
of computers is itself a language. In a field in which language and
languages are both the subject and the instrument of research,
confusion due to purely verbal analogies and to the metaphorical
use of the word language is very frequent. Fortunately Andreev,
a Russian, and Mounin, a Frenchman, have, each in his own
fashion, contributed to a better understanding of this complex
subject in which the utmost exactitude in terminology is essential.
 FROM SOURCE LANGUAGE TO TARGET LANGUAGE 47
Georges Mounin rightly distinguishes between pseudo-
languages—of which Esperanto is the classic example—intended
to be speakable, and interlanguages, designed for use as auxiliary
languages, such as the interlingua of Peano or that of Gode and
Blair. He points out that the present work on machine translation
must lead to the study of the problem of languages “which can be
used as common intermediaries, as central links in the chain of
translation from any language into any other”. [21]
Here we must make a distinction between two ideas: between
that of an interlanguage set up a posteriori as the result of research
undertaken on a number of pairs of languages analysed in view
of automatic translation from one to another—which would in
fact express the highest common denominator of the means of
expression of all these languages—and that of an interlanguage
conceived a priori as a universal translation programme applicable
to all languages. Whereas the first would be an outcome, the
second would be a starting point. This means that before designing
and working on translating machines it would be necessary first
to draw up a universal programme. Is such a programme really
desirable? Is this a priori possible?
Booth, Brandwood and Cleave [7] have demonstrated the weak-
ness of the economic arguments invoked in favour of the universal
programme thesis. For N languages, we are told, we should need
N-1 programmes in order to translate, without an intermediary
language, from each one of these languages into another. And to
translate N languages into N-1 languages, we should need
N (N-1) programmes, i.e. almost N2. On the other hand, it is
maintained, the use of an artificial intermediate language M would
require only N programmes to pass from N languages into this
language M and as many again to go from M into N languages,
i.e. 2N programmes in all instead of N2.
Booth easily refutes this by a mathematical argument. If the
turntable language selected is not an invented language M, but a
real language L1, we should need not 2N but 2N—2 (i.e. two less)
programmes to translate, via language L1, all languages into all
other languages. Thus a natural language would be a more
economical turntable than an artificial one.
Booth also puts forward arguments founded on common sense
and observation. We are told that the artificial language would
48 MACHINE TRANSLATION
facilitate the passage from language A into language B if the two
languages are very dissimilar in structure. If we examine this
contention closely, we shall see that language M, the entirely
artificially-created language, does not in fact simplify anything.
The elaboration of a translation programme A→B represents a
sum of work less great than that of a double programme A→M
and M→B. Moreover, if, as is proved by observation, programmes
are not in fact reversible, the establishment of programmes
A→M→B and B→M→A would be more costly than that
of programmes A→B and B→A.
Thus we are brought back to the comparative and empirical
study of languages A and B as the only practical method of setting
up translation programmes A→B and B→A. And it is
certainly true that if ever an intermediary language (or a universal
programme of translation) becomes possible, this intermediary
language is more likely to be of real use if its structure and its
characteristics are experimentally based on the comparative study
of multiple bilateral programmes of the type A→B and B→A.
Both the British and the Russians are, however, moving to-
wards the use of their own languages as natural turntables,
justifying their choice by various practical considerations. In
spite of certain objections due to the phonetic and structural
ambiguities of English, Booth sees in his own language a possible
pivot, even if this should involve slight modifications of current
English in the interests of its universal use. Panov, for his part,
envisages the use of Russian as the basic language. Andreev [1]
employs very strong theoretical arguments to justify the use of a
language closely related to Russian. Leaving aside all questions of
national ambition, can this coincidence be purely fortuitous?
When the work of analysis of one or more foreign languages has
been successfully completed, is not each national team naturally
inclined to seek the general solution most convenient to itself?
This solution naturally consists in using as turntable that pre-
linguistic state prior to output when linguistic data are still
expressed in code and are on the point of being transformed into
English for some and into Russian for others. So that we are forced
to admit, leaving aside for the moment Andreev's theoretical
arguments, that other natural claims to the title of turntable might
equally well be put forward. In point of fact any language spoken
 FROM SOURCE LANGUAGE TO TARGET LANGUAGE 49
by a nation dynamic enough to carry out a number of bilateral
translation programmes into its own language can also claim to be
a natural turntable language. As we shall see below, however, the
optimum choice may depend on certain factors inherent in the
structure of this language.
While an intermediary translation machine language remains,
for the moment, if not a somewhat academic dream, at best a very
long-term project, this is not true of a metalanguage for the use of
specialists in automatic translation. The automatic transposition
of language A into language B can only be performed in accord-
ance with a strict system of instructions given to a machine,
which responds to numerical codes representing the letters of
words, grammatical forms and syntactic relationships between
words. The programme of operations is also controlled by nu-
merical codes. The analysis and synthesis of linguistic data
necessitates an exact inventory of these data, and this inventory
must be made in terms which can be reduced into codes. In this
context it is possible to employ the word metalanguage in a
restricted but active sense, as does Andreev:
1. We call a metalanguage any linear system of signs used
for the written designation of the elements in a particular
system of ideas and the relations between these elements.
2. The class of metalanguages at the present time comprises
mathematics, physics, chemistry, formal genetics, and
symbolic logic.
3. A special metalanguage in the symbols of which the facts
and relationships of the language systems may be described
that are subject to equivalent comparison, needs to be
developed for the preparation of algorithms for machine
translation.
4. The symbols used in the metalanguage of machine transla-
tion are regarded as metalanguage words and grouped in
categories analogous to the parts of speech. [28]
Thus the metalanguage envisaged by Andreev is a system of
symbols expressing linguistic elements and the relations between
these elements, the study of which should make it possible to
prepare automatic translation programmes. These symbols
would be to automatic translation what H2 O and NaCl are to
50 MACHINE TRANSLATION
chemistry. They would be immediately comprehensible to
specialists, but of no use to the uninitiated. An international
language requiring no translation, they would be immediately
transposable into machine codes. In this restricted sense, the idea
of a metalanguage resembles that of a strictly closed semantic
system of expression such as algebraic or chemical symbols.
It is not unlike the “linguist to computer” code of instructions
recently devised by M.I.T. research teams under the name of
COMIT. Such a metalanguage can never be a substitute for
natural languages, but may make it possible to study such lan-
guages with a degree of precision not easily attainable by the use
of ordinary language, in which meaning is frequently distorted by
analogy, metaphor and all the fluctuations of semantics.
Translating-machine metalanguage would be a highly specialized
instrument for use by those engaged in facilitating translation;
it would help them to define and fix their ideas with greater
precision and also to communicate these ideas to the machine in
completely unambiguous form; the symbols of this metalanguage
would set in motion the machine operations corresponding to
linguistic programmes or sub-routines. On this level it is not a
question of philosophy, linguistics or logic, but of the creation of
an instrument by means of which the human mind would be able
to explore the realities of language and at the same time control
and discipline the mechanical forces which unconsciously simulate
certain functions of the conscious mind. It would be an instrument
of communication between certain specialists, an instrument for
the control of the machine—a language used as a precision tool,
working on language as its object, making possible by its very
precision the exploration of this object, at the same time respecting
the imprecisions and illogicalities inherent in its nature.

THE LINGUISTIC MOULD OF REPRESENTATION

Applied to linguistic data this instrument should make it possible
to reconstitute the meaning of the sentence translated, that is to
say to enable the reader of language B to participate in the
representation of the speaker or writer of language A. The question
the translator most frequently asks himself is: “What does this
mean?” and linguistic data are to him inseparable from their
meaning. The machine, even more than the human translator,
 FROM SOURCE LANGUAGE TO TARGET LANGUAGE 51
will have to approach meaning, so closely bound up with the
representations of writer and reader, by the narrow path of
material signs expressing representation within the limits of the
expressive systems of the languages employed. These limits are
imposed by the choice and meaning of words, by the forms such
words may take, by the relationships existing between them:
semantics, morphology and syntax all mould and frame thought.
Each language is in effect a collective system of expression, a
framework or mould largely prefabricated by the social life of a
human group. It is rare when the rigid outlines of this mould
completely coincide with the individual forms of the representa-
tion we are trying to communicate.
Now the moulds for the expression of a thought offered by
two different languages are rarely identical. The prefabricated
elements of these moulds have neither the same shape nor the
same dimensions, except in the case of strictly scientific texts for
which scientists have created means of expression not far removed
from metalanguage.
Therefore, in order to reconstruct the meaning of a sentence
in language A with the semes, grammatical forms and syntactic
structures of language B, the machine must identify each of the
linguistic elements of the sentence, including the relationship
between words, must assign to each of these a code number, which
in turn must be able to call for and deliver at output the semes,
inflexions and other morphemes, as well as the equivalent syn-
tactic relationships of language B, thereby making a compre-
hensible sentence in that language. The mesh of different linguistic
elements in the sentence in language A, carefully disentangled by
the machine, will be replaced in the memory-registers of this
machine by a series of codes sorted out into strict order, to be
recombined afresh in a different network composed of those
elements of language B which reproduce the meaning of the
original sentence. How can this be accomplished?

HIEROGLYPHIC CONVERSION
The problem being that of respecting meaning to the maximum
extent compatible with the necessity of rendering the sentence
pliable to the demands of the machine’s codes, the first step is to
decompose the sequences of words and their inter-relationships
52 MACHINE TRANSLATION
in language A, and to express them by a series of equivalent
figures, capable in turn of being transformed later into a meaning-
ful sequence of words in language B. Andreev gives the name
“hieroglyphs” to these numerical codes which represent semes,
forms and structures. He divides them into three classes: semantic
hieroglyphs, formal hieroglyphs, and tectonic hieroglyphs.
When the input text is coded, numerical symbols are
obtained for the ideas contained in lexical units (semantic
hieroglyphs), numerical symbols for grammatical morphemes
and symbols for link words (formal hieroglyphs), and numerical
symbols for word order and syntagmatic relationships between
words which are not expressed phonemically (tectonic hiero-
glyphs). In decoding, corresponding hieroglyphs determine the
choice of words, their grammatical formation, and the methods
of their combination in the output language, [1]
The main task of the machine—the fundamental linking factor
in machine translation—is hieroglyphic conversion. The basic
pattern of automatic translation consists of three principal
phases: analysis, or coding of information given in the input
sentence; conversion, or the substitution of one code for another;
synthesis, or the decoding of the converted information into a text
in the output language. In bilateral translation programmes of the
type A→B, analysis and conversion are carried out simultane-
ously and are conditioned by the need to arrive at the linguistic
framework of language B.
It is, however, possible to envisage other methods, for example
that of analysing all the inherent forms and grammatical re-
lationships of language A, entirely disregarding any “output”
language B. Language A would be coded in complete isolation,
every word being analysed according to all the inherent gram-
matical forms of that language, whereas in A→B translation
analysis is restricted to the differences between the characteristics
of the two languages.
For instance, to translate into French her fine clothes, it is
necessary to effect an analysis which will determine the fact that
fine is plural although without any visible plural sign, so that the
French adjective will be made to agree with the noun habits when
the moment comes for synthesis. If, on the other hand, the
 FROM SOURCE LANGUAGE TO TARGET LANGUAGE 53
machine were translating from English into a language in which
adjectives do not agree, such an analysis would be superfluous.
Analysis must also determine the number and gender of the word
habits, so as to obtain correct translation of the possessive adjective
her, whereas in translating into French it will not be necessary
to determine the gender of the possessor, even though this is
indicated in English. If the output language declines substantives,
further analysis of clothes will be required in order to indicate
whether the noun which translates it is in the nominative, genitive
or any other case. Such analysis is not needed for translation into
French.
If the aim of the analysis is to translate from one language into
any other language, it is obvious that it must be as complete as
possible for each part of speech, whereas for translation of the
type A→B, it need only be partial, its extent depending on the
degree of parallelism between the structures of the two languages.
Given the hypothesis of an analysis for universal application for
translation of the type A→X, it will be sufficient, after a com-
plete analysis of language A, to draw up for each pair of languages,
conversion tables for the hieroglyphs representing the elements
of the analysis. Andreev reminds us that:
Analysis and synthesis are constant values for every language,
and are determined exclusively by the norms of a given language
and by the principles of coding. Hieroglyphic conversion, on
the other hand, is a variable value, and is a function of both
input and output hieroglyphs, [11]
Three main types of difference exists between input and output
hieroglyphs:
(1) The suppression of superfluous hieroglyphs: German die
Sprache—language—coding of the article die is superfluous; her
dress, sa robe—the hieroglyph for the gender of the possessor is
superfluous.
(2)Introduction of additional hieroglyphs: Japanese ani no hon
—elder brother's book—Japanese having no genitive case, an
additional hieroglyph denoting this case is required and must be
introduced for English; her dress—sa robe—the hieroglyph
denoting the gender of the object possessed is required and must
be added.
54 MACHINE TRANSLATION
(3) Modification of the type of hieroglyph: to catch cold—
French s'enrhumer, Russian prostudit'sja—one of the English
words, catch, is represented by a semantic hieroglyph whereas in
French and Russian this must be replaced by a formal hieroglyph
denoting the reflexive verb: to make clear—clarifier—the semantic
hieroglyph for make must be replaced by a formal hieroglyph of a
verbal type.

LINGUISTIC ANALYSIS BY MACHINE

The accounts of Soviet experiments provide the best concrete
illustrations of the methods employed to transform words into
“numerical equivalents” or codes stored temporarily in the
memory-registers of the machine and making it possible to
synthetize the sentence in another language. In order to under-
stand these illustrations fully it is essential to be acquainted with
the broad outlines of the translation programme drawn up for
the B.E.S.M. computer by the Institute of Precise Mechanics
and Computer Techniques of the Academy of Sciences of the
U.S.S.R.
Figure 2 gives the general layout of the programme: input of
English text; analysis, divided into two main phases—vocabulary
and parts of speech; synthesis—also divided into vocabulary and
parts of speech; output, or printing out of Russian text. Opera-
tions take place in descending order. Input is in Baudot telegraphic
code.
At this point a brief description of the electronic dictionary
here employed must be given. As shown in Figure 2, it has two
sections—English and Russian. The first part contains, in
numerical code, the English words accompanied by:
(a) Certain permanent information concerning each word,
including an order number indicating its place in the English
section of the dictionary.
(b) Where applicable, the order number of the corresponding
Russian word, making it possible to obtain certain permanent
information concerning this word.
(c) If the Russian word is not given, an instruction referring
back to the polysemantic dictionary;
(d) Appropriate instructions concerning the next sub-routine
to be performed.
56 MACHINE TRANSLATION
The Russian section of the dictionary contains the numerical
codes representing the letters which compose the Russian words,
and the permanent information concerning these words. It
enables the machine to construct Russian sentences at output, by
reassociating the stems of these words with the endings in accord-
ance with the indications recorded in the memory register during
the analysis of the English sentence.
English words having more than one meaning are placed in a
polysemantic or supplementary dictionary which makes it possible
to choose, from among several meanings of the same word, the
correct Russian equivalent for the given context. When the
machine looks up a polysemantic word in the English section of
the dictionary, it finds not a Russian order number but a code
instructing it to look for this word in the polysemantic dictionary,
where it will receive supplementary instructions setting in motion
the particular sub-routine for deciding the meaning of this word
(see below, sub-routine to determine the meaning of of).
If the spelling of an input word corresponds exactly to that of a
word in the dictionary the word is identified and the programme
continues immediately in accordance with the dictionary indica-
tions. If, on the contrary, the word is not in the dictionary (e.g.
the word walked) the machine immediately embarks on the sub-
routine for the reduction of inflected endings (see Figure 3)
which, in the case of the English perfect tense, makes it possible
at the fourth attempt to detach the ending -ed and subsequently to
look up and identify the “base” walk in the dictionary.
We must now return to Figure 2 and the analysis of the English
sentence. This analysis consists of a series of sub-routines, all
designated by the names of parts of speech, with two exceptions—
the sub-routines for formulae and syntax. This programme was
designed for the translation of mathematical texts, and mathe-
matical formulae (in which Russian M.T. scholars include foreign
words and some proper names) require no translation. The name
for the “syntax” routine is self-explanatory.
For the various parts of speech, a special analytical sub-routine
aims at identifying and inscribing in the machine's memory-
register, in coded form, all the information necessary for the
construction of a grammatically correct Russian sentence equiva-
lent in meaning to the English sentence. When a word is identified
 FROM SOURCE LANGUAGE TO TARGET LANGUAGE 57
in the dictionary as being a preposition, the sub-routine for pre-
positions is put into action, and the machine proceeds to a series
of check-ups concerning the context of the word; such check-ups
are to determine whether the word is to be translated into Russian
or not, what case is governed by the Russian preposition which
translates it, or, if it is not to be translated, what case the following
word should be, etc.
Similarly, the sub-routine “English nouns” submits the English
noun or pronoun on which the machine is working, once it has
been identified as a substantive or its equivalent, to a series of
check-ups to determine its grammatical role, its case, gender,
number, etc. These check-ups are made by the method of dicho-
tomy: that is to say by asking a series of questions which can
receive either a negative or an affirmative answer. If the reply is
affirmative, the machine classifies the information received in its
memory-register. If it is negative it continues its search in a pre-
determined order. In the memory-register two “cells” are allocated
to each word: in the first cell is inscribed on the right the order
number of the English word, then, in fixed positions from left to
right, the affirmative or negative indications received in reply to
each of the questions applicable to this word. The sub-routines are
performed in a pre-established order enabling all the divisions of
the cell to be filled in. The number of divisions varies with the
parts of speech concerned: two for adverbs and prepositions, four
for conjunctions, eight for cardinal numbers, seventeen for nouns,
eighteen for adjectives and verbs. These divisions are always
followed by a four-figure division for the order number of the
word. The second cell contains only the order number of the
equivalent Russian word in the Russian section of the dictionary.
It will be observed that the sub-routine for ordinal numbers—
"numerical adjectives"—is justified not by any peculiarities in their
behaviour in English, but by the fact that in Russian they are
declined. Let us observe, too, the dotted line that links “verbs”
and “adjectives”: the Russian verb has numerous adjectival forms
which are declinable. Their behaviour is therefore both verbal and
adjectival, with declension and conjugation, and it is therefore
necessary to return to the verbal sub-routine after analysis of
certain adjectival forms such as English participles.
Once the analysis is complete, each English word in the sentence
58 MACHINE TRANSLATION
appears in the memory register in the form of an order number,
accompanied by all the coded information defining its grammatical
role and all the information required to decline or conjugate the
corresponding Russian word. Panov calls this the numerical
equivalent of the English word. It remains only to apply the sub-
routine “change of word-order”, according to the indications
contained in the divisions of the first cell.
SOME EXAMPLES OF SUB-ROUTINES
A few examples will help us to visualize in concrete fashion some
of these sub-routines. The first example [23] illustrates a case
of multiple meaning: the determination of the exact meaning
in Russian of the English preposition of. In the table given
below the notation 1.(3,2) signifies that if the result of the first
operation is affirmative operation (3) should be performed; if the
result is negative, then the machine should perform operation (2),
and so on. The notation 3.(0, 0) signifies that the sub-routine is
terminated and that the result should be recorded in the memory-
register.
1.(3, 2) Check preceding word for is, are, was, were, be.
2.(3, 4) Check following word for formula, or course.
3.(0, 0) This particle is not translated.
4.(5, 6) Check preceding word for idea or discussion.
5.(0, 0) Prepositional case: Russian translation o.
6. (7, 8) Check preceding word for true or productive.
7.(0, 0) Russian preposition dlja followed by genitive.
8.(9, 10) Check preceding words for fall short or in place.
9.(0, 0) Preposition not translated. Following noun in genitive
case.
10.(11, 12) Check preceding word for out.
11.(0,0) Genitive case: translation iz.
12.(13, 14) Check preceding word for incapable.
13.(0, 0) Preposition k plus dative case.
14.(15, 16) Check following word for necessity.
15.(0,0) Dative case: translation po.
16.(17, 17) Genitive case.
17.(18, 20) Check preceding word for noun.
18.(19, 20) Check following word for noun, cardinal number, or
formula not followed by a noun.
 FROM SOURCE LANGUAGE TO TARGET LANGUAGE 59
19.(0, 0) Preposition not translated. Following word in genitive
case.
20.(21, 22) Check the preceding word for consist, each, one, some
and the following word for all or them.
21.(0, 0) Translation iz.
22.(0, 0) Translation ot.
Close study of this example will show that most possible trans-
lations of “of” are included, beginning with those occurring
least frequently and ending with those most often encountered.

Grammatical analysis. Another series of examples will show how

the figures representing grammatical characteristics which will, at
the appropriate moment, control the synthesis of the Russian
sentence, are inscribed in the memory register. Panov, in his
brochure Automatic Translation [25], demonstrates his method
by means of the following English sentence:
This is true certainly of the vast category of problems
associated with force and motion.
We shall here limit ourselves to examining his illustrations for
the words this, certainly, of (twice) and associated, completing
them in certain details with the aid of Muhin’s brochure. [23]

This
The machine looks up and finds in the dictionary the English
order number—1115—but fails to find a corresponding order
number for a Russian word. The absence of such an order number
refers it to the supplementary dictionary for polysemantic words.
The information entered in the cell now reads as follows:

The polysemantic dictionary sets in motion the sub-routine for

determining the Russian meaning of this which will make it
possible to fill in the first cell of the memory register by five
figures signifying: 1: noun; 1: singular; 3: neuter; 1: nominative;
0: hard stem.
60 MACHINE TRANSLATION

Having now been identified as a noun for the requirements of

the translation programme, the word is submitted to the check-ups
prescribed by the sub-routine “English nouns”. As a result the
Russian order number of the word eto is entered in the second
cell and the first cell is completed as follows (reading from left
to right):*1: noun; first 0: is declined like an adjective;† 1:
invariable; 1: singular: 1: nominative; 3: neuter; 1: there is an
indication of number; 1: subject; 0: hard stem (in Russian); the
last 0 signifies: absence of an instruction to "omit word".

Certainly
The dictionary immediately provides the order number for the
Russian word bezuslovno—2257—together with the indications 5:
adverb; 1: interpolated word; English order number 0132. The
Russian order number being given, there is no need to consult the
supplementary dictionary. The cells are filled in as follows:

Of (the vast category)

The dictionary indicates: preposition, English order number
0472: it orders the sub-routine “English prepositions”: the cell
reads as follows:

the 6 indicating “preposition”. In the course of executing the sub-

routine for this preposition, which involves determining which
one of its multiple meanings is required, the machine finds
*Zeros in sub-divisions 3, 6, 9, 10, 13, 15, 16, of the cell are not relevant to
the case under consideration—although of course each sub-division must be
filled, either by a zero or by another figure.
† This means in fact that this ‘noun’ eto is a pronoun, and the following sub-
division indicates that eto is invariable.
 FROM SOURCE LANGUAGE TO TARGET LANGUAGE 61
(operations 6 and 7 of the example quoted above) the translation
dlja followed by the genitive, and enters in the second cell the
Russian order number for dlja, 5046. The first cell is now filled
in as follows:

the 2 signifying “governs the genitive”

Of (problems)
The machine operates exactly as for the preceding example and
in executing the sub-routine for prepositions receives the answer
“is not translated, governs the genitive” (see above, operations
17, 18, 19). The absence of a Russian order number in the second
cell means that the word of will not be translated at output, but the
inscription of a 2 in the appropriate division of the first cell will
ensure that the following word (the translation of problems) is
declined in the genitive case. The first cell will thus be filled in
exactly as in the case of the preceding of, and the Russian cell will
be empty, reading as shown below:

Associated
The dictionary check up for this word revealing no equivalent,
the programme for the separation of inflected endings is put into
action (see Figure 3); when the ending -ed has been eliminated, the
word associat- is found, accompanied by the following information:
2: verb; 1: 1st conjugation; 4: governs the accusative; 0: im-
perfective aspect; 3: has a desinence -ed; English order number
0085; the fact that there exists a Russian order number (2140)
refers the machine to the sub-routine “English verbs”. The cell
is now partially completed thus:
62 MACHINE TRANSLATION
The machine executes the sub-routine “English verbs” and
finds the information shown below in its numerical form in the
18 sub-divisions of the cell and explained step by step. The
figures in brackets are the numbers of the sub-divisions, which
have been added in order to facilitate reference.

Sub-division
number
(1)3: adjective (the participle being the adjectival form of the
verb, the 2 originally recorded in this division is now
replaced by a 3);
(2) 1: soft stem,
(3) 1: 1st conjugation (we are dealing with a verb and this
information is needed for synthesis of the participle);
(4) 0: stem ending neither in a sibilant nor a guttural;
(5) 0: variable word;
(6) 0: plural;
(7) 0: not a predicate;
(8) 2: genitive;
(9) 2: feminine;
(10) 0: designates an inanimate object;
(11) 1 : takes the shortened form;
(12) 0: has no indication of number;
(13) 3: participle;
(14) 0: past;
(15) 0: not subject;
(16) 0: not governing a case;
(17) 0: no indication “omit this word”;
(18) 1: the English word has ending -ed.

ANALYSIS AS PRE-SYNTHESIS
These examples demonstrate how analysis and hieroglyphic
conversion are combined in the programmes illustrated: we are
dealing with bilateral programmes of the type A→B, and not
with a universal programme of the type A→M→B. Here,
 FROM SOURCE LANGUAGE TO TARGET LANGUAGE 63
analysis is pre-synthesis—the study of the English sentence being
directed towards and conditioned by the needs of the Russian
synthesis. If the machine records a code meaning “feminine
gender” after the word associated, it is only because the Russian
word corresponding to problems is feminine.
It is perfectly normal and legitimate thus to undertake a
thorough check-up on the role of the English words in the sentence
solely with a view to synthesis into Russian. But it will be ob-
served that to translate into French the word group the vast
category of problems would require a quite different and much
simpler analysis. The Russian translation omits the and of, French
would translate them. Russian declines vast, category, problems.
In French these words present problems of number and gender
only. Analysis for translation into French would be shorter and
simpler than for translation into Russian. The same is true of
certain polysemantic problems: in both instances the word “of”
would be translated by de in French.

TOWARDS A MULTILATERAL PROGRAMME?

It is, however, possible to see how the Russian method illustrated
above, although bilateral, is potentially capable of generalization.
When the memory cell for each word of a sentence has been filled
in, the analysis being complete and the synthesis not yet begun,
the machine is in the same state as a translator who has looked up
all his words and analysed all their inter-relationships, but has
not yet begun to write. His translation is “in his head”. The
machine has “in its memory” numerical hieroglyphs representing
semes, grammatical forms, syntactic structures—which it can, at
a given signal, transcribe into Russian words correctly ordered,
conjugated and declined.
Is it possible that instead of Russian words (language B) it
could align French words (language C) corresponding to the
English sentence (language A) and to the hieroglyphs representing
it?
The machine can certainly do this if the number of categories
according to which the analysis has been made in view of language
B is superior or equal to the number of categories necessary for
language C, or conceivably if it brings into play a supplementary
programme of analysis B→C when language C requires a more
64 MACHINE TRANSLATION
complex analysis than that which was required for language B.
It is equally clear that a universal programme must include in
its analysis all conceivable grammatical categories for each part
of speech and all their combinations in order to provide all the
necessary indications for synthesis into any target language. Only
practical research spread over a number of languages can make it
possible to draw up such a programme, and this research must at
first be of a bilateral type, i.e. language A into language B.
The Academy of Sciences of the U.S.S.R. sees in its methods
of analysis and synthesis a practical and almost immediately
realizable solution to the problem of an intermediary language.
Panov has illustrated this belief in the following formula. If E
stands for English, R for Russian, F for French, V for vocabulary,
A for analysis and S for synthesis, the phases of the programme
E→R can be represented as follows:

If the double operation of English-Russian vocabulary and

analysis is summed up by the symbol VAER, it will be seen that,
once this stage is passed, the Russian equivalents of words are
available, ready for synthesis.
“It is evident”, writes Panov, “that when once VA is accom-
plished, we are fully provided with Russian equivalents, with all
their grammatical indications attached. Thus we can easily pass on
to, say, a Russian-French dictionary and get the corresponding
French equivalents together with their necessary grammatical
indications. Thus using only the French synthesis programme we
can obtain a French sentence automatically translated from
English.” [26]
It is, in effect, possible, by using a synthesis programme SF
and a Russian-French dictionary, to obtain the following formula,
in which the continuous lines represent the normal programme
and the dotted lines the proposed variations:
 FROM SOURCE LANGUAGE TO TARGET LANGUAGE 65
Russian is, of course, the interlanguage proposed in this solution,
or rather that kind of pre-Russian constituted by analysis of E→R
in which Russian grammatical categories are superposed on or
substituted for those of English. It is clear that Russian, being
a synthetic language and rich in grammatical categories, is thus
particularly well placed to be used as the basis for a machine-
interlanguage. Since in Russian the indications of grammatical
categories are attached to the words themselves instead of being
concealed in the intricacies of word order, Russian is better able
to respect the individual concrete character of the original sen-
tence than is a language less rich in inflexions and in which word
order, which is necessarily rigid, plays an important part in com-
municating the meaning of the sentence.
The method of multi-lingual translation proposed by Panov is
ingenious but not without its drawbacks. It is based on an analysis
of English, with Russian as the goal and as the tool. The figures
are recorded in the memory cells according to a system based upon
the Russian language. Since he who can do more can do less, it is
possible that this system may give satisfactory results for trans-
lation into languages less richly inflected than Russian. But this
method does not answer the objections raised by Booth. Is not a
programme E→F preferable, if only for reasons of economy,
to a programme E→R→F? Such a programme, while perhaps
of some value in Moscow, London or New York, would certainly
not be suitable, for instance, for a French organization engaged in
the production of scientific translations into French from a number
of foreign languages. For translation from English, Romance
languages, etc., such a team could utilize analysis programmes
much simpler than those designed by the Russians, though based
on the same principles. This simplicity would be due to the
greater degree of resemblance between the structures of the
languages involved.

PRIORITY OF BILATERAL PROGRAMMES

The Russian method makes it possible to have only one single
synthesis programme for any target language. Thus a national
workshop translating into the native language from a number of
foreign languages has everything to gain by drawing up its own
synthesis programme and by preparing for each source language
66 MACHINE TRANSLATION
an analysis programme in conformity with the demands of the
target language.
For theoretical as much as for practical reasons, bilateral pro-
grammes are at present the vital ones. The sum of knowledge
required for the eventual establishment of universal programmes
can be acquired only by comparison of numerous bilateral pro-
grammes, which should be worked out as the first priority in all
countries interested in machine translation for practical purposes.
The execution of translation work at high speeds and without
waste of precious time will depend on the economy of the pro-
grammes, that is to say on the degree to which they can avoid all
check-up routines having no immediate utility for the translation
in process. Even if it takes but a thousandth or even a millionth
of a second to fill a sub-division of a memory cell, economies of a
thousandth or a millionth of a second will be of great value in
programme making, as they will occur an infinitely great number
of times in each translation.
One of the great merits of the Russian system is that it makes
possible the use of whole sections of bilateral programmes for
work on different languages. With progress in self-programming
it is possible to look forward to a day when the machine will itself
choose from among several programmes those applicable to a
given language, its choice being governed by considerations of
maximum economy. At this point Andreev’s metalanguage would
tend to become one and the same thing, at least in this context, as
a recorded set of instructions enabling a machine to choose, from
among a universal but empirically elaborated programme of
translation, those elements required for the translation of language
A into language B and vice versa, and to perform this translation
with the greatest possible measure of economy. But much time-
consuming practical work is essential before this point is reached.
 CHAPTER V

Syntax and Morphology

DURING the early days of research the priority of lexis over
morphology in preparing the way for machine translation was taken
for granted. In drawing up the first automatic dictionary, Booth
and Richens gave scarcely a thought to grammar. Its true im-
portance became evident only as partial, word-for-word trans-
lation was abandoned little by little in favour of an attempt to
produce genuine, fully automatic translation. In the meantime the
work of preparing programmes for the machine clearly revealed
certain weaknesses in traditional grammars and opened up new
horizons as to possible fresh classifications of linguistic data.
Reifler’s work in Seattle, that of the Wundheilers at the Illinois
Institute of Technology, of Bar Hillel, Chomsky and Victor
Yngve, Oswald and Fletcher’s work on German grammar, ex-
panded later by Booth and Brandwood in London, Brandwood’s
further work on French—the accumulated experience and con-
clusions of all this research certainly contributed considerably
to the subsequent rapid progress made by Soviet linguists and
technicians. The role of morphology and of syntax was hence-
forward clearly defined. Automatic analysis of the function of the
word in the sentence, illustrated in the preceding chapter by
Soviet examples, was rendered possible by the previous research
of these pioneers.
IMPORTANCE AND LIMITS OF GRAMMATICAL PROBLEMS
As we have already seen, the somewhat naive optimism of the
early stages as to the usefulness of providing rough word-for-word
translations soon came up against the problem of inflected
endings, and later against the necessity of determining the gram-
matical values of uninflected words. It soon became clear that a
sequence of semes is insufficient to communicate the meaning of
a sentence, even to a specialist in the subject matter concerned.
68 MACHINE TRANSLATION
Translation was impossible without prior knowledge of the gram-
mar of the original language. One major reason for this was the
high proportion of polysemantic words.
Setting to work on this problem, Yngve [17] took as his starting
point the improvement of word-for-word translation and evolved
a series of basic principles. The information necessary for the
solution of the problem of multiple meaning, he observes, resides
in the context, that is within the sentence itself. Basing himself
on Zipf's work on word frequency, he noted that words very
frequently used are those that have the most meanings. The fifty
most frequently used words account for about one-half of the run-
ning words of a text, so that it is clear that a solution to the
the problem of multiple-meaning for the fifty most frequently
occurring words would go more than half-way towards solving all
problems of multiple meaning. These common polysemantic words
proved to be grammatical tools—or “cement words”—articles, pre-
positions, conjunctions, auxiliary verbs, pronouns, etc.—the very
words which constitute the grammatical structure of language in
which the nouns, verbs, adjectives, adverbs, etc., are contained.
Each sentence being almost separate, both grammatically and
syntactically, from its neighbours, a sentence would probably
be a suitable basis for translation and it would rarely be
necessary to go further afield to find a solution for the problems
of multiple meaning involved in the fifty most frequently used
words.
To lend support to arguments based on Zipf’s law, Yngve
produced evidence likely to convince unbelievers by recounting in
detail the results of an experiment he had himself conducted. A
partial translation of German was prepared, taking as text 750
running words of a review of an American work on mathematics.
The 250-word vocabulary of this text was put on cards and
Yngve, without prior knowledge of the text, translated each card
in alphabetical order. The mathematical vocabulary was correctly
translated but words like der and sein proved to be untranslatable
out of their context. They were therefore left in the original
German, and German word-order and flexional endings were
also left unaltered. The “translation” was typed out, the English
words in capitals and the remaining German elements in lower
case letters. Here is part of the result:
 SYNTAX AND MORPHOLOGY 69

“Die CONVINCINGe CRITIQUE des CLASSICALen

IDEA-OF-PROBABILITY IS eine der REMARKABLEen
WORKS des AUTHORs. Er HAS BOTHen LAWe der
GREATen NUMBERen ein DOUBLEes TO SHOWen:
(1) wie sie IN seinem SYSTEM TO INTERPRETen ARE;
(2) THAT sie THROUGH THISe INTERPRETATION
NOT den CHARACTER von NOT-TRIVIALen DEMONS-
TRABLE PROPOSITIONen LOSEen. CORRESPONDS der
EMPLOYEDen TROUBLE? I AM NOT SAFE, THAT es dem
AUTHOR SUCCEEDED IS, den FIRSTen POINT so IN
CLEARNESS TO SETen, THAT ALSO der UNEDUCATED
READER WITH dem DESIRABLEen DEGREE OF EXACT-
NESS INFORMS wird . . .”
The full text of this partial translation was then put before
two categories of reader: those who knew no German, who were
able to understand the subject matter only, without grasping the
meaning, and those who knew some German grammar and who,
once they had recovered from their amusement, demonstrated
that they had understood rapidly and well.
The next principle enunciated by Yngve was therefore: “Con-
centrate on the grammatical problems since they account for the
majority of multiple meaning problems, and the specialized field
glossary can cope with most of the rest.”
This experiment also demonstrated another point—namely, the
importance of word order. Since the problems of multiple meaning
are bound up with problems of syntax, the context of language B
can help us to understand the meaning of words in this language
only if the word order of language B is respected in translation
from language A into language B. Morphology and syntax can
generally provide the solution to those problems of multiple
meaning which occur most frequently.
This brief account illustrates two facts with which all subsequent
work on automatic translation has had to reckon. Morphological
and syntactic problems are paramount and no genuine translation
is possible until those problems have been solved. They are, how-
ever, relatively restricted in scope, and once the solutions have been
found for a given pair of languages A→B, they will be applicable
to all translations of A into B.
An experiment recounted by Bel’skaja [5] strongly confirms the
70 MACHINE TRANSLATION
fact that morphological and syntactical problems are at once of
primary significance and strictly limited scope. In an important
article in Research she describes the first English-Russian diction-
ary used by the Academy of Sciences of the U.S.S.R. for its
experiments. This dictionary comprised about 5,000 words, almost
equally divided into English and Russian. Since it was designed
for work on Russian mathematical texts, the vocabulary was, of
course, highly specialized. “As to the grammar part of the transla-
tion programme”, writes Bel’skaja, “it has very little, if at all, been
affected by the fact that a very limited field, that of mathematics,
had been chosen for machine translation. Indeed, the grammatical
programme has proved to be universally applicable.”
In support of her contention, the Soviet linguist quotes some
remarkable experiments made in order to discover whether the
same grammatical programme could be applied to a text as far
removed from mathematics as, say, an article from The Times, or
a passage from Dickens. “The experiments have proved the
success of our ideas on the possibility of having a universal
grammatical programme for the machine translation of any two
languages; in the vocabulary field a series of specialized diction-
aries, covering different fields of human activities, are unavoid-
able.”
The only thing with which the machine can be reproached in
this experiment is that it was obliged to leave in English the words
for which no translation had been entered in its dictionary—but
in a sentence such as the following only one English word (in
italics) remained in the Russian text:
“It made a great impression on me, and I remembered it a long
time afterwards, as I shall have occasion to narrate, when the time
comes.”
“Eto proizvelo bol’šoje vpečatlenije na menja, i ja pomnil eto
potom, dolgoe vremja, kak ja budu imet’ slučaj rasskazat', kogda
pridet vremja." (David Copperfield, Chap. XVI.)
Here we have proof that grammatical analysis of the sentence
is valid for all sentences, provided they respect the grammar and
above all the syntax of the language in question. A grammatical
programme established for two languages A→B can therefore be
used for all translations from A to B. It must, however, include all
constructions normally employed in that language. Machine-
 SYNTAX AND MORPHOLOGY 71
translation linguists will have to pay more attention to morphology
in languages which are analytical and richly inflected, whereas in
synthetic, poorly inflected languages the problems of syntax
will be paramount.

MORPHOLOGY AND THE MACHINE

Figure 3 shows how the problem of exploring word forms has
been solved in a language as poor in inflexions as English. English
has only six flexional endings: -‘s or –s’, -s for the plural or 3rd
person singular of the present tense, the verbal-endings -ing and
-ed, -er for the comparative, -est for the superlative and -th for
ordinal numbers. The Russians added a false inflexion, essential
for the operation of their dictionary: -e as in to love, to clothe, etc.,
and in adjectives like true. These words figure in the dictionary
as lov-, cloth-, tru-, thus permitting the identification of lov-es,
lov-ed, lov-ing, tru-er, etc.
Similarly in the French translation programme drawn up by
Mel’čuk and Kulagina on the basis of mathematical works by Paul
Appel and Emile Borel, stems or bases* and flexional endings have
been entered separately in the dictionary, the determining criteria
being empiric in nature and not historical. The base of travail is,
for instance, entered as trava- because it appears in the two forms
trava-il and trava-ux; parler is given as parl-, finir as fini-, and so
on. The common verbal endings have been grouped and classified
as in Brandwood’s work on French [7] in ending tables which
make it possible to analyse the word grammatically as soon as the
base has been identified.
Thus both for source and for target languages, for each variable
part of speech—nouns, verbs, adjectives, pronouns, ordinal
numbers, etc.—the machine will have an ending table making it
possible both to separate the bases after input and to identify
them, and to conjugate and decline the words at output by adding
the endings to the bases. For the separation of endings, the
machine will always follow the plan laid out in Figure 3—the
only difference being that Russian or German will require more
sub-routines than English, in which the maximum number of
attempts was seven until Bel’skaja added various refinements
* “Base” is a more correct name than “stem” for the graphically invariant
part of an inflected word.
74 MACHINE TRANSLATION
making it possible to trace the bases of Latin words currently used
in English, as well as plurals like busmen, carmen, etc. For the
synthesis, when endings are added to a base without its being
modified, as in aimer-ai, -ons, etc., a single base will be entered in
the dictionary. When the base is modified by certain endings, it is
preferable to enter both forms in the dictionary; alternatively,
when a base is regularly modified (as in the case of Russian words
ending in sibilants or gutturals) the machine can be instructed to
alter certain consonants before certain endings. Here again the
only criterion is programme economy.
In all poorly inflected languages—and even in those with a
number of inflexions—the endings do not always provide informa-
tion on the grammatical role of the word in the sentence. As
Mel’čuk and Kulagina [18] have observed, French contains forms
which can be either verb or noun: la forme, il forme, le fait, il fait,
la limite, je limite, etc. The ending fails to provide the required
clue, just as it fails to distinguish between il limite and je limite.
Here we must turn for a solution to the form classes established
by the structural linguists Bloomfield, Harris and Fries. The
question is to determine objectively to which form class the word
limite belongs in any given sentence. It is either a singular noun,
or a 3rd or 1st person singular verb (this the Russians call 4th
person until a sub-routine has been able to determine whether it
is 3rd or 1st).
When a word, such as limite, can be either verb or substantive
(VS) the following possibilities exist:

(A) VS preceded by a determinant (article, demonstrative or

possessive adjective) other than le, la, en, is a substantive.
(B) VS preceded by le, la, en:
(1) if the construction is: Nominative or accusative noun or
nominative pronoun, + le, la, en, +VS,
(a) in the absence of a comma or a co-ordinating conjunction
before le, la, en, VS is a verb.
(b) if there is a comma or a co-ordinating conjunction in
this construction,
(i) VS is a verb if there is no verb between it and the end
of the sentence or between it and a conjunction,
(ii) VS is a substantive if it is followed by a verb.
 SYNTAX AND MORPHOLOGY 75
(2) if VS preceded by le, la, en is not part of a construction of
this type then it is a substantive.
(C) In all other cases VS is a verb.
This is only a preliminary analysis, making no claim to com-
pleteness or finality, but illustrating a method inspired by struc-
tural linguistics. While Reifler divides German words into “form
classes” and records them in separate memories, the Russians of
the Steklov Mathematical Institute enter in the dictionary a
distinctive numerical indication for each class, this number being
in fact an instruction code which refers the machine to the appro-
priate sub-routine.

STRUCTURAL ANALYSIS
By analysing typical structures of whole sentences, or parts of
sentences, it can be made possible for the machine to translate
uninflected or partially inflected words. Syntax takes over when
morphology offers no solution. The machine then analyses the
positions of words in relation to one another. Here are some
examples from Mel’čuk and Kulagina.
“Pas, point, are negative particles if they come immediately
after the verb, or are separated from it only by an adverb, and in
constructions of the type ne+pas+infinitive. In all other cases
these words are substantives.
“Ensemble after a determinant or a preposition (from which it
may be separated by adjectives, adverbs and co-ordinating con-
junctions) is a substantive; otherwise it is an adverb.”
The formal rule is here incomplete. Ensemble is not an adverb
in “Ensemble de premier ordre, les Petits Chanteurs à la Croix de
Bois ont. . . .” The rule requires completion for cases where
ensemble, not preceded by a determinant, is followed by an
adjective, or else a general rule on appositions must modify this
rule. Nevertheless, we have here excellent examples of rules
which can be embodied in dichotomic sub-routines.
A more complete example is that of the English noun (and
pronoun) analysis, as practised in the Panov-Bel’skaja translation
programme. As in the sub-routine quoted in the previous chapter,
1.(2, 7) means: “Perform operation 1. If the reply is affirmative,
proceed to 2. If it is negative, proceed to 7.”
76 MACHINE TRANSLATION

English Nouns
1.(2, 7) Check given word for us.
2.(3, 5) Check following word for noun.
3.(0, 0) Produce sign of dative case.
5.(6, 13) Check immediately preceding word for let.
6.(0, 0) Produce sign of nominative case.
7.(8, 13) Check given word for it.
8.(13, 10) Check it for presence of sign of gender.
10.(0, 0) Take gender from nearest preceding subject.
13.(14, 15) Check for presence of sign of singular or plural
number.
14.(0, 21) Check for presence of any sign of case.
15.(16, 19) Check for ending -s.
16.(17, 17) Produce sign of plural number.
17.(18, 14) Check preceding word for formula without the
sign=.
18.(0, 0) Produce sign of genitive case.
19.(16, 20) Check preceding word for much.
20.(14, 14) Produce sign of singular number.
21.(22, 23) Check preceding word for let.
22.(0, 0) Produce sign of nominative case and subject.
23.(24, 28) Check immediately preceding word for sign of
similar conjunction.
24.(28, 25) Check word immediately preceding and following
similar conjunction for adjective.
25.(26, 27) Check all words for same word as the given word.
26.(0, 0) Take case from noun found.
27.(0, 0) Take case from nearest preceding noun.
28.(18, 29) Check for ending -s.
We see here how, in order to determine the case of the noun or
pronoun, the machine performs a series of explorations of immedi-
ate context of the word in question and of the structure of the
sentence. The sub-routine is, in fact, based on Yngve’s recom-
mendation: “to make the needed information that is implicit
in the context explicit at each word position in the sentence.”

CLASSIFICATION AND COMPARISON OF STRUCTURES

In richly inflected languages where word order is relatively free,
 SYNTAX AND MORPHOLOGY 77
morphological endings provide the information needed; in
languages like English or Chinese where the sequence of words is
strictly ordered, analysis must depend mainly on structure and
word order. Thus programmes or sub-routines appropriate to the
types of the two languages with which we are dealing will have
to be prepared for all the basic structures of each and for the
conversion of the structures of language A into those of
language B.
For the preparation of such programmes, the work of Jespersen,
Bloomfield and Fries has proved most useful. From our point of
view, the great merit of their researches, undertaken long before
there was any question of machine translation, was that they
examined grammatical structure without reference to meaning,
and sought to define structures independently of meaning,
because, as Fries pointed out, meaning provides no means of
identifying and distinguishing structures. This applies also to the
machine which, when faced with the sentence “the maid gave the
cat meat” is incapable of identifying the grammatical role of cat
or meat except by reference to structure, whereas any schoolboy
learning English will guess their respective roles by the meaning
of the sentence.
The work of Fries, which was largely directed towards the
teaching of English to foreigners, leads even more surely than
Jespersen’s Analytic Syntax to the expression of structures in
algebraic formulae.
English words have been grouped into four classes defined
according to their role in the structure of the sentence. Although
these form classes do not exactly coincide with the four main
parts of speech, it can be said, for the sake of brevity, that Class (1)
consists of most nouns, (2) of the majority of verbs, (3) of most
adjectives and (4) of almost all adverbs. Fries has also identified
fifteen groups of auxiliary or functional words designated by letters
A to O, and comprising in all 154 words. Complete sentences
can be expressed algebraically by means of these nineteen
symbols, to which must be added + for the plural, — for the
singular, -d for the preterite, -ing for the present participle, etc.
The sequence of symbols D 1 2 –d 4, for example represents

sentences of the type: The pupils ran out. The ships sailed away.
78 MACHINE TRANSLATION
Soviet linguists have made a thorough study of this system and
have classified Russian words into seventeen groups which they
have related to the English language classes and groups. Mološnaja
[19] has shown that structure-formulae can be grouped according
to types or models.
Elementary units of English words, associated in structural
patterns, having been classified, these structures were translated
into Russian as simply and directly as possible. The Russian
constructions were then analysed, reduced to formulae and com-
pared with the equivalent English constructions, whenever
possible by two-member word combinations. When two English
elements failed to coincide with two Russian elements, additional
symbols were introduced as required. For instance in the following
English absolute participle construction, the symbol J (subordinat-
ing conjunction) has been added to express the difference in
structure:
The rain having ruined my hat, I had to buy a new one.
Tak kak dožd' isportil moju šljapu, ja dolžen byl kupit' novoju.

English Russian
1 2-ing 12 J1212
These are cases where an English construction may have
several Russian equivalents:
(He) looks pale: 2±3 (On) vygljadit blednym: 2±3
(the stone) lay deep (in
the water): 2±3 (kamen') ležal gluboko
(v vode): 2±4x
a group of children: 1 F 1 gruppa detej: 1 1
a book with pictures: 1 F 1 kniga s kartinkami: 1 F 1
Structures being treated as entities like words, one can thus
make a dictionary of structures and identify cases of polysemantic
or homonymic structures requiring the formulation of special
rules to identify their target language equivalents, these special
rules being treated as sub-routines just like those for resolving
grammatical homonymy.
Complex structures can be simplified by reduction. For instance
 SYNTAX AND MORPHOLOGY 79

The old man 3 1

becomes D since the adjective plus the noun
D 3 1
1
are reducible to a nominal unit; this is very useful since in
Russian these three words can be translated by a single Class 1
word, and in French by two (determinant and noun).
Thus it is easy to see how, once all possible English syntagmas
have been classified and inventoried and the same processes
completed for a second language, for example Russian, it will be
possible automatically to identify each structure in a complex
sentence, to reduce it to a simple formula, which is then converted
into a language B formula and later expanded according to the
rules peculiar to that language, After reduction, conversion and
expansion of the formulae, it will remain only to “unroll” the
Russian sentence by means of the semes stored in the machine’s
memory and the morphological indications which were entered
in the memory when the sentence was analysed.
This method of syntactic analysis requires further intricate
research and much elaboration of detail. It has been advanced by
Mološnaja as a working hypothesis which the machine must verify
by application to a large corpus of text. It would appear to be
capable of providing a solution to the problems of automatic
analysis by the machine of the syntactic elements constituting
complex sentences.
It might make it possible to complete sentence analysis by
separating temporarily, where necessary, the various constituent
elements of expression: semes, which are identified by the diction-
ary; morphemes, identified either by form classes (if they are
independent words) or by the stem-ending tables (if they are
inflexions); syntagmas (expressed in formulae) identifiable by
means of the inventory of the various possible structures of a given
language.
It is clear that such a method cannot do more than facilitate the
translation of word combinations of unambiguous structure. A
phrase like “the King of England’s Empire” will always remain
enigmatic to the machine, since it contains no graphic or structural
clue to determine whether it refers to the Empire of the King of
England, or to the King of the Empire of England. In such cases
a reviser must remain the only final resort, as in cases of lexical or
80 MACHINE TRANSLATION
morphological ambiguity not reducible to objective graphic
criteria.

STRUCTURAL MEMORIES
Meanwhile, it is clear that the translation machine will have to add
to the memories already described—lexical memory or dictionary,
morphological memory or stem-ending tables,—a structural
memory, permitting comparison of structures received at input
with structures held in this structural memory. A comparison of
this sort, while not of any great interest for simple sentences, will
be essential for complex sentences and will make it possible to
reduce to a minimum all cases of ambiguity inherent in the
structure of the sentence itself. In human languages, it is syntax
—the compulsory pattern for the combination of words into
sentences—that is slowest to change. A complete inventory of the
structures of a given language will require ingenuity in classifica-
tion and finesse in establishing the order in which the machine is
to conduct its explorations, rather than patience in drawing up
the inventory itself. Such an inventory can be drawn up for each
language and only when this has been done will fully automatic
translation become possible. Compared with lexis, the problem is
one which is relatively limited in scope.
For the same reasons, the problem of registering the rules of
morphology and syntax in the memory of the machine does not,
from the point of view of computer technique, present any pro-
blem more difficult to solve than those involved in programmes of
management and scientific calculation already treated by machines.
The data to be entered are not appreciably more numerous than
the rules which must be stored for the execution of a sequential
series of scientific calculations. A magnetic drum computer will
probably be able to enter on its drum all the morphology and all
the syntax of two languages—the only remaining questions in the
field of technology being rapidity of access to the rules thus
entered, and in the field of programming, the order of entry of
these rules and of access to them.
 CHAPTER VI

Lexical Problems of Automatic

Translation
WHILE the problems of morphology and syntax are relatively
restricted in number, the same cannot be said of questions raised
by vocabulary. These are, indeed, very considerable in extent if
not in complexity. Taking into account variations in the meaning
of words, the rapid evolution of scientific and technical vocabulary,
slang and local speech, the number of words per language may be
so high as to challenge the skill of electronic memory constructors.
In recent concise dictionaries the vocabulary of the English
language comprises some 60,000 word entries: this number may
run four times as high if each meaning of each polysemantic
word is entered separately. So that a dictionary in which every
form of every word would constitute a separate entry might well
number over half a million words in a modern inflected language.
We are thus faced immediately with the problem of lexical
content. This is closely followed by questions of classification
(should there be one dictionary only, or several, according to
subject?), and of order of classification (alphabetical, logical, con-
ceptual, or numerical according to the increasing or decreasing
number of characters in a word, etc.?). Finally come the specific
problems of translation—multiple meaning, idiom; and—sooner
or later—the problem of style, or styles, of the choice of words
for reasons peculiar to the author.
MEMORIES: TECHNICAL ALTERNATIVES

A solution to any one of these problems involves a choice, or a

series of choices, inevitably limiting possibilities in other direc-
tions ; all the more so, in that lexical problems, even more than those
of morphology, are closely bound up with the technological aspects
of computer construction. Memory capacity, rapidity of access,
these are important considerations in the choice of solutions. Even
82 MACHINE TRANSLATION
if we set aside for the moment, for practical reasons, the objections
of those who maintain that the choice of the right word by the
translator is a matter of taste, of personal judgment, and that the
machine will never be able to exercise such judgment, we can still
not affirm at the present time that an ideal solution has so far been
found to the lexical problems of mechanical translation. But the
empirical method of partial solutions has been applied with
increasing success. It has enabled research to continue while
technicians pursue the study of recording processes, thanks to
which it will eventually be possible, where required, both to store
a very great number of words and to have very rapid access to
them.
One or more magnetic bands or a battery of magnetic discs can
contain an entire dictionary: but access time to any given word is
relatively long, and this means a slowing-up in the matching of
input words with words stored in the dictionary. It has been
calculated that the lexical memory of the machine should provide
random access to any word in not more than 10 milliseconds; this
would allow for the matching of dictionary words with all the
words of a sentence of average length (20 words) in one-fifth of a
second. Since tapes and discs can only give sequential access, it is
the magnetic drum that at present appears the most suitable
method of making vocabulary as well as rules of syntax and
morphology instantly available for machine operations. The
weakness of the drum is that its capacity is limited. Other types of
memory with high capacity and rapid random access are, however,
likely to be available shortly.
During the past two years Erwin Reifler and Lew Micklesen at
Seattle have been concentrating on the linguistic work connected
with the use of a large-scale Russian dictionary recorded on
Gilbert King’s photoscopic memory, which provides immense
capacity and very rapid access, but the logical circuits necessary
for translation have not yet been added to this memory. Similarly,
at Harvard, Oettinger and others have been working on a Russian
automatic dictionary on a Univac I computer. Both teams have
made considerable progress in the lexical analysis of Russian and
the logical treatment of vocabulary, and their methods, when
applied to more modern computers, should lead to very rapid
progress.
 LEXICAL PROBLEMS OF AUTOMATIC TRANSLATION 83
The work of Reifler and Oettinger clearly shows that there could
be no excuse for awaiting the improvement of memories before
beginning the basic linguistic work. All that is necessary is to
ensure that lexical research is undertaken in the order best calcu-
lated to exploit the properties of existing machines while not
losing sight of future potentialities. We may then hope that many
of the purely linguistic aspects of the organization of lexical work
will have been dealt with by the time the electronic technicians
come forward with their optimum solutions.
Let us suppose that a large magnetic drum is used as support
for a lexical memory. If the average length of a word is six letters
(to which must be added grammatical indications and programme
instructions appropriate to each word) we must allow for six
characters  six bits per character + the indicators and instruc-
tions, that is in all some 250 bits per entry. Certain programmes
may require as many as 1,000 bits per dictionary word. The
magnetic drums of the type employed in the 704 computer now
functioning in the I.B.M. Paris office store between them 294,912
bits—that is, according to the type of programme, between 300
and 1,200 words; the drum of the Gamma 60, containing 786,432
bits, has a capacity of 785 to 3,200 words. To execute a minimum
programme a translating machine must therefore have a capacity
at least as high as that of the I.B.M. 704 and may require twice
this capacity unless it is arranged for the machine to have, for
certain parts of the programme, rapid access to other types of
memory which are called into play in certain cases only.
Several different types of lexical memory can be used. It is
conceivable that one or more vocabularies comprising a very large
number of words could be recorded on magnetic tapes, of un-
limited capacity but sequential, and therefore slow, in access; that,
for the special requirements of any given translation, one such
vocabulary (or section of a very large vocabulary) could be trans-
ferred, for the duration of the operation only, on to either a drum
or ferrites or any other form of memory providing rapid or ultra-
rapid random access. In the course of one translation, it would
then be possible, on receipt of a given signal, to call into play such
and such a specialized vocabulary, registered on magnetic tape, to
transfer it for a few minutes only on to a drum or ferrites, and
to replace it some instants later by another similar vocabulary.
86 MACHINE TRANSLATION
The time of transfer being relatively negligible, a rational organiza-
tion of vocabularies by subject is perfectly compatible with the
simultaneous utilization of slow sequential memories and rapid
random ones (see Figure 4).

CONSULTING THE ELECTRONIC DICTIONARY

The first question to be raised was how to classify words in an
electronic dictionary in such a way as to ensure as rapid a look-up
as possible. Words being represented by a binary numerical code,
several alternative methods of classification have been tried:
arrangement in order of decreasing frequency, alphabetically by
sections, etc. Booth [7] has described the method recom-
mended both by himself and by the Russians. “Suppose,” he
writes, “that the dictionary contains N entries arranged in
ascending order of numerical magnitude in locations 1, 2 . . . , N
and that N is some power of two. The incoming word is first
subtracted from the entry in N/2. Then if the result is positive,
the required entry is in the ‘first half’, i.e. between 1 and N/2.
If negative, however, it is between N/2 and N. Now, assuming a
negative result at the first stage, subtract the word from the
middle entry of the last half, i.e. that in N/2+N/4. If the result
is negative the equivalent must lie between entries N/2 and
N/2+N/4 and, if positive, between 3N/4 and N. This comparison
process is repeated until the correct location is isolated and it is
seen that this requires log2N steps.” With a dictionary of 10,000
words (104), 14 operations (4log210) are required; for 20,000
words (2104) 15 operations, and for 1 million words (106) 20
operations. For a machine of which the combined access and
subtraction time is 1 millisecond, the look-up time for one word
in a million is about 20 milliseconds.

CODE COMPRESSION
A further technical refinement of great importance for the whole
conception of the dictionary is code compression. In order to
save memory space and thereby augment memory capacity,
methods employed in cryptography and telegraphy have been
adopted by mechanical translation mathematicians. One such
method consists in adding together the code numbers of each six-
 LEXICAL PROBLEMS OF AUTOMATIC TRANSLATION 87
letter group in the input word and treating the resulting total as
representing the word.

LINGUISTIC PROBLEMS
Classification of words in ascending order of magnitude of their
code and subsequent code compression are mathematical solutions
to technological problems of recording linguistic data in the
machine or of achieving greater speed of access. The real linguistic
problems are no less urgent—for example the fundamental
question of multiple meaning in relation to the dictionary. Should
the dictionary contain as many entries for each word as that word
has meanings? Would this drastic solution, which is perhaps com-
patible with a gigantic memory having very rapid access, be
appropriate for a programme of sentence analysis based, as at
present seems desirable, on the necessity for solving problems of
multiple meaning by exploration of context? Or have we not
rather arrived at the point where the complexity of programmes
and the necessity of keeping such programmes flexible, argue in
favour of restrictive dictionary size and concentrating on speciali-
zation by subjects or groups of subjects?
In all experiments to date certain precise limits have been
imposed in order to achieve effective results without sacrificing
the balance which it is desirable to maintain between the pro-
portions of the machine and the relative size of the subject of the
experiment.

A FRENCH-RUSSIAN DICTIONARY
Kulagina and Mel’čuk have constituted, according to this prin-
ciple, an experimental electronic French-Russian dictionary for
the translation of mathematical texts. Their method differs little
basically from English and American electronic dictionaries. The
texts of Paul Appel and Emile Borel on which the dictionary is
based comprised 20,500 running words of which 2,300 were
different. The 1,000 words occurring more than four times each
were entered in the dictionary. Without any statistical survey,
about 50 words that “were obviously needed” were added,
together with another 50 French “grammatical tool” words. This
gave a total dictionary of about 1,100 words. Each stem in the
dictionary was accompanied by a dictionary entry containing:
88 MACHINE TRANSLATION
(1) the Russian translation; (2) French data including (a) a part-
of-speech notation, (b) an idiom notation, (c) the preposition code,
(d) grammatical characteristics; (3) Russian data, including (a) a
notation on selection of Russian stem, (b) grammatical character-
istics; (4) a notation on the choice between two French stems. [18]
This method of noting the characteristics of each word is
similar to that described by Panov and Bel’skaja for English
vocabulary (see Chapter IV); in the case of French, the gram-
matical indications are, for example, for nouns: gender, formation
of plural; for verbs: transitive or intransitive, conjugated with
être or avoir, conjugation number, etc.
The preposition code corresponds to the peculiar problems
presented by French prepositions which can be translated in
many different ways, governing a number of different cases.
This code refers the machine to special preposition translation
tables. Preposition codes are given for nearly all verbs and many
adjectives and nouns, the same preposition code number being
given to all words governing the same preposition.
The notation on the “choice between two stems” consists of an
indication that a choice of stems is involved and a notation giving
the address of the alternative stem: for example point, noun and
point, negative particle, will be accompanied by this indication so
that the machine, having completed the look-up and the pro-
cessing of idioms, can, by using the rules for distinguishing
homographs, “decide” which of the two stem entries applies in
that particular case.

IDIOMS AND HOMOGRAPHS

The indication “idiom” means that a word may, in association
with others, form a group, the meaning of which is not dependent
on the analysis of each word in the group, so that a literal trans-
lation would either be meaningless or would convey a wrong
meaning. This indication refers the machine to a special dictionary,
where all idioms containing the word bearing that indication are
listed. They are divided into integral (e.g. de plus en plus, à
présent) and non-integral idioms (e.g. aussi . . . que, à . . . près)
arranged in the alphabetical order of the meaningful word of each
idiom. Under the same meaningful key word, which may occur
in several idioms, they are arranged in decreasing order of the
LEXICAL PROBLEMS OF AUTOMATIC TRANSLATION 89
number of words in each idiom. A special indication in the stem
dictionary gives the number of idioms listed under each key word,
so that the search programme may come to an end when the list
is exhausted.
Thus when the machine finds the word plus, the stem dictionary
refers it to the idiom dictionary. If the machine identifies a group
of input words (for instance de plus en plus) with one of the idioms
listed, it finds the translation; if not, it returns to the word plus
and translates it in accordance with the instructions of the stem
dictionary.
A distinction is thus established between the analytical con-
stituents of language—which the speaker is still free to combine as
he wishes—and fossil or vestigial constituents, which, while they
are not single words, are nevertheless units of meaning which can
no longer be analysed. This distinction shows up the true nature
of idiom in our modern languages: as fossilized survivals of
expressions which were originally analytical.
Idioms do not, in scientific language at least, present a problem
of any great magnitude. It is necessary only to catalogue them and
record them in a special memory. The problem will doubtless be
very different when we come to everyday language and to that of
plays and novels. Since the use of idiom introduces an extra-
linguistic element into language—the evocation of a situation
which has a special meaning for a given social group, a systematic
study will have to be made of idioms, clichés and all metaphorical
use of words or groups of words—and Flaubert’s original idea
of a Dictionnaire des Idées Reçues will perhaps enjoy new popu-
larity and expansion! A study of this type would make it possible
to decide which idioms form part of everyday language, and must
therefore figure in the idiom dictionary; which can be translated
literally into certain languages, and which must perforce be left
in the original language as being totally untranslatable since they
refer to social situations of uniquely local and limited significance.
The problem of homographs is relatively restricted. We have
seen in the previous chapter how grammatical analysis of context
enables the machine to choose the right translation for most
homographs. In the rare cases where such analysis is insufficient,
they will probably have to be classified with genuinely poly-
semantic words.
90 MACHINE TRANSLATION

GENUINE POLYSEMY
There are many words, apart from those with idiomatic usage and
those whose meaning varies with their grammatical function,
which are in fact truly polysemantic. Is the English plant a French
plante or usine? Is the French temps to be translated time or
weather? Should champignon be rendered by fungus, by mushroom
or by toadstool? Grammatical analysis is of no assistance, nor
at first sight is the idiom dictionary.
What does the translator do when faced with such a problem?
If he understands the subject perfectly he chooses the translation
which appears to him to correspond to the overall sense of the
context. In a sentence dealing with poisoning, he will translate
champignons by toadstools—although he will be understood if he
says fungi. But scientific and technical translations are full of
traps for the human translator not fully conversant with the subject
of his text. Only constant and close collaboration between trans-
lator and specialist will ensure that the right translation is always
given—above all in texts on modern technical subjects where the
vocabulary is in constant evolution.
The translation machine cannot hope to do better than the
human translator in this respect; if the text fails to provide the
machine with recognizable, objective criteria signalizing meaning,
then the translation will, for the present, have to list all the
meanings of indistinguishably polysemantic words, and a specialist
will have to choose the right meaning from this list before the
final version is made. It is, however, obviously desirable that the
machine should be able to solve the majority of polysemantic
problems. It should be able to choose the right meaning. In cases
of grammatical multiple meaning and of homographs, we have
seen that the micro-context—the study of the immediately
surrounding words—has made it possible to choose automatically
between several meanings.
How can the context help to determine the correct English
rendering for a polysemantic word like champignon or to decide
on the correct English equivalent for temps? It is possible to
imagine a general dictionary for a given language A containing a
translation for every single word in language B. A word would be
defined as a meaningful group of signs, alphabetic and non-
 LEXICAL PROBLEMS OF AUTOMATIC TRANSLATION 91
alphabetic (spaces, hyphens, etc.) which have one meaning only.
A word M with four distinct meanings in language A and requiring
four different words for translation into language B, would thus
figure four times in the dictionary, as M1, M2, M3, M4, with
appropriate indications making it possible to identify with one
of these four meanings, according to the context, its correct
translation. But for this it is essential that the sentences to be
translated should contain objective criteria enabling the machine
to choose between M1, M2, M3, and M4, and only extensive
analyses of context will make it possible to see to what extent
they so do.
Polysemantic words which have exact polysemantic equivalents
in another language are of no great importance: the overall sense
of the context will provide the reader with the means of choosing
between the four meanings of a given English word provided these
four meanings coincide exactly with the four meanings of a given
French word. Problems arise where the multiple meanings do
not coincide between two given languages, that is to say where
there are differences between the connotations of a word in lan-
guage A and those of the word which normally translates it in
language B.

MICROGLOSSARIES
Of all the solutions so far suggested, the most practical seems to
be the idea of idioglossaries or microglossaries. It had originally
been thought that by listing in the output translation all possible
meanings of a word in language A, the reader could select the
correct one according to context. Research since 1949 has led to
the provisional conclusion that in scientific texts non-grammatical
polysemantic nouns and verbs do not present any great difficulty
within the limits of the restricted vocabulary of any given science
or technical subject. Thus special restricted dictionaries—
microglossaries—should be constituted, having the double
advantage of reducing the size of the dictionary necessary for a
given translation to dimensions compatible with the operational
memory of present-day computers, and also of limiting the
number of cases of non-grammatical polysemantic words.
The Academy of Sciences of the U.S.S.R. considers a dictionary
of 6,000 words quite sufficient for translating any mathematical
92 MACHINE TRANSLATION
text. They think it reasonable to expect that other fields will not
require much larger vocabularies. This estimate is borne out by
the statistics showing that 95% of English texts can be understood
by a reader knowing 6,000 words. The specialized mathematical
dictionary established at the Academy for the translation of
Milne's The Numerical Solution of Differential Equations was
divided into three independent sections:
(1) Technical words, i.e. mathematical terminology—approx-
imately 400 words.
(2) Non-technical mono-semantic words, amounting to 1,800
words.
(3) Polysemantic words, amounting to 300.
The technical words of a subject being thus recorded in a
special memory, it becomes relatively easy to find their exact
translation for this subject, it being assumed that multiple meaning
is rare within the limits of one scientific subject. The one remaining
problem is that of multiple meaning of words which have one
meaning in mathematics, for instance, and another in physics, in
a sentence dealing with both mathematics and physics. Here the
machine is at a disadvantage compared with the specialist trans-
lator, but not greatly so compared with the non-specialist.
Andreev suggests determining the particular meaning of a poly-
semantic word by a system of “semantic keys”, of the type
employed by lexicographers for identifying particular acceptances
of words. In an article on agriculture, for example, the word luk
in Russian has every likelihood of meaning onion and not bow; in
an article on astronomy, vozmuščenie will almost certainly mean
perturbation, a change in the orbit of one celestial body under the
influence of another, and not the mental state indignation. While
secondary meanings cannot be absolutely excluded in such cases,
their probable incidence according to Andreev [1] is close to zero.
“Hence the percentage of errors resulting from disregarding the
secondary meaning will in general not be greater than the usual
percentage of typographical errors.” When receiving the text
for translation the machine will be provided with a semantic key
permitting it to select immediately from a general dictionary the
particular meaning of a polysemantic term corresponding to the
subject of the text. This, of course, would be particularly suitable
 LEXICAL PROBLEMS OF AUTOMATIC TRANSLATION 93
in a large-size dictionary such as that on which Reifler has recently
been working in Seattle.
Andreev further recommends sub-dividing the dictionary into
separate fields: mathematics, chemistry, zoology, music, etc., each
with its own semantic key or numerical code establishing a relation
between a word and the subject of the appropriate section. The
translation will proceed by successive look-up operations in the
different sections of the dictionary, beginning with the main
subject, i.e. the mathematical section if the text relates to mathe-
matics, and so forth. The general dictionary will be consulted
only after the special sections. It will itself be sub-divided accord-
ing to indications provided by a statistical study of vocabulary,
into the following sections: (1) commonly used words, (2) words
of average frequency, and (3) rarely used words. Dictionary
search will proceed in the numerical order of these three categories,
and only words which are not found in the first category will be
looked for in the second and so on. An appreciable amount of
time will thus be saved in dictionary look-up.

STATISTICS OF WORD MEANINGS

All these various solutions to the problem of genuine polysemy
involve a pre-selection of the meaning of words by means of the
macrocontext, while use of the microcontext makes it possible
to solve problems of grammatical polysemy and to identify
idioms. Both approaches are founded on a probabilist attitude
towards the problem of meaning: it is therefore essential to make
thorough statistical studies, based on numerous and varied texts,
of the exact meaning of words in language A and their equivalents
in language B. Such a study should be focused not only on the
frequency of words as represented by alphabetical signs, as are
those of Zipf and Estoup, but also on the frequency of the various
meanings of polysemantic words. Macrolinguistic analysis will
be brought to bear on sign/meaning combinations and not on
signs alone.
Linguists who become automatic translation programmers will
have to be trained in probabilist methods. If a word has a certain
meaning in 95% of cases and alternative meanings in 2% and 3%
of cases, it may be necessary to risk translating it by the first
meaning and to give the other two only in parentheses, or to
94 MACHINE TRANSLATION
ignore them altogether. The bilingual dictionary of automatic
translation will be based on this principle. Such calculated
acceptance of risk is also necessary in organizing human transla-
tion, which is never altogether devoid of erroneous shades of
meaning: as in typography, it is the low percentage of error which
determines the quality of the work—total absence of such errors
is very rare. It is probable that by making systematic inventories
of vocabularies and synonyms it will be possible, for scientific
texts, to isolate a relatively restricted number of cases where only
specialists in the subject will be capable of determining the correct
translation in a given context.
Reifler, in one of his studies, rightly emphasizes the importance
of comparative semantic studies for the eventual reduction of the
role of the reviser of automatic translations. One aim of such
studies might be to determine which word in language B translates
most completely all the meanings of a word in language A. This
translation would then be adopted as the most satisfactory from the
point of view of the reader’s comprehension of the meaning of the
original text. For instance the English fungus could always be used
to translate champignon, as being the only word communicating the
total connotation of the French word. Thus automatic translation,
for the sake of communicating meaning, would seek a simplifica-
tion and concentration of vocabulary, similar to that observed
during all the great classical periods of history, when meanings
with wide connotative values in the given community are preferred
to individual and local semantic fantasies.
How far should such studies be carried? We can do no more
here than to suggest a few of the avenues open to us. The question
has been asked, how can the machine distinguish between the
meanings of the French temps, time and weather? If we turn to
Littré and see how many times temps suggests the passage of time
and how many times it refers to sun and rain, we find that the rare
examples quoted by Littré where temps means weather can all be
considered colloquial or idiomatic, the special meaning being
identifiable with the help of objective criteria present in the micro-
context. The study of idiom and that of the particular meanings of
certain words leads straight into the field of comparative and
statistical semantics, and it is here that the key to many poly-
semantic problems seems likely to be found.
 LEXICAL PROBLEMS OF AUTOMATIC TRANSLATION 95
We thus arrive at the need for a taxonomy of the meanings of
words, and for logical definition and classification of concepts and
the ideas which express them—to the synchronic study of changes
in meaning by metonymy and synecdoch.

THE THESAURUS METHOD

That is broadly the direction taken by the British team working
in the Cambridge Language Research Unit under the direction of
the logician Margaret Masterman. They have had the idea of
solving problems of meaning equivalence between languages by
a device based on the conception of Roget’s Thesaurus of English
Words and Phrases, in which words are classed according to the
ideas they express. Ideas are classified logically, generally in
dichotomic form, under numbered headings. Thus Roget had
established a concordance between a logical classification of con-
cepts and a numerical system which can be processed through a
computer. The same English word with several meanings will
appear under several headings, accompanied by other words
expressing the same or closely related concepts, and thus narrowing
down the possible meaning.
In order to reach the exact language B equivalent of a word in
language A, the Cambridge research unit have thought of improv-
ing on the Thesaurus, and of searching by machine for the word in
language B which is common to all the Thesaurus headings under
which the words in the immediate context of this word can be
found. The method is attractively ingenious but proves somewhat
disappointingly clumsy in application. Numerous systematic
trials alone can show whether it can be of real service. It seems,
however, open to a fundamental criticism: is it necessary or useful
to look for the correct word in the output language, by a method
fraught with hazards? Is it not simpler to look deliberately in the
electronic dictionary for the translation of a given word, having
used the context in the original language to define the meaning
of that word? However seductive and original some of the ideas of
the Cambridge unit, it seems paradoxical and contrary to the
necessary respect for the intentions of the writer, to place the
emphasis on the output language in lexical search for the right
word. The philosophy of thought and of its means of expression
which is behind the work of this group derives from the disciplines
96 MACHINE TRANSLATION
of logic rather than of linguistics and psychology, and is therefore
somewhat divorced from the empirical approach which the problem
of translation requires.
Nevertheless, this Cambridge research leads in the direction of
a new kind of bilingual or multilingual dictionaries, in which
words would be classified according to the ideas they suggest, with
a numbering system referring to their logical position within a
taxonomy of concepts. This would be a refined variant of
Andreev’s suggestion, based not on notions of simple frequency
but on the relationship between words and a logical classification of
ideas. The Thesaurus method, applied not to the operations of the
machine at the output stage—where it would run the risk of over-
looking certain fundamental requirements of translation—but to
the classification of words in the dictionary, would probably
facilitate the search for exact meaning in certain cases of polysemy
by bringing the immediate context to bear more completely on
the polysemantic word.

SCIENTIFIC AND TECHNICAL DICTIONARIES

Apart from the thousand or so common words indispensable to
any translation, the vocabulary of science and technology will be
the first to be subjected to terminological classification for mech-
anical translation. It constitutes a high priority field in view of the
urgent needs of science, and a particularly propitious one in that
words generally have only one meaning within the limits of the
glossary of a single scientific subject.
The lexicography of mechanical translation is now aiming at the
construction of specialized glossaries magnetically or otherwise
recorded. The work of Oettinger at Harvard, Panov and Bel’skaja
in Moscow, Reifler in Seattle, shows the way. But, as Oettinger's
latest report proves [29] the construction of a dictionary is in-
separable from work on morphology and syntax and from the
many aspects of linguistic research involved in and facilitated by
computer analysis.
Like the Harvard team, the research workers at the Rand
Corporation of Santa Monica in California have undertaken under
Kenneth Harper and David Hays a systematic inventory of words
and rules of translation—in spite of all attempts at analysis, the
two remain indissoluble—and have defined the guiding principles
 LEXICAL PROBLEMS OF AUTOMATIC TRANSLATION 97
of their work. The inventory begins with 20,000 to 50,000
running words taken from one or more scientific texts in one
language and dealing with one subject. These texts are systematic-
ally analysed by means of punched cards and electronic machines.
Morphological and syntactical rules with a bearing on translation
are inventoried; the vocabulary is classified into monosemantic
and polysemantic grammatical words, common words, and words
peculiar to the subject of the text. Note is taken of terms which,
having a different meaning in other contexts, are to be the subject
of separate study. A translation is then made and treated by the
same procedure of word and context analysis. Thus we have a first
vocabulary of words in common use in each of the two languages
under examination, plus a bilingual dictionary of the scientific
subject chosen. A second corpus of text is translated mechanically
by means of the vocabulary and rules drawn up from the first
translation; this new translation is revised and improved where
necessary, the decks of punched cards are completed in accordance
with those improvements wherever necessary and the whole
process is then repeated. The same operation, carried out as often
as necessary on other texts as closely as possible related in subject
to the first text, will make it possible to constitute step-by-step a
complete bilingual vocabulary of this scientific subject, the
common vocabulary being both enriched and defined in the pro-
cess. For this purpose a corpus of some 250,000 to 500,000 words
on the same subject would be subjected to analysis by this cyclic
process. Acceptable translations are a by-product of this method,
clearly illustrating its value as a means of accumulating objective
knowledge of language.
By moving on to another subject—proceeding from mathematics
to a branch of physics or astronomy, for example—the bilingual
scientific vocabulary is gradually extended, and a series of micro-
glossaries can be created in which the enrichment of the vocabulary
goes always side-by-side with the statistical analysis of results:
word frequency, incidence of exact language-to-language equiva-
lences, nature and frequency of polysemantic words, etc.
A NATIONAL TERMINOLOGICAL CENTRE AND TRANSLATION LABORA-
TORY
This method, inseparably bound up with ultra-rapid electronic
98 MACHINE TRANSLATION
recording processes, contains in embryo the solution to the basic
problems of scientific and technical translation, whether by man
or by machine. In the first place it will permit the constitution of
technical dictionaries which can be constantly kept up to date
and available for rapid and reliable consultation. We are, however,
dealing with collective means of production, the complexity and
costliness of which will make it necessary to operate on a nation-
wide or even international scale.
When once an inventory has been made of all the Russian words,
for instance, used for one technical or scientific subject and their
French equivalents, and it has been magnetically recorded in
binary code, a national electronic centre for terminology could
continue to receive from specialized research bureaux, in the form
of typewritten or better still punched cards, both requests for
technological information and new acquisitions in technological
terminology. The electronic dictionary for any given scientific
subject would thus be kept up to date at regular intervals by the
automatic insertion of new words or meanings on magnetic bands,
which would, in due course, be used to “charge” the magnetic
drums or ferrites of the rapid access memory of translation
machines. These magnetic tapes would serve a dual purpose: not
only would they constitute the permanent technological memory
of translating machines, but they could also be "read" in reply to
questions received from research centres, and this “reading”
would enable the reply to be punched or typed on the incoming
cards, which can then be returned to the bureaux from which
they came. A central automatic translation workshop would thus
necessarily be at the same time a national or regional focal point
for terminological information and would replace or complete
lexicographical cards of the conventional type (see Figure 4).

FROM METALANGUAGE TO THE UNTRANSLATABLE

In what order should this work of drawing up a terminological
inventory, of cataloguing the sum of human knowledge, be
undertaken? The facts dictate a certain order. The specialized
vocabularies should be explored in the order defined more than a
hundred years ago by Auguste Comte, that of the decreasing
exactitude of the sciences. The effort required will increase with the
decreasing degree of precision of each branch of knowledge. It is
 LEXICAL PROBLEMS OF AUTOMATIC TRANSLATION 99
no accident that has led the Soviets to begin with mathematics.
Astronomy also brings to automatic translation clear and distinct
concepts, an international terminology free of individual fan-
tasies, analogies and figures of speech, those traps set by non-
Cartesian thought on the path of all translation which seeks to be
exact and faithful. Next will come the natural sciences, in their
descriptive and mathematical aspects, the concepts of which are
often expressed in formulae or in exact definition of objects; their
terminology will be all the simpler to catalogue in bilingual form
in that their vocabulary describes substances, recognized facts
and relationships which can be perfectly expressed without any
admixture of metaphoric language.
With the human sciences, social and psychological, the problems
become more complicated and the lexical work in particular
becomes extremely arduous, since it must bestride two languages
and yet take into account the semantic variations introduced by
individuals in the use of words and in the creation of vocabulary,
inevitable when new ideas are to be expressed. These sciences are
already far removed from formulae and metalanguage, and are a
fertile field for image and figures of speech.
Petroleum research engineers, for whom a “carrot” is the
contents of the boring tool taking soundings from the subsoil,
atomic scientists for whom the atomic pile which generates
fissionable isotopes is a “breeder”, face the translator with prob-
lems as hard as those set by the statesman making a speech at the
United Nations who interlards his words with proverbs and images
drawn from his national folklore, without equivalent in any
foreign tongue. These men are thinking in terms of imagery, not
in exact definitions. Brilliant improvisations translating these
images will be much admired and long remembered: only the
untranslatable really requires translation. You do not need to
translate H2O and ax2+ bx + c. The more closely language reflects
the reality it expresses, the less it is necessary to translate from one
language into another. It is metaphorical neologisms, not the clear
and distinct ideas of the sciences, which cause disputes in termin-
ology centres and offices for the standardization of technical terms.
They too necessitate the most laborious research on the part of the
translators of international organizations. The art of the translator
begins at the point where thought diverges from the descriptive
100 MACHINE TRANSLATION
and analytical methods of the sciences, into the twilight zone
where judgment and feeling play as great a part as knowledge. In
this area successes are sometimes easy to score, but traps are many,
disputes often violent and personal preferences vehemently
expressed.
This is also the zone of genuine polysemy, which taxes equally
machine and translator, since the lack of objective criteria is bound
to leave open several alternative choices. No such criteria are
available to guide the translator faced with the metaphorical use
of words (like carrot or breeder) such as are constantly found in
technical texts, and in the writings of sociologists, economists,
psychologists, psycho-analysts and even linguists, all of whom
tend to confuse language the tool of their analysis, with language
the object of their study, because the subject of their work has no
material being other than in words.
In such cases the translating machine, in spite of its special
requirements, can be of some service. The author of obscure texts,
in which the slippery shifts of semantic meaning are never signal-
ized, is always inclined to blame the translator for not following
every fluctuation in the meaning of the words he uses. Will he be
able to blame the machine if he has used the same word in two
different senses? The machine will be able to translate writings on
the human sciences only after thorough preparatory work on their
terminology. In this respect it will not differ from its human
predecessor. Such preparation will be facilitated if authors will
consent to define their terms and to add a glossary to their books
and articles.

STYLES AND VOCABULARIES

Nor do the prospects opened up by lexical research for automatic
translation machines end here. Certainly it will be relatively easier
for the machine to translate scientific and technical works—and
these are at once the most urgently needed and the most economic-
ally rewarding translations. But the very logic of the work of
programming for the translating machine may lead on to bolder
enterprises.
Is there in fact any definite and firm line of demarcation be-
tween the translation of scientific and technical prose and that of
literary prose? Where does scientific vocabulary end and that of
 LEXICAL PROBLEMS OF AUTOMATIC TRANSLATION 101
literature begin? If, beginning with scientific prose, it is necessary
in order to solve the problems of automatic translation to break
down the vocabulary into different compartments so that the
memories of the machine can more easily digest it, this same
method is capable of extension to all vocabularies until all possible
groupings of words by subjects, or indeed from any other con-
ceivable angle, have been exhausted. The lexicography of the
translation machine thus leads finally to a general logical classifica-
tion of knowledge, that is to say of the words expressing knowledge,
the overriding law of such classification being that of commodity
of access to the data necessary for the translation of any given text.
The principal, if not the sole superiority of strictly scientific
texts over those of general literature probably consists in the
stricter limitation of the theoretical scope of their vocabulary, in
a more exact equivalence between the objects described and the
words used to describe them, that is in the higher quality of their
information content due to the more exact definition of the
meaning of each term.
But these are differences not of kind, but of degree. Beyond the
first subdivision of vocabulary by scientific disciplines, how far
will it be practicable to pursue the ramifications of a general
classification of all words, by subjects, areas of geography, periods
of history, social environments, etc.?
Only a thorough study of the best techniques of storage and rapid
access can reveal what is practicable. But theoretically at least, as
long as we keep to legitimate subjects of study and translation,
the possibilities are limitless. Geography, history, anthropology,
ethnology will each contribute their specialized vocabularies and
the terms used to describe the objects of these sciences will be
infinitely numerous and susceptible of varied cross-referencing.
In the same way that any good translator makes his own dictionary
for each new author or subject, a great deal of the time of automatic
translation programmers will be devoted, once relatively simple
texts have provided the groundwork, to the selection of those
different compartments of the electronic dictionary which must
be called into play for any given work.
Shall we compile a special sixteenth- or seventeenth-century
dictionary for the best historical novels? An Anglo-French dic-
tionary of the works of Victor Hugo? A German-Japanese one for
102 MACHINE TRANSLATION
Goethe? Will there not be room for an English-French Shakes-
pearian glossary, and so on? In translating the great authors will
it not be useful and perhaps even practicable to identify the exact
meaning of the terms employed by them at the various epochs of
their life? We are not here in the realm of science fiction; these
are practical possibilities for the day, which is not far off, when
critics, men of letters and literary translators come to make use of
machines and magnetic recording systems. The present writer is
gratified to find that on this as on other related subjects his
personal views have led him to conclusions very similar to those
of Professor Panov and Miss Bel’skaja. [25, 29]

THE SEMANTIC ATLAS

Lexical research for the translating machine opens up the way
towards a vast collective study of vocabulary, towards the enumer-
ation and classification on a national scale of all the words of a
language, arranged, in order of frequency and date of occurrence,
in specialized compartments of the electronic dictionary which
can be called into play as required by each translation programme,
or consulted like an ordinary dictionary on particular points of
detail. This lexical work would make it possible for research
workers to accomplish for the vocabulary of any language what
Gilliéron has done for French phonetics—to draw up an atlas of
meanings. A history of the changes in the meanings of each word
should also be possible. If each nation, each linguistic group,
participates in this study according to methods defined by mutual
agreement, language-to-language equivalences of meaning can be
patiently surveyed and recorded on magnetic tape, until they
constitute a collection of bilingual electronic dictionaries which will
enable automation to be applied progressively to an increasing
number of languages and subjects and to the literatures of the
present, past and future.
 CHAPTER VII

Future Prospects
LIMITATIONS OF THE MACHINE
WITHIN limits, automatic translation is already possible; all that
is required is that sufficient time and talent should be devoted to
the preparation of bilateral programmes. Despite differences in
theoretical approach between various schools of thought, success
will be achieved provided that, according to the rules of scientific
empiricism, all theories are turned to account.
It is now certain that the machine can transpose into a second
language, correctly—that is to say respecting the rules of grammar
and of syntax—a sequence of sentences written in an original
language. Naturally it will not at first be able to avoid displeasing
repetitions of the same word; it will not clarify ambiguities in the
original text; it will not always avoid facing the reader with a choice
between several alternative translations of a single word. It will
have no particular style, or, if it has, that style will be a somewhat
simplified style, that is to say it will transpose faithfully sequences
of words or groups of words without seeking those short cuts,
paraphrases and euphonies which a good translator who “rethinks”
the original will always find. The degree of semantic sophistication
of the machine will depend on that of its electronic dictionary; it
will correspond to the degree of complexity permissible in the
lexical programmes of the machine, which means, in the last
analysis, on the number of numerical indices by means of which
it is possible to determine useful choice between several alternative
meanings of a word without unduly burdening and slowing down
the programme.

ROLE OF THE MACHINE

Within these limits, what then is to be the role of the machine?
In the field of scientific and technical translation the automatic
translator will clearly be increasingly useful as and when the
104 MACHINE TRANSLATION
vocabularies of the various branches of science are inventoried and
recorded in bilingual form, according to accepted norms of bi-
lateral automatic translation. Already Oettinger’s Harvard auto-
matic Russian dictionary is proving its usefulness. The number of
semantic gaps will decrease as programmes become more sophis-
ticated and as bilingual vocabularies are completed. Such gaps
will be of two kinds: words not yet appearing in the dictionary—
which will be reproduced in the translation in their original form
—and rare idioms not yet registered in the “idiom dictionary”.
Such idioms will be translated word-for-word, a solution which,
according to Bar Hillel, will prove acceptable in a surprising
number of cases, when what is required is communication of
meaning and not a sophisticated rewrite bringing into play all the
arts of persuasion.
The circumstances in which the meaning of a scientific state-
ment is communicated to a specialist are, in fact, somewhat
peculiar. Thanks to their intimate knowledge of the same subject,
reader and author already enjoy, in most cases, a certain com-
munion of thought. The art of persuasion, rhetoric, plays only a
minor role in this meeting of minds. While the rough translations
provided by Booth and Richens in 1952 were already of some slight
value to scientists, those of Dostert and I.B.M., of Panov and
Bel’skaja, of Oettinger and the Harvard group, are largely adequate
for current scientific needs.
Moreover, it is possible, and probably essential, to foresee
several grades of presentation for texts translated by the machine.
Although it is true that translation will not be truly automatic
unless we can reduce to a minimum the role of the reviser improv-
ing the translated text, nevertheless, we cannot accept without
reserves the disappearance of the post-editor. The machine can
dispense with his services as regards the syntax and grammar of
the target language. Its raw output can be communicated to the
scientist, to a restricted circle of interested specialists. But if this
same text is to appear in a learned journal, or to be shown to a
meeting of company directors, it will probably have to be touched
up in order to reply in advance to the objections which the purists
will not fail to raise. If it is to be more widely diffused, then it is
certain that it will have to be carefully edited. And this is, in fact,
exactly the procedure adopted for conventional translation.
 FUTURE PROSPECTS 105

OPERATING COST
Clearly the great advantage of the machine lies in its speed of
operation: reliability will be a second advantage, once the vocabu-
lary has been established. Well-trained and experienced trans-
lators, whose translations nevertheless require revision and editing
before presentation to a relatively exacting public, normally
translate 300 words an hour, counting the time required for
research and careful preparation for the work. Even these good
translators are liable to distort shades of meaning when not fully
conversant with every detail of the subject. It is estimated that,
provided they have been adequately programmed, existing
machines could translate at the rate of 20,000 words an hour.
Such a speed of output will require long and painstaking prepara-
tion and will demand a considerable investment in human effort
and intelligence. Once achieved, this output will increase with the
greater potentialities of machines now in preparation and about to
be put into operation. Not only will this investment be amortized
over a considerable number of years, since it is a permanent one—
to all intents and purposes indestructible provided elementary
precautions are taken—but the effort of preparation will preclude,
or at least reduce, technical errors of meaning.
Moreover, good translators, of whom there are not enough at
the present time to satisfy scientific requirements, will be em-
ployed either for the preparation of the lexical programmes of the
machine or as revisers. Their productivity will be increased; once
again mechanization of the purely repetitive elements of a complex
activity will concentrate attention on other elements of that same
activity requiring intelligence and invention. The net gain for
science will be to render rapidly accessible works which today are
available without undue delay only to those specialists who are also
linguists: the division of labour in scientific research will be
improved—hence new possibilities of creative thinking and
cross-fertilization of minds.
Will the translating machine ever be a paying proposition?
This will obviously depend on the speed of execution of its pro-
grammes and on the use made of these machines once the pro-
grammes are established. The most recent estimates forecast
machine translations at a price definitely lower than the present
106 MACHINE TRANSLATION
cost of scientific translations. An American estimate quotes a
maximum cost price of the order of $.005 (half a cent) a word; a
more recent English calculation puts it at a maximum of 2s. a
thousand words [16] the cost of automatic translation of scientific
texts. Outside translations of such texts cost at present up to
£3 10s. 0d. per thousand words, counting the incidental overhead
expenses of commissioning outside translation. The English
author quoted puts at £4 2s. 0d. a thousand words the price of
outside translation, including administrative overheads. So that a
cost of $5 a thousand words for mechanical translation, unrevised
but technically perfectly correct, would be well worth while,
above all taken in conjunction with the high speed of the machine
which would greatly reduce the time lag. A cost of 2s. a thousand
words would mean a sensational saving, even if the price of the
preparation of the programmes and a normal revision fee had to be
added.
Apart from all question of purely commercial values, automatic
translation of languages brings appreciable advantages to the
linguistic group or national community. Once the initial effort has
been made and programmes established, it should free intellectual
ability for more productive work than that of run-of-the-mill
translation. Just as accounting machines perform mechanical
work formerly done by men, the translating machine will assume
the worst drudgery of the sometimes somewhat sterile business of
translation. The translator is often a man capable of invention,
of literary creation, of understanding subjects the complexity of
which requires a high level of general culture. Think for a moment
of the time such a man must devote to transposing from one
language to another the personal pronouns, definite and indefinite
articles, prepositions, conjunctions, everyday words and auxiliary
verbs. When we read sentences, which the translator must
translate from beginning to end, let us stop to consider the idea
of redundancy in language with which the mathematical theory of
information has made us familiar. Of all the words in a given
sentence, how many are essential to the transmission of the
author’s message and how many are simply conventional signals
forming part of the linguistic mould of thought, but not of the
actual thought expressed by the author? If the whole work of
translating this mould can be mechanized, and if, in addition, the
 FUTURE PROSPECTS 107
automation of translation processes can be brought to bear on all
or almost all those words of which the meaning really matters,
may we not then expect to see a significant release of energy and
talent? The tool now being forged will soon become an indis-
pensable part of the intellectual equipment of every nation and
its use should speed up the rhythm of the acquisition of knowledge
and lead us to a wider and more equitable distribution of en-
lightenment.
LITERARY PROSE
Is it too soon to envisage the extension of its use to tasks normally
considered as literary—the translation of general information, of
books of travel and geography, of novels, of philosophical and
critical works? Which of us has not in the past read translations
of foreign literature, under the auspices and even under the
signature of well-known authors, in which wrong shades of
meaning and misconstructions have abounded because the
translator—or his hack—had translated the words without
understanding their meaning, or failed to recognize idioms,
relying haphazard on his dictionary, or worse still, on his own
intuition? Though frequently of blatantly poor quality, literary
translation, when it plays its proper role, serves to build a bridge
between different cultures. What counts is the imaginative effort
to transpose not words, but representations. This is a supra-
linguistic effort, delicate and complex in that the details of the
representation of the reader rarely coincide with those of the author
as between one culture and another. The role of the translator is
to establish between the two a zone of intercommunication
bounded by the evocative value of words. Here we enter the
domain of supra-semantic evocation, of subconscious association,
of harmonics and the magic power of words. Are we bold enough
to trespass with our machine into this sacred realm of the in-
dividual and the imponderable?
Let us imagine that it is desired to translate into French a
contemporary novel written in Hindi, the action of which takes
place in a village in the Punjab. If the translator has in fact
participated in the life of such a village and if, therefore, all the
words of the original have for him their full local evocative power,
he will at once find himself faced with the problem of translating
108 MACHINE TRANSLATION
the everyday vocabulary, not to mention for the moment the
philosophical and religious undertones and the rhythmic and
euphonic values of the original. Is he to speak of the "maire" and
the "garde champêtre"? Should he call the houses, the familiar
objects, cows and horses, officials and trees by their Hindi names
transcribed into French, or should he seek equivalents in the
everyday vocabulary of France? Neither of these two solutions is
without its drawbacks: to take the reader completely out of his
own element renders understanding impossible, whereas to
gallicize everything is completely to destroy all local colour and
feeling.
Similar problems must have faced the first translators of Russian
novels, and it will be remembered how much the solution of
partial translation left to be desired. In reality, we are dealing here
not with translation pure and simple, but with wide cultural
exchange and with the interpretation of one culture to another.
We have to create for the French reader an atmosphere which he
is only partially prepared to perceive, to take him out of his own
element without permitting him to get lost: we must enable him
to follow the thread of events and feel at least a part of the inner
meaning of each scene. The brunt of this delicate mission will fall
upon vocabulary. Only by weaving and reweaving the threads
of his translation will the translator succeed in finding the right
proportion of Hindi words to convey local colour and of French
words to facilitate the adaptation of his reader to the change of
scene.
How far will it be possible to “pre-fabricate”, so to speak, this
vocabulary when preparing a programme of automatic translation,
by establishing in advance a mixed vocabulary peculiar to such a
translation? There is nothing inconceivable about such an opera-
tion, which might even prove to be a paying proposition for a
work of considerable proportions or for a series of works of lesser
size. The machine would in no way take the place of man, but
would perform for him certain ultra-rapid tasks in accordance
with directives determined by the translator, who would then
improve the detail of the machine’s rough draft, as does any good
reviser of human translations. The machine would simply have
“devilled” for the man, but would no doubt prove to be more
docile than a human hack and would lay fewer traps for the reviser
 FUTURE PROSPECTS 109
in the form of wrong shades of meaning and plausible mis-
translations. All is here a question of proportion, of common
sense and of comparative costs.

COLLECTIVE METHODS
We shall perhaps see the modern equivalent of certain successful
translation teams of old reconstituted round the machine. But the
work of translation will be distributed somewhat differently: the
machine will replace the hack translator, while part of the team
concentrates its energies on thorough preparation of vocabulary
and others are detailed to work on revision and stylistic im-
provement of the translations coming from the robot at great
speed.
Here again the machine leads to collective methods of work,
by concentrating and accelerating certain means of production
which in the modern world are rarely suited any longer to in-
dividual work and are therefore best used collectively. Electronic
memories, by their reliability and their speed of reference, make
it possible to devise methods of translation which will associate
the best specialists of languages and of the sciences, each col-
laborating in the common task and contributing to it his individual
knowledge. The collective methods of research laboratories are
those needed for this literary and artistic work. Thanks to elec-
tronic memories this collaboration will be effective not only now,
but will carry over from one generation to another, as does the
work of the great lexicographers: each programme, each diction-
ary, once established, can serve an indefinite number of times and
can be improved and transformed through the centuries while
safeguarding all that is best and most worthwhile of what has been
established earlier.
POETRY

And now we must come to a question which has long lain in wait
for us. Will the machine translate poetry? To this there is only
one possible reply—why not? All of us have done it in our
schooldays, when neither our Latin syntax, nor our grammar,
nor our vocabulary, nor our sense of rhythm, nor our skill in
rhyming could rival those of the electronic machines of to-
morrow. Do not let us ask the machine to do more than a
110 MACHINE TRANSLATION
minimum, but let us see what this minimum may be, how we
can, if possible, improve upon it, and what lessons it can provide
for the future.
The task of the translator becomes progressively more com-
plicated and sophisticated as the text for translation grows further
removed from straight description or narrative, as its vocabulary
becomes more connotative and less denotative, and as extra-
linguistic elements, such as the elements of situation in dialogue
of novels or plays, take precedence over those strictly linguistic
markers sought in the sentence by the programmers of automatic
translations. In dialogue, in poetry, in “stream of consciousness”
writing, in everything which suggests a momentary individual
representation rather than a Cartesian expression of a clearly
defined concrete or abstract reality, the task of the translating
machine will become extremely difficult. Not that there is any
difference in kind between this language and the language of
scientists; but the search for lexical equivalents between two
languages becomes more problematical and depends on a greater
number of factors, some of them extralinguistic. The private
understanding between author and reader of a scientific text does
not necessarily exist between the reader and the author of a poem.
The more the choice of words becomes an individual matter
instead of being dictated by the constraints of a scientific dis-
cipline, the greater will be the number of sub-routines required
by the machine or by the translator for lexical research, and the
less likely it is that such research will prove economic or even
possible. How can the machine succeed in a domain where the
magic of sound and rhythm, of extraneous semantic evocation are
the imponderable guides of the sensitive translator?
Between metalanguage and pure poetry, from the clear and
distinct expression of a scientific representation to the synthetic
expression of the vibrations of the poet’s ego at the centre of
his individual universe, there exists a whole vast range of un-
translatables. All translation is an approximation, because
language alone is translated while metalanguages require no
translation. If we dare to reply “why not?”, it is because from
the Cartesian absolute of metalanguage to the mystic absolute
of pure poetry, there are differences not of kind but only of
degree.
 FUTURE PROSPECTS 111
Poet or geometrician, the true writer gives to language both
its full connotative and musical harmonics and its full denotative
value. Is it not possible that by tackling boldly the difficulties of
poetic translation, we may hit upon the solution to some of the
more profound problems of scientific or narrative translation?
It would be foolish to assert that the machine will translate poetry
as successfully as a handful of poets have done, but it would be
worse than folly, once the instrument has been forged, not to
make use of it to take the measure of its failures as well as of its
successes, and to find out where and why it fails.

STUDIES IN POETICAL SEMANTICS

Such an analysis would be based on many aspects of language
which it has not been possible to discuss in the present study.
Together with work on written language and its translation we
have mentioned briefly the existence of research into the phonic
aspects of language. Sooner or later the study of written language
and that of spoken language will combine not merely for oral
control of the movements of elevators, for automatic simultaneous
interpretation and stenotyping, but also for the study of the
rhythms and sounds of poetry and of prose. Electronic recording
will make it possible to conserve and to classify sounds and
rhythms, and to submit them to exact and objective study of a
type impossible for the critic working with only intuition or card
indexes to guide him. In the same way studies of comparative
semantics, indispensable for the constitution of electronic bi-
lingual dictionaries for machine translation, will be directed be-
yond the bounds of utilitarian requirements towards the semantic
analysis of poetry, which will take into account both suggested and
expressed meaning. Matila Ghyka, noting Mallarmé’s poetic
predilection for the words azur, vierge, or, cristal, glacier, has
sketched the outline of a semantic study of the sound and “shape”
of words, showing the way to a purely disinterested research to
which the new machine methods can contribute. Thorough studies
of comparative semantics will make it possible to determine to
what extent it is possible to find equivalents, in other languages,
for the connotative value of words, so vital a factor in poetry.
Already Reifler’s work contains interesting suggestions in this
direction. Automatic translation workshops should be so managed
112 MACHINE TRANSLATION
that they can put at the disposal of literary research part of the
time of their machines and of the experience acquired by their
personnel in thorough and penetrating analysis of all aspects of
language.
Electronic machines have composed lyrics and verses, have
invented rhymes and composed music. Will there one day emerge
from the sorry mechanical monotony of these attempts, as some-
times from anthologies of anonymous poets of bygone ages, lines
which at least evoke, let alone create, the fleeting thrill of human
emotion? This is at least as statistically probable as the recreation
of the works of Shakespeare by a monkey blindly typing through
eternity on the keys of the typewriter. Will man be able to lead
the machine beyond this random search for beauty, and so direct it
that he can speed up its creative course, thus suspending, at least
for a moment, the rule of chance? To attempt to translate poetry
by machine, after full analysis of the constitutive elements of
poetry, is a more alluring proposition than to teach a robot to
make rhyming couplets. It is a proposal that should attract lovers
of poetry as well as iconoclasts—all those, in fact, who wish to
penetrate the secrets of verbal creativeness.

LITERARY ANALYSIS
In fact the problems of automatic translation is here only one of
the many aspects of the application of electronic machines to
literary analysis. Linguistic data systematically registered in
magnetic memories or on punched cards or tape can be subjected
to a number of analyses in the same way as can scientific problems
or questions of management in complex business enterprises. If
the translation of scientific texts represents a utilitarian aspect of
this new science—the analysis of discourse with methods offered
by electronics—it is also true that computers make it possible to
submit all the elements of spoken and written discourse to a
systematic study impossible with the individual, manual methods
of yesterday. Father Roberto Busa has demonstrated this by his
studies of the work of St Thomas Aquinas and the Dead Sea
Scrolls. He has perfected a method of making, within a relatively
short period, a concordance of the Summa Theologica and an
index of the Dead Sea Scrolls. In the latter case the mechanical
analysis programme has permitted the reconstitution of words
 FUTURE PROSPECTS 113
missing in the manuscript, on the basis of studies of the frequency
of certain word groups. The application of similar techniques to
literary, juridical and scientific studies is still in its infancy, but
an attempt is already in progress to extend it to automatic
summary records and abstracting.
In a short chapter in their book, [7] Booth, Brandwood and
Cleave have summarized the various aspects of contextual and
structural analysis made possible by the use of electronic com-
puters: frequency counts on the lines of those of Estoup and
Zipf, but employing more rapid and more reliable techniques;
biblical and other concordances; the constitution of a dictionary
of syntactic structures. They deal at greater length with stylistic
analysis, the chronology of the works of Plato and the mechanical
study of rhythms and syntax in the Dialogues. Their book provides
us with glimpses of a whole new technique shortly to be at the
disposal of literary research, which cannot afford to neglect these
new methods.
How much time will such analyses require in detailed pre-
paration for the high-speed work of the machine? This, Booth
replies, will depend on the number of persons working on the
problem. It is no longer a case of a work of laborious scholarship
undertaken by one man at the beginning of a lifetime of patient
work: there must be a new division of labour, with a hierarchy for
the formulation of exact rules to be strictly applied by all. A
hierarchic division of labour on these lines is already apparent in
the team of Father Busa at the Aloysianum at Gallarate, where
leadership is in the hands of the inventor of the research while the
tasks of execution are spread among less ingenious technicians.
Thus, in order to make use of modern techniques, literary research
will have to become collective, as scientific laboratory research
already is. This evolution must be fully comprehended and con-
trolled so that we do not fail to safeguard the essential: respect for
and knowledge of the creative genius of man, the secret of which
both translator and analyst are endeavouring to fathom. Machines
bring us the means to know and to understand writers better, in
an age when there is a greater need than ever before for the human
community to affirm the right of all men to culture and to know-
ledge, and to see to it that every nation and every individual has
the means to make this right a reality.
114 MACHINE TRANSLATION
SPEEDING UP CULTURAL EXCHANGE
Chance alone has not decided Soviet scientists to work on the
elaboration of automatic translation programmes into Russian from
Arabic, Burmese, Chinese, Hindi, Indonesian and Vietnamese.
Nor is the importance attached to Asian languages by Soviet
linguists entirely political. Newly independent nations need to be
able to read in their own language the works of other countries.
They would also like to see their own literature translated into
other tongues. The automatic translation programme into Russian
represents the first step in the process of exploring the linguistic
relationships between languages which will end in two-way
translation programmes. As we observed in Chapter I, automatic
translation can accelerate the contacts of these young nations with
other peoples, helping them to affirm their personalities and to
safeguard their own cultural heritage, enriching it at the same
time by contact with other cultures. This would seem to be the
only way of crossing before it is too late, without compromising
the originality and diversity of cultures, the barriers raised be-
tween peoples by linguistic difference. In a world where cos-
mopolitan currents establish themselves thanks to swiftness of
communication, and where a small number of languages might, for
economic and strategic reasons, come to impose their hegemony
over the whole world, to multiply authentic translations is one way
of defending the profound originality of national cultures, against
the tendency towards standardization brought about by more and
more uniform techniques and by the world-wide spread of a
universal technical terminology.
WHAT REMAINS TO BE DONE?
The translation machine, together with revolutionary new
techniques in linguistic analysis, is now on our doorstep. In order
to set it to work, it remains to complete the exploration of lin-
guistic data by means of comparative lexical and structural
analyses, first bilateral, then multilingual. If, in these pages, we
have devoted more space to language than to machines, it is
because the prime necessity is the adaptation of linguistic studies
to the new techniques. It is impossible to emphasize this point
too strongly: machines capable of translation already exist—it
remains for men to learn how to make use of them.
 FUTURE PROSPECTS 115
To inventory words and meanings, to undertake statistical
studies of semantic frequency, to catalogue inflexions and their
grammatical functions, to analyse word order and its exact
significance or value, to list types of structure and their meaning—
these first tasks can be greatly facilitated by the use of tabulators
and computers. Simultaneously, or perhaps subsequently, it will
be necessary to plan, organize and keep up to date a large electronic
dictionary fully adapted to the potentialities of existing machines,
and constantly to improve the operation of this dictionary in the
light of the evolution of machine techniques. It will be useless to
await the perfect machine before setting to work. Nature did not
await the human brain before creating nervous tissue. As soon as a
dictionary designed and compiled for the translating machine is
registered on magnetic tape, it will be an easy matter to transfer
it on to any alternative form of memory offering greater speed of
access or higher capacity. Technically, the go-ahead signal has
been given and we can count on the new techniques to facilitate
progress in the tasks they have opened up to us.
The third task, which can be tackled at the same time as the
first two, will be the construction of a machine specially adapted
to translation needs. Booth has somewhat optimistically estimated
its cost as between £50,000 and £100,000. Such a machine would
be of no avail without programmes, and without it the best
programmes would serve no useful purpose. Meanwhile, it would
be a waste of more powerful machines, capable of more complex
operations, to use them continually to perform translations de-
manding no operation more complex than addition and sub-
traction. Thus programming can begin before the ideal machine
is available, but should be undertaken only in close collaboration
with the technicians who know existing machines and are able to
design new ones fully adapted to translation.
No attempt has been made here to conceal or omit all that still
remains to be done. We have tried to give a synthesis of recent
work without wearying the reader with too many technical de-
tails. A great many problems in fact remain unsolved. But the
way is now open and one solution often leads to another. An
attempt has been made to explore the complexity of linguistic
data, and it has already been established, for example, that the
translation of a twenty-word sentence may require as many as
116 MACHINE TRANSLATION
10,000 logical machine operations. If the translation of more
complex sentences is to be performed rapidly enough to be
economically interesting, we shall have to discover more ex-
peditious methods, making it possible to reduce such operations
to a minimum and economize some millionths of a second per
word. The exploration of linguistic structures will have to be
pursued to the very end, so that we may discover whether it is
really possible to translate not word-by-word but clause-by-
clause, as anticipated by the structuralists. We shall have to solve
the problems of self-programming, so that the machine can choose
for itself the most effective programme for a given structure.
If much remains to be done in the field of simplification and
rationalization of programmes, the same is true of the acceleration
of machine input methods. Tape-punching is slow and relatively
costly, particularly for Cyrillic, Arabic and Asian scripts and for
ideographic languages. Direct reading by the machine, with
automatic coding of printed text, without human intervention,
would be the ideal solution, and work is already in progress to
this end. These are but a few examples of the type of problem with
which it has been impossible to deal here since such questions
are subsidiary to the fundamental analysis of language for auto-
matic translation. Finally, it must be added that if the perceptron,
a new machine, based on the ideas of Ross Ashby, fulfils its early
promise and can be trained to recognize patterns, it should
provide the solution to the remaining problems of syntax and
structure.
While a great deal remains to be done, it can be stated without
hesitation that the essential has already been accomplished. Before
broaching the problem and taking stock of it, men had to free
themselves of taboos; that having been done, the rest is a matter
of technique only. Booth’s bold thinking, Reifler’s patient lin-
guistic ingenuity, Panov’s scientific dialectics and empiricism,
have won acceptance for a new attitude towards the study of
language—an attitude which, while respecting the individual
qualities of a spoken or written text, is nevertheless fully de-
termined to explore these qualities and as far as possible to
imitate them in another language. The translating machine recalls
to mind a very simple tool, the use of which has long since ceased
to be considered sacrilegious: the pantograph, with which a
 FUTURE PROSPECTS 117
workman can copy, in the material of his choice, the marble
Venus of Milo, without disrespect for the inspiration and genius
of the unknown sculptor. Translating machines will soon take
their place beside gramophone records and colour reproductions
in the first rank of modern techniques for the spread of culture
and of science.
 Postscript to the English Edition
THIS book was completed in its original French version by the
end of December 1958 and published in May 1959. In June of this
year the author attended the Unesco conference on information
processing, which enabled machine translation specialists to meet
and exchange views on the state and prospects of their work.
With few notable exceptions most of the schools of research
mentioned in the foregoing pages were represented.
The conference provided an opportunity to have a mechanical
translation made, without previous trial or preparation, of a
foreword specially written in French for this English edition so
that it might be so translated. No choice of programme or of
machine was possible. It so happened that Mr A. F. R. Brown of
Georgetown University had brought to Paris his recorded French-
to-English translation programme and vocabulary, designed for
the translation of texts on chemistry and nuclear energy; the
vocabulary is of some 4,000 words and the programme operates
at the speed of 5,000 to 10,000 words per hour, using an IBM 704
computer available in Paris at the I.B.M. headquarters.
The text of the foreword was given in French typescript to
Mr Brown at 5.30 p.m. on 19th June 1959. He proceeded to the
I.B.M. headquarters where he keypunched it; the text as entered
into the machine is shown in Figure 1. The figure “1” following a
letter means “acute accent” while the figures “2” and “3” con-
ventionally designate the grave and circumflex; “$ FIG” means
that the signals following it are figures and not letters of the
alphabet, “$PAR” meaning “new paragraph”.

The translation, produced and handed over by 6.30, is also re-

produced in Figure 5. From it the reader can deduce that the words
have been analysed into stems and affixes in such a way that the
mark of the acute accent in “spécialisé » is retained in English in
the translation “specialised”. This seems due to the literal trans-
cription of stems which are alike in English and French. Words not
recorded in the memory of the machine because they are not
120 MACHINE TRANSLATION
relevant to the vocabulary of chemistry and nuclear physics, such
as “calculatrices”, “traductions”, “langues”, etc., appear un-
changed in the English output. One word, “encore”, seems to
have caused some technical hitch, coming out as it did as “000017”.
This programme is clearly still inadequate for the proper
translation of present participles used as adjectives, such as
“satisfaisant” and “exigeant”, which have been treated as verbs
and not subjected to the rule for changing word order—unlike
“spectaculaire” which has been correctly translated and placed.
Similarly, prepositions require further programming—e.g. to
avoid such renderings as “to present at the reader” ; past participles
(définies) and nouns having the same forms as past participles
(découvertes) also, as do pronominal verbs (s’améliorer) and some
adjectives (large Britain).
The machine does not work without humour. The programme
being originally designed for chemistry, words such as “brom-
ure”, “iod-ure”, “carb-ure” are stripped of their chemical affix
and rendered by “bromide”, “iodide” and “carbide”. Hence—
and that is a fault in programming—the feminine adjective form
“future” has been wrongly stripped of a false affix and a new
chemical has been invented by the machine, “futide”!
One alteration was suggested by Mr Brown before keypunch-
ing: the addition of “en” before “croissant.” But no attempt was
made by the author to simplify style or to avoid idioms such as
“se tenir au courant” or “mise au point”, mechanically translated
by “hold at the current” and “put at the point”.
Two Russian-language versions of the same Avant-Propos
were made by Mr Michael Corbe: one of them follows the French
original almost word for word; the other is on the contrary free
and easy. Mr Corbe left both versions with Dr Don Swanson of
the Ramo-Wooldridge Corporation at Los Angeles, where both
versions were in their turn translated into English by means
of Dr Swanson’s experimental programme for the translation of
Russian physics texts. This programme is stated to be capable of
producing considerably better than “word-for-word” translation.
It is of course subject to the same lexical shortcomings as all such
programmes, and indeed as any human being, in that it cannot
translate words which it has never encountered before, and which
are therefore not recorded in its memory. On the other hand it is
 POSTSCRIPT TO THE ENGLISH EDITION 121
well-equipped for syntaxical analysis, on which Dr Swanson’s
team has concentrated, and it has an efficient method for stripping
off and interpreting in English the flexional endings of Russian.
The Appendix contains, set in four columns, four of the texts
involved in this multiple experiment: column I shows in sequence,
numbered vertically, each word or coherent group of words of the
original French text. Facing each French horizontal line, the
reader will find in column II, presented in the same manner,
Mr Brown’s machine translation into English, together with (in
italics) those indispensable editorial amendments where French
words had been used by the machine instead of their English
equivalents. Here the wide lexical similarity of English and
French through their common Latin stock of words, has un-
doubtedly facilitated the task of the machine and of the post-
editor.
Column III shows, similarly laid out, Michael Corbe’s “human”
word-for-word translation. Finally column IV gives Dr Swan-
son's machine-made English translation from that particular
Russian text; it has been so arranged as to show in italics those
editorial improvements introduced at Ramo-Wooldridge Corpor-
ation in pencil on the actual machine output, as part of the cumu-
lative process of lexical amelioration of the programme. It will be
seen that those lexical editorial changes are quite numerous
whereas the grammatical structure has on the whole been very
accurately rendered by the machine. Moreover, for each Russian
polysemantic word the machine has given alternative translations;
the one selected by the post-editor has been given first. Such
editorial choices are later recorded in the memory with con-
textual information as part of the cumulative improvement
process. A similar procedure is applied to idioms, such as deržat'sja
v kurse, the correct rendering of which will now have been re-
corded for future use in the Ramo-Wooldridge programme.
Careful examination of those four columns in the Appendix will
it is hoped give the reader a fairly precise idea of the present state
of machine translation and of the practical means by which it is
being improved.
The Ramo-Wooldridge version of Mr Corbe’s free and easy
Russian translation has not been added here as a further illustra-
tion. It differs from the translation shown in column IV by its
122 MACHINE TRANSLATION
better flow of English phrase—while it suffers from the same
lexical shortcomings.

The part of the conference and the symposium devoted to

mechanical translation permits certain conclusions which streng-
then the main thesis of this book. Machine translation is now not
only possible, it is actually being carried out, but not as a finished
product and mainly in an experimental and fact-finding spirit. Its
end-products are and will be many, and perhaps at this stage the
least important is translation itself. The method of gradual accu-
mulation of carefully checked data, using the machine as a means
of objective analysis of language, is now consecrated in the work
of Harper and Hays at Rand Corporation, of Oettinger and his
team at Harvard; it appears to be gaining ground even among the
research groups with a more theoretical approach. This method
makes the fullest use of the electronic computer and its ancillary
machines, sorters, tabulators, etc., to subject the data recorded
about words in a text and its translation, to successive analyses
from various points of view. Language data are indeed processed
not only with translation in mind but with the aim of obtaining
the widest and deepest penetration of such facts as the relation-
ships between words. Harper and Hays have in particular pre-
sented a method of analysis of structures based on the dependency
and precedence relationship between words which promises
considerable simplification in structural analysis by machine.
It was unfortunate that neither Miss Bel’skaja, who had sub-
mitted a remarkable paper, nor Professor Panov who had prepared
the survey on the present state of machine translation, could
attend the conference owing to illness. Russian work in this field
was ably represented by Miss Kulagina, whose approach is based
on mathematical theory—this at a time when the more empirical
approach of Panov and Bel’skaja appears to be winning support
even from the more theoretically-minded Western research teams.
It is to be hoped that the original views expressed by Bel’skaja
on the feasibility of translating poetry by machine will be de-
veloped and clarified, and also that she will be able to give wider
dissemination to her sub-routines for the analysis of English
sentences.
 POSTSCRIPT TO THE ENGLISH EDITION 123
From these meetings it clearly emerges that M.T. has reached
a stage where theories must temporarily recede into the back-
ground while practical laboratory work and machine processing of
long consecutive texts is indispensable to further research and to
the development both of practical programmes and of new working
hypotheses, based on the study of numerous facts rather than on
intuitive preconception. This perhaps did not require proof, but
it will not be amiss, for the future of research in language-data
processing, that this international conference brought it out in the
full light.
The critical attitude of M.T. specialists towards traditional
grammars was moderate and tactful; it was felt that they need
improving and completing rather than scrapping, and that to start
from them and work on their improvement is better than to start
from scratch.
Another fact which struck members is the international character
of this type of research, and the need for close co-operation
between national centres conducting language studies. More
instructive perhaps than the public meetings were the small
informal gatherings in which specialists—most of them very young
—exchanged information about details of their methods of
analysis and of recording facts in the memories of the computer
and of arranging programme routines.
The general conclusion is one of optimism. Machine trans-
lations today are still very imperfect. But the way to perfecting
them is clear. The field is attracting talented people in increasing
numbers. One of the major problems is to produce programmes
which do not ramify into excessively time-consuming sub-routines
while solving most of the problems of sentence structure if not of
polysemy. And the growing experience of programmers points
to man’s ability to observe the behaviour of the machine and give
it a chance to solve simply problems which at first baffle the mind
because we have not yet learned to state them simply. We can trust
the machine to teach us precisely that, because its fundamental
methods are simple.
 Bibliographical Notes
[1] Andreev, N. D., “Mechanical Translation and the Problem
of an Intermediary Language” (in Russian) in V.J., 6 (5).
(See below No. 30.)
[2] Ashby, William Ross, Design for a Brain, Chapman & Hall,
London, 1952.
[3] Ashby, William Ross, An Introduction to Cybernetics, Chapman
& Hall, London, 1956.
[4] Belevitch, Vitold, Langage des machines et langage humain,
Hermann, Paris, 1956.
[5] Bel’skaja, I. K., “Machine Translation of Languages,” in
Research 10 (10), October 1957, Butterworths, London.
(See also No. 29 below.)
[6] Booth, A. D. and K. H. V., Automatic Digital Calculators,
Butterworths, London, 1953.
[7] Booth, A. D., Brandwood, L., and Cleave, J., Mechanical
Resolution of Linguistic Problems, Butterworths, London,
1958.
[8] Booth, K. H. V., Programming for an Automatic Digital
Calculator, Butterworths, London, 1957.
[9] Cary, Edmond, La traduction dans le monde moderne, Georg,
Geneva, 1956.
[10] Cherry, Colin, On Human Communication, Technology Press
of the MIT and John Wiley & Sons, New York, Chapman
& Hall, London, 1957.
[11] Communications Research Centre, University College,
London, Aspects of Translation, Secker & Warburg,
London, 1958.
[12] Delavenay, E. and K. M., Bibliography of Mechanical Trans-
lation, Mouton & Co., The Hague (in preparation).
126 MACHINE TRANSLATION
[13] Fries, Ch. C., The Structure of English, an Introduction to the
Construction of English Sentences, Harcourt, Brace &
Company, New York, 1952.
[14] Guilbaud, G. Th., “La cybernétique.” Collection Que sais-je?
No. 638, Presses Universitaires de France, Paris, 1957.
[15] Hays, David G., Harper, K. E., and others, RAND research
memoranda: M.T. Studies 1-9, RAND Corporation, Santa
Monica, California. (See also No. 29 below.)
[16] Liebesny, F., “Economics of Translation,” in ASLIB
Proceedings, 10 (5), London, May 1958.
[17] Locke, W. J., and Booth, A. D., and others, Machine Trans-
lation of Languages. Technology Press of the MIT & John
Wiley & Sons, New York, Chapman & Hall, London, 1955.
[18] Kulagina, O. S., and Mel’čuk, I. A., “Machine Translation
from French into Russian” (in Russian) in V.J., 5 (5). (See
below No. 30.)
[19] Mološnaja, K. T., “Some Syntactic Problems in Machine
Translation from English into Russian,” (in Russian) in
V.J., 6 (4). (See below No. 30.)
[20] Mounin, Georges, Les Belles Infidèles, Cahiers du Sud, 1955.
[21] Mounin, Georges, “Pseudo-langues, interlangues et meta-
langues,” in Babel, 4 (2), June 1958.
[22] M.T. = Mechanical Translation, devoted to the translation
of languages with the aid of machines, Massachusetts
Institute of Technology, 1954-.
[23] Muhin, I. S., Opyty avtomatičeskogo perevoda na elektronnoj
vyčislitel'noj mašine B.E.S.M., Academy of Sciences of the
U.S.S.R., Moscow, 1956.
[24] Murphy, John S., Basics of Digital Computers, Rider, New
York, 1958.
[25] Panov, Dmitrij J., Avtomatičeskij perevod, 2nd edition,
revised and augmented, Academy of Sciences of the
U.S.S.R., Moscow, 1958.
 BIBLIOGRAPHICAL NOTES 127
[26] Panov, Dmitrij J., Concerning the Problem of Machine
Translation of Languages, Academy of Sciences of the
U.S.S.R., Moscow, 1956 (in English).
[27] Sestier, A., Les calculateurs numériques et leurs applications,
Hommes et techniques, Neuilly-sur-Seine, 1958.
[28] Tezisy konferencij po mašinnomu perevodu, 15th-21st May
1958, Ministry of Higher Education, Moscow, 1958.
[28a] Tezisy soveščanija po matematičeskoj lingvistike, 15th-21st
April, 1959, Ministry of Higher Education, Leningrad,
1959.
[29] Unesco: Proceedings of the International Conference on
Information Processing. Unesco & Oldenbourg, Munich;
Butterworth, London; Dunod, Paris (in preparation).
[30] Voprosy Jazykoznanija, Academy of Sciences of the U.S.S.R.,
Moscow, 1955 and after.
 Glossary
ALGORITHM or Algorism: in general, the art of calculating with
any species of notation; in particular the word is used by
computer programmers to designate the numerical or
algebraic notations which express a given sequence of
computer operations, define a programme or routine
conceived to solve a given type of problem.
BINARY CODE: A binary system of numbers or other marks (e.g.
electronic pulses, etc.), used to represent either decimal
digits or letters of the alphabet. See binary system below for
the binary coding of decimal digits.
BINARY SYSTEM: A system of counting using two as the radix, or
base, whereas the more familiar decimal system uses ten
as the base. Only two characters or symbols are used, 0 and
1, or + and —, or, in electronic circuits, pulse and no pulse,
whereas the decimal system uses ten characters. The
following table shows the conversion from decimal symbols
to binary:
Decimal Binary Decimal Binary
0 0 5 101
1 1 6 110
2 10 7 111
3 11 8 1000
4 100 9 l001
Because electronic pulses and magnetic states have binary
form (on or off, pulse or no pulse, + or ― ) they lend
themselves easily to the recording of data in computers.
COMPLEMENTARY SIGNALIZATION: A device consisting of adding
conventional signals to existing alphabetical signs, first
developed by Professor Reifler to “pre-edit” sentences
prior to machine-translation. This system was abandoned
with the progress of computers and of language analysis
130 MACHINE TRANSLATION
for M.T., as it was found that where the human mind
unconsciously recognized signals in written sentences, the
machine can in most cases be programmed so as to recog-
nize them too.
CRYPTOGRAPHY: The act or art of writing in secret characters or
cipher: the science or techniques of cipher.
CYBERNETICS: A word derived from Greek kybernētikē, the art
of steering a ship, helmsmanship ; used by Ampère (1834)
to designate the study of means of government; then by
N. Wiener in Cybernetics, or Control and Communication in
the Animal and in the Machine, Hermann, Paris 1948.
Wiener chose the word under the influence of the Watts
“governor” on the steam engine, one of the earliest feed-
back controls ever invented. Cybernetics is used with
precision to mean the science of control mechanisms, and
loosely to designate the theory and practice of automata
and calculating machines, “thinking” machines, etc. See
Bibliography, J. Th. Guilbaud, for a sober appraisal of
this new science.
DESINENCE: Termination, ending of a word—more properly the
inflected ending of a word.
DIACRITICAL SIGNS: Distinguishing signs, marks, points or other
signs, attached to a letter or symbol to distinguish it from
another of similar forms: e.g. accents in French.
DIGIT: (Latin digitus, finger): each of the numerals below 10 in
decimal counting, 0 to 1 in binary (q.v.) counting.
DIGITAL COMPUTER: See DIGIT. As opposed to the analog computer,
which simulates the problems it is asked to solve, the
digital computer, derived from Pascal’s arithmetic machine
and from the desk calculator, works out numerical solutions
to problems, by calculations made with and on digits.
HOMOGRAPH: One or two or more words identical in spelling, but
of different derivation or meaning: e.g. French la route,
suivez-la, etc.
HOMONYM: A word having the same pronunciation as another, but
differing from it in origin, meaning and possibly spelling.
 GLOSSARY 131
Hence, HOMONYMY. Homophones are homonymous in
sound alone, homographs are homonymous in spelling.
HOMOPHONE: A letter or word having the same sound as another,
but differing from it in meaning and/or spelling. E.g.
French ou, où, houe; English: Rome, roam; bare, bear.
INFORMATION THEORY: An approach to the study of messages
transmitted in a language, based on the mathematical
theory of communication, as developed by communication
engineers in their search for economy and efficiency in
transmission of messages. The statistical study of language
in so-called information theory bears mainly on the
frequency of reference of graphemes and phonemes, but
can be developed in various directions more directly useful
to applied linguistics and M.T.
INPUT ORGAN: Any organ of a computer through which data are
fed into it—e.g. a punched-card reader, tape-reader,
photo-electric reader, etc.
INVARIANT: An invariable quantity—a term from the vocabulary
of modern mathematics. A semantic invariant is a constant
semantic fact which is found in languages having historical
or other connections, such as the evolution, in accordance
with certain laws, which would seem to be universal, of the
meanings of words first designating an object but later
acquiring an additional meaning. For instance French
“mouton”, sheep, “moutonner” said of a cloudy sky where
the small clouds suggest a flock of sheep: a similar semantic
evolution is found in Chinese.
LEXICAL CONTENT: The word-content of a language or a sentence,
book, etc.—its vocabulary.
LEXIS: A greek word meaning word, phrase, diction. Used here,
and by M.T. linguists, to designate the words of a language,
contained in its dictionary or lexicon, as opposed to the
morphology and syntax of that language.
MACROSCOPIC, MICROSCOPIC: Greek makrós, great, micrós, small
As opposed to microscopic study, which rivets its attention
132 MACHINE TRANSLATION
on infinitely small details, macroscopic study concentrates
on large-scale aspects of phenomena—for instance macro-
scopic linguistics bears on very general statistical rules of
language (e.g. Zipf’s law) rather than on individual aspects
of language or speech. Information theory (q.v.) has so
far studied mainly macroscopic aspects of language.
MICROSECOND : One millionth of a second.

MILLISECOND: One thousandth of a second.

MORPHEME : A form which cannot be analysed into smaller forms,
together with its corresponding meaning.
MORPHOLOGY: That branch of language study which deals with
the functions of inflexions and derivational forms, hence
MORPHOLOGICAL, pertaining to —.
M.T.: Machine translation, or mechanical translation.
PHONEMICALLY: A phoneme being a group of variants of a speech
sound (e.g. the e sound in get, tell, say, any, send)—phonemic
means “of the nature of a phoneme”—also “significant,
distinctive” (of sounds). Hence “phonemical”, “phone-
mically”.
PHONETOGRAPH: An instrument designed to record the sounds of
speech in the form of typewritten sequences of letters of
the alphabet.
POLYVALENT: A word from the vocabulary of chemistry, where it
means “having multiple valence”. By extension: potentially
capable of fulfilling several functions, playing several roles.
POLYSEMY: Multiplicity of meaning.
POLYSEMANTIC: (A word) having several meanings. Polysemantic
dictionary: a dictionary of words which have the common
characteristic of each having several meanings.
SEMANTICS (also semasiology): The branch of philology which
deals with meanings. Used here in contrast with syntax,
morphology and even with lexis, which is the set of words
of a language, as opposed to the various meanings of a
word, studied by semantics.
 GLOSSARY 133

SEMANTIC INVARIANT: See INVARIANT.

SEMANTIC UNIT: A unit of meaning, as opposed to units of
vocabulary, or to phonemes or morphemes. In amavi, am-
expresses “love” and "avi" the 1st person in the preterite:
both have semantic value, each is a unit. In je n'ai qu'un
livre, ne ... qu' is a semantic unit made up of two words.
SEMANTIC VALUES: The word “dog” has different semantic value
when it designates the animal, or a “fire-dog”.
SEME : From Greek sēmeion, a mark, sign, a unit of meaning.
STATISTICAL SEMANTICS: Statistical study of meanings of words
and their frequency and order of recurrence.
STRUCTURAL LINGUISTICS: A form of the scientific study of
language which concentrates on structures or patterns
(“he—the—a—” is such a pattern, which can be filled in
as “he gave the car a push”, “he found the boy a drink”,
etc.). Some schools of structural linguistics tend to dis-
regard meaning as unessential to the study of structure;
all emphasize the paramount importance of patterns in the
development of language and in its teaching.
SYNTACTIC: Pertaining to SYNTAX q.v. Syntactic value, significance
in terms of relations between words, of their syntactic
link (position, preposition, etc.).
SYNTAGMA: Arrangement of units in a syntactic construct, such as
“actor+action+goal” (the dog bit the man) .
SYNTAX: That branch of linguistics which deals with the arrange-
ment of syntagmata. Syntactic analysis for M.T. is the
analytical study of word-arrangement, with a view to
programming translation work in such a way that sentence-
for-sentence translation will be possible, as opposed to
word-for-word.
TECTONIC : Structural, pertaining to the structure of the sentence
or group of words.
 AVANT - PROPOS.
PRE1SENTER AU LECTEUR QUI N EST SPE1CIALISE1 NI
DANS L E1TUDE DE LA LING UISTIQUE NI DANS LA
CONNAISSANCE DES CALCULATRICES E1LECTRONIQUES ,
LES PROBLE2MES ACTUELS DE LA TRADUCTION
AUTOMATIQUE DES LANGUES , TEL EST LE BUT DANS
LEQUEL CE LIVRE A E1TE1 CONCU .
NOMBREUSES SONT LES DIFFICULTE1S QUI SE DRESSENT
ENCORE SUR LE CHEMIN AVANT QU UNE TRADUCTION
SATISFAISANTE POUR UN LECTEUR UN PEU EXIGEANT
PUISSE SORTIR D UNE MACHINE ,.. DE GRANDS
PROGRE2S ONT E1TE1 RE1 ALISE1S DANS L ANALYSE
DES LANGUES , ET LES PRINCIPALES ElTAPES DE
L E1TUDE DU LANG AGE EN VUE DE LA TRADUCTION
AUTOMATIQVE SONT MAINTENANT DE1FINIES .
LA RECHERCHE A PROGRESSEl DE FAÇON SPECTACULAIRE
DEPUIS $FIG 1955 ,.. DES TRADUCTIONS UTILES SONT
FAITES PAR DES MACHINES ET LEUR NOMBRE IRA EN
CROISSANT , LEUR QUALITEl S AME1LIORERA
CONSTAMMENT .
MAIS CERTAINES DE1COUVERTES SONT NE1CESSAIRES
POUR QUE CETTE RECHERCHE ENTRE BIENTO3T DANS UNE
NOUVELLE PHASE , CELLE DE L AUTOMATISATION. A2
$FIG 98 OU $FIG 99PCT .
$PAR LE LECTEUR QUI SOUHAITE SE TENIR AU COURANT
ET SUIVRE CETTE FUTURE E1TAPE DU PROGRE2S
SCIENTIFIQUE TROUVERA ICI UNE MISE AU POINT DE L
E1TAT ACTUEL DES TRAVAUX TEL QU IL RESSORT DES
OUVRAGES ET ARTICLES PARUS DE PUIS $FIG 1955 AUX
E1TATS - UNIS , EN GRANDE - BRETAGNE ET DANS L
UNION SOVIE1TIQUE .

FIG. 5. A SPECIMEN OF MACHINE TRANSLATION

(a) A Foreword to this book, as typed out in its original French in

the course of its mechanical translation on I.B.M. 784 computer.
This Foreword was written for the sole purpose of being so translated.
See page 119 for an explanation of figures in words and other
conventional symbols.
BEFORE - REMARK .
TO PRESENT AT THE READER WHICH IS SPECIALISED
NEITHER IN THE STUDY OF THE LINGUISTIC NOR IN
THE KNOWLEDGE OF THE ELECTRONIC CALCULATRICES ,
THE PRESENT PROBLEMS OF THE AUTOMATIC TRADUCTION
OF THE LANGUES , SUCH IS THE AIM IN WHICH THIS
BOOK HAS BEEN CONCEIVED .
NUMEROUS ARE THE DIFFICULTIES WHICH SET UP
THEMSELVES 0000170N THE PATH BEFORE A TRADUCTION
SATISFYING FOR A READER A LITTLE REQUIRING CAN
EXIT FROM A MACHINE /COLON/ LARGE ADVANCES HAVE
BEEN REALISED IN THE ANALYSIS OF THE LANGUES ,
AND THE PRINCIPAL STEPS OF THE STUDY OF THE
LANGUAGE IN VIEW OF THE AUTOMATIC TRADUCTION ARE
NOW DEFINITE .
THE RESEARCH HAS PROGRESSED IN A SPECTACULAR
MANNER SINCE 1955 /COLON/, USEFUL TRADUCTIONS ARE
DONE BY MACHINES AND THEIR NUMBER WILL INCREASE
CONTINUOUSLY , THEIR QUALITY WILL IMPROVE ITSELF
CONSTANTLY .
BUT CERTAINES DISCOVERED ARE NECESSARY FOR THIS
RESEARCH TO ENTER SOON A NEW PHASE , THAT OF THE
AUTOMATISATION AT 98 OR 99PCT .
THE READER WHICH SOUHAITE HOLD AT THE CURRENT
AND FOLLOW THIS FUTIDE STEP OF THE ADVANCE
SCIENTIFIC FIND HERE A PUT AT THE POINT OF THE
STATE PRESENT OF THE WORK SUCH THAT IT BE
EVIDENT FROM THE WORK AND ARTICLE PARU SINCE
1955 AT THE STATE UNITE , IN LARGE - BRITAIN AND
IN THE UNION SOVIE1TIC .

(b) Reproduction of the actual machine-translation of the same

Foreword, as typed out by the I.B.M. 704 computer in Paris on
19th June 1959. The French-to-English translation programme used
was conceived and designed by Mr A. F. R. Brown of Georgetown
University for the translation of texts on chemistry and nuclear
energy. A fuller explanation will be found on page 119.
See also Appendix for a comparison of this translation with a
machine-translation of the same text from a Russian version.
 APPENDIX
Two Machine Translations of the Same
Preface to this Book
The Comparative Tables below show:
I. French original of a Preface written for this book; II. Machine trans-
lation of this original from French into English by means of a programme
devised by Mr A. F. R. Brown of Georgetown University, on an I.B.M.
704 computer; translation made in Paris on a computer of this type;
III. Word-for-word Russian version of the same French Preface, pre-
pared by Mr Michael Corbe; IV. English version of this Russian text,
machine-made at the Ramo-Wooldridge Corporation, Los Angeles, on
an IBM 704 computer, under the direction of Dr Don Swanson.
Each text is presented vertically, one word or coherent group of words
at a time on one line. Comparison can thus be made between any two or
more vertical columns to see what happened to a given word.
KEY to column II:
ROMAN TYPE: machine output.
ITALICS : English translations, supplied by the author, of those words
for which the machine did not possess a translation in its memory.
KEY to column IV:
ROMAN TYPE: machine output in English, as accepted by Ramo-
Wooldridge post-editor.
ROMAN TYPE in square brackets: alternative machine output in English,
rejected by post-editor.
SMALL CAPITALS: machine output in Russian, i.e. words which the
machine could not translate because they were not in its memory.
ITALICS: English words supplied by the post-editor as part of the
process of cumulative improvement of his programme.
I II III IV
Author's Original Brown's Trans- Corbe's Transla- Swanson's
French lation, I.B.M. 704 tion, manual Translation,
I.B.M. 704
1 Avant-propos. Before-remark. PREDSILOVIE. Preface.
2 Présenter To preesnt PREDSTAVIT' To present [to
represent]
 APPENDIX 137
I II III IV
Author's Original Brown's Trans- Corbe's Transla- Swanson's
French lation,I.B.M.. tion, manual I.B.M. 704
704
3 au lecteur at the reader ČITATELJU to the reader
4 qui which KOTORYJ which (that)
5 n' NE is [there is, there
are]
6 est is EST'
7. spécialisé specialised SPECIALIZIROVAN SPECIALIZIROVAN
(specialised)
8 ni neither NI neither [nor,
not]
9 dans in v in
10 l'étude the study IZUČENIJ the study
11 de la linguis- of the linguistic LINGVISTIKI of the LING-
tique VISTIKI (of ling-
uistics)
12 ni nor NI nor [neither,
not]
13 dans la in the V ZNANII in the ZNANII
connaissance knowledge (knowledge)
14 des calculatrices of the electronic ELEKTRONNYH of electronic
[electron]
15 électroniques, calculatrices, VYČISLITELNYH numeral MAŠIN
(computers) MAŠIN, (digital
computers),
16 les problèmes the present AKTUAL'NYE actual
17 actuels problems PROBLEMY problem-s
18 de la traduction of the automatic AVTOMATIČES- of automatic
KOGO
19 automatique traduction PEREVODA translating
(translation)
20 des of the of
21 langues, langues, JAZYKOV, JAZYKOV
(languages) (languages)
22 tel such TAKOVA is such
23 est is EST' is [there is,
there are]
24 le but the aim CEL' the purpose
25 dans in V in [into, to]
26 lequel which KOTOROJ which
27 ce this ETA this
28 livre book KNIGA book
29 a été has been BYLA was
30 conçu. conceived. ZADUMANA. ZADUMANA
(conceived).
31 Nombreuses Numerous MNOGOČIS- Are numerous
LENNY
32 sont are SUT' are
33 les difficultés the difficulties TRUDNOSTI difficulty-s
138 MACHINE TRANSLATION
I II III IV
Author's Original Brown's Trans- Corbe's Transla- Swanson's
French lation, I.B.M. tion, manual Translation,
704 I.B.M. 704
34 qui which KOTORYE which
35 se dressent set up them- VSTAJUT VSTAJUT (arise)
selves
36 encore 000017 (still) EŠČE yet [still]
37 sur on NA on
38 le chemin the path PUTI the way [means]
39 avant before PREŽDE before
40 qu' CEM
41 une a ODIN one [alone]
42 traduction traduction PEREVOD PEREVOD
(translation) (translation)
43 satisfaisante satisfying UNDOVLETVOR- satisfactory
ITEL'NYJ
44 pour for DLJA for
45 un a ODNOGO one [alone]
46 lecteur reader NEMNOGO of not much
47 un peu a little TREBOVATEL' TREBOVATEL'
NOGO NOGO
(demanding)
48 exigeant requiring ČITATELJA reader
49 puisse can MOŽET can
50 sortir exit VYITI VYITI (emerge)
51 d' from iz from [of]
52 une a ODNOJ one [alone]
53 machine; machine; MAŠINY. MAŠINY .
(machine)
54 De grands Large BOLŠIE Large
55 progrès advances PROGRESSY PROGRESSY
(progress)
56 ont été have been BYLI were
57 réalisés realized DOSTIGNUTY attained
58 dans in V in
59 l'analyse the analysis ANALIZE the analysis
60 des langues, of the langues, JAZYKOV, of JAZYKOV
(languages) (languages)
61 et and I and
62 les principales the principal OSNOVNYE (the) principal
63 étapes steps ETAPY stages
64 de l'étude of the study IZUČENIJA of the study
65 du langage of the language JAZYKA of the JAZYKA
(language)
66 en in s with
67 vue view CEL'JU the purpose
68 de la traduction of the automatic AVTOMATIČES- of automatic
KOGO
69 automatique traduction
(translation) PEREVODA translating
 APPENDIX 139
I II III IV
Author's Original Brown's Trans- Corbe's Transla- Swanson's
French lation, I.B.M. tion, manual Translation,
704 I.B.M. 704
70 sont are SUT' are
71 maintenant now TEPER' now
72 définies. definite. OPRBDELENY. are determined.
73 La recherche The research ISSLEDOVANIE The investiga-
tion [research]
74 a progressé has progressed PROGRESSIRO- PROGRESSIRO-
VALO VALO (has
progressed)
75 de façon in a spectacular EFFEKTNYM by the EFFEKT-
NYM (effectively
76 spectaculaire manner OBRAZOM way
77 depuis since S from [with]
78 1955. 1955. 1955. 1955.
79 Des traductions Useful POLEZNYE Useful
[effective]
80 utiles traductions PEREVODY PEREVODY
(translations) (translations)
81 sont are SUT' are
82 faites done DELAEMY are maked
[doing] (made)
83 par des by machines MAŠINAMI MAŠINAMI (by
machines machines)
84 et and I and
85 leur nombre their number IH ČISLO their number
86 ira will POIDET POIDET (will go)
87 en croissant, increase VOZRASTAJA, increasing,
continuously,
88 leur their IH their
89 qualité quality KAČESTVO quality
90 s'améliorera will improve
itself BUDET will
91 ULUČŠAT'SJA be improved
92 constamment. constantly. POSTOJANNO. constant-ly.
93 Mais But NO But
94 certaines certaines
(certain) NEKOTORYE certain [some]
95 découvertes discovered OTKRYTIJA OTKRYTIJA
(discoveries) (discovery-s)
96 sont are SUT' are
97 nécessaires necessary NEOBHODIMY are necessary
98 pour que for ČTOBY that
99 cette this ETO this
100 recherche research ISSLEDOVANIE investigation
[research]
101 entre to enter VOŠLO VOŠLO (enter)
102 bientôt soon
103 dans v into [in, to]
140 MACHINE TRANSLATION
I II III IV
Author's Original Brown's Trans- Corbe's Transla- Swanson's
French lation, I.B.M. tion, manual Translation,
704 . I.B.M. 704
104 une a ODNU one [alone]
105 nouvelle new NOVUJU new
106 phase, phase, FAZU, phase,
107 celle that ETU this
108 de l'automatisa- of the automati- AVTOMATIZACII AVTOMATIZACII
tion sation (automation)
109 à at NA on
110 98 98 98 98
111 ou or ILI or
112 99 99 99 99
113 pour cent. pct. PROC. PROC. (per cent)
114 Le lecteur The reader ČITATEL' The reader
115 qui which KOTORYJ which
116 souhaite souhaite (wishes) ŽELAET desires
117 se tenir hold DERŽAT'SJA DERŽAT'SJA (to
keep abreast)
118 au at v in [into, to] (to
keep abreast)
119 courant the current KURSE KURSE (to keep
abreast)
120 et and I and [also]
121 suivre follow SLEDOVAT'ZA to follow during
122 cette this ETIM these [this]
123 future futide (future) BUDUŠČIM future [willing,
will be]
124 étape step ETAPOM stage
125 du progrès of the advance NAUČNOGO of the scientific
126 scientifique scientific PROGRESSA PROGRESSA
(progress),
127 trouvera find NAIDET will find
128 ici here ZDES' here
129 une a ODNO one [alone]
130 mise au point put at the point UTOČNENIE refinement
131 de l'état of the state AKTUAL'NOGO of the actual
132 actuel present SOSTOJANIJA state [condition,
position]
133 des travaux of the work RABOT of works
[papers],
134 tel qu' such that TAKOGO such
135 il it KOTOROE which
136 ressort be evident VYTEKAET flows out
[follows]
137 des from the IZ from [of]
138 ouvrages work TRUDOV works [treatise,
difficulty]
139 et and I and
140 articles article STATEJ articles
 APPENDIX 141
I II III IV
Author's Original Brown's Trans- Corbe's Transla- Swanson's
French lation, I.B.M. tion, manual Translation,
704 I.B.M. 704
141 parus paru (published) POJAVIVŠIHSJA of appearing
142 depuis since S from [with]
143 1955 1955 1955 1955
144 aux at the v in [into, to]
145 Etats state SOEDINENNYH connected
(United States)
146 Unis, unite, ŠTATAH ŠTATAH (United
(United States) States)
147 en in V in [into, to]
148 Grande- large-britain VELIKOBRITANII VELIKOBRITANII
Bretagne (Great Britain) (Great Britain)
149 et and I and
150 dans in v in
151 l'Union the union SOVETSKOM by SOVETSK-
152 soviétique sovietic. (Soviet SOJUZE. SOJUZ. (Soviet
Union) Union)
 Index
A IKEN , PROF . HOWARD H . 19 Cary, Edmond 125
Ampère, J.-J. 130 Chamoun, Camille 4
Andreev, N. D. 43, 48ff, 52-3, Cherry, Colin 125
66, 92-3, 96, 125 Chomsky, Noam 30, 67
Appel, Paul 71, 87 Cleave, J. P. vi, 31,47,113,125
Ashby, William Ross 116 Comit 50
Communications Research Cen-
tre, University College,
BABBAGE, CHARLES 19, 21 London 125
Bar Hillel, Yehoshua 29, 67, 104 Comte, Auguste 98
Belevitch, Vitold 125 Corbe, Michael 120-1, 135
Bel'skaja, I. K. 8,31,69-70,75,
88, 96, 102, 104, 122, 125
BESM computer 19, 30, 54 DEAD SEA SCROLLS 112
Birkbeck College, London 31 Delavenay, E. K. M. 125
Blair, Hughe 47 Dickens, Charles 70
Bloomfield, Leonard 74, 77 Dostert, Léon E. 29, 38ff, 104
Borel, Emile 71, 87 Dreyfus Graf 6
Booth, A. D. vi, 5,27-8,31,34-5,
40-1, 44, 47-8, 65, 67, 86,
104, 113, 115-6, 125-6 EDSAC COMPUTER 19
Booth, K. H. V 125 ENIAC computer 19
Brandwood, L. vi, 31,47,67,71, Estoup,J.B. 93,113
113,115-6,125-6
Brown, A. F. R. 30, 119ff, 135
Bull, Compagnie des Machines FLAUBERT, GUSTAVE 89
19 Ferranti Ltd 19
Busa, Father R. 112-3 Fletcher, Stuart L. 29, 67
Fries, Charles C. 32, 42, 74, 77
126
CALIFORNIA,INSTITUTE OF
TECHNOLOGY 31
California, University of 31 GAMMA 60 COMPUTER 19-21,
Cambridge Language Research 83
Unit 31, 95ff Garvin, Paul 29, 38
 INDEX 143

Georgetown, University of 38ff, Koutsoudas, A. 30

43, 119 Kulagina, O. S. 31, 71, 74-5,
Ghyka, Matila 111 87, 122, 126
Gilliéron 102
Gode, Alexander 47
Goethe 102 LEBEDEV, SERGEJ A. 42
Guilbaud, G.-Th. 8, 126, 130 Leningrad State University 31-2
Leningrad State University, Ex-
perimental Laboratory of
HARRIS, ZIEGFELD S. 74 Machine Translation 43
Harper, Kenneth E. 31, 96, 122, Leo computer 19
126 Liebesuy, F. 126
Harvard University 30, 82, 96, Littré 94
104, 122 Ljapunov, A. A. 32, 43
Hays, David G. 31, 96, 122, 126 Locke, William N. vi, 30, 126
Hill, W. Ryland 30
Hugo, Victor 101
MAISON LYONS 19
Mallarmé, Stéphane 111
ILLINOIS INSTITUTE OF Mark I computer 19
TECHNOLOGY 67 Mark III computer 19
Institute of Linguistics of the Massachusetts Institute of Tech-
USSR 43 nology (MIT) 29-30,36,44,
International Business Machines 50, 126
(IBM) 19-20, 31, 38ff, 43, Masterman, Margaret 31, 95ff
104, 119 Melčuk, I. A. 31 71, 74-5, 87,
IBM701 29, 39 126
IBM704 20-1, 83, 135 Mercury computer 19
International Standards Organ- Michigan, University of 30
ization (150) vi Micklesen, Lew R. 30, 82
International Telemeter Cor- Milne, numerical solution of
poration, Los Angeles 21, differential equations 92
30 Mološnaja, K. T. 32, 78-9, 126
Mounin, Georges 46,126
Muhin, I. S. vi, 59, 126
JACQUARD 19, 21 Murphy, John S. 126
Jespersen, Otto 32, 42, 77

NATIONAL SCIENCE FOUN-

KING, GILBERT W. 21, 31, 82 DATION, WASHINGTON,
Korfhage, R. 30 D.C. 31
Korolev, L. N. 31 Nuffield Foundation 31, 40
144 MACHINE TRANSLATION

OSWALD, V. A. 29, 67 Shannon, Claude E. 29

Oettinger, Anthony G. 30, 82, STRELA computer 42
96, 104, 122 Swanson, Don 120-1, 135

PANOV, PROF.D.J.vi, 9, 30-1, TROJANSKIJ, P. P. SMIRNOV

41-4, 48, 58-9, 64-5, 75, 27
88, 96, 102, 104, 116, 122,
126-7
Pascal Biaise 12,130 UNESCO 119, 127
Peano, J. 47 United Nations, General As-
Pegasus computer 19 sembly 4
Pennsylvania, University of 19 Univac Computer 19, 82
Perrot, Nicolas (Unfaithful USSR Academy of Sciences vi,
Beauties) 2 31,40,64,70,91,126-7
Plato, Dialogues 113 USSR Institute of Automation
Princeton Institute for Advanced and Telemechanics 27
Study USSR Institute of Precise
Mechanics and Computer
Technology 30-1, 41, 43, 54
RAMO-WOOLDRIDGE COR- USSR Steklov Mathematical In-
PORATION, LOS ANGELES statute 31-2, 42,75
31,111,135
Rand Corporation, Santa Mon-
ica, Cal. 31, 96
Rasumovskij, S. N. 31 WALL, ROBERT E. 30
Reichenbach, Hans 28 Watts, James 10, 130
Reifler, Ernwin vi, 28, 35ff, 41, Washington State University,
44,67,75,82,93,94,96 Weaver,Warren 5,8,27-28
Remington Rand 19 Wiener, Norbert 130
Richens, R. H. 28, 31, 34-5, 44, Whatmough, Joshua v
67, 104 Wundheiler, Alex and Luitgard,
Rockefeller Foundation 5, 27, N.67
29
Roget, P. M. 95
YNGVE, VICTOR H. 30, 67ff,
76
ST THOMAS AQUINAS 112
Sestier, A. 127
Shakespeare 112 ZARECHNAK, M. M. 30
Sheridan, Peter 29 Zipf, J. K. 68, 93,113,132