Machine Translation Again? 
Yorick Wilks, Jaime Carbonen, David Farwell, Eduard Hovy and 
Sergei Nirenburg 
Department of Computer Science 
New Mexico State University 
Las Cruces, NM 88001 
Machine translation (MT) remains the paradigm task for 
natural language processing (NLP) since its inception in 
the 1950s. Unless NIP can succeed with the central task 
of machine translation, it cannot be considered successful 
as a field. We maintain that the most profitable approach 
to MT at the present time is an interlingual and modular 
one. MT is one the precious few computational tasks fal- 
ling broadly within artificial intelligence (AI) that combine 
a fundamental intellectual research challenge with enor- 
mous proven need. To establish the latter, one only has to 
note that in Japan alone the current MT requirement is for 
20 billion pages a year (a market of some $66 billion a year). 
The vulgarized version of the history of MT is as fol- 
lows: In the 1950s and 1960s large funds were made avail- 
able to US MT which proved to be an umitigated failure. 
The ALPAC report (1966) said MT was impossible and 
doomed all further US funding. MT work then moved to 
Canada and Europe where it partly succeed, which was 
then followed by highly successful exploitation in Japan. 
The truth, of course, is not at all like that. 
MT work did not stop in the US after ALPAC: the 
AFOSR continued to fund it in the US and there were and 
are enormous commercial developments (the best known 
systems being SYS'rRAN, ALPS, LOGOS, METAL and 
SMART). 
ALPAC did not say MT was impossible nor that the 
work done was no good: only that at that point history, 
with the cost and power of 1960s computers, human trans- 
lation was arguably cheaper. 
MT work did not really move to Europe, since it 
stopped there also in response to the ALPAC report. The 
UK believed the ALPAC report, and only in France did 
serious work continue, and the GETA system in Grenoble 
became the foundation for a range of others, including the 
major Japanese university system (Mu) and aspects of the 
Eurotra system, which was designed to be a multilingual 
system between the languages of the EEC. 
The GETA system, like SYSTRAN, date their origins 
from the very earliest period of pre-ALPAC MT. The 
longevity of such systems is proof of the need of stamina 
and persistence in MT to achieve serious results, but also 
the need for periodic redesign, pretty much from scratch, 
since old formalisms and software reach a point where they 
cannot be further optimized, a point reached long ago with 
SYSTRAN itself. One way in which all MT work is in 
SYSTRAN's debt is that it is the main existence proof: it 
convinces doubters that there that machine translation now 
exists, albeit in primitive form, and can be purchased on a 
large scale and at a quality that many users find acceptable 
for their needs. A key defect in the ALPAC report was that 
it underestimated how large a market there was for par- 
tially accurate, low quality MT, and SYS'rRAN filled that 
market. The point now, of course, is to move on to the 
huge market for higher-quality MT. But even now the pro- 
portion of internal EEC documentation translation for 
which a preliminary draft version is done by machine is 
rapidly growing using a modified version of the SYSTRAN 
system. 
It is certainly not the case that most major MT installa- 
tions in the world are now Japanese. In the list given in the 
JEIDA report only one Japanese system occurs among the 
list of major installed systems in the world outside Japan. 
All the rest are American. However, that list is becoming 
quickly dated, as Japanese system are being researched, 
developed and deployed at a much faster rate, reflecting a 
lopsided ten-to-one total R funding skew in favor of Japan 
over America. Moreover, some commercial American MT 
efforts are being purchased by the Japanese; witness 
BRAVIS's purchase of WEIDNER COMMUNICATIONS, 
and a (partial) purchase of SYSTRAN. A crucial differ- 
ence between US and foreign strategies to date has been 
that the Japanese government made machine translation 
central to the Fifth Generation effort, and the European 
Community began ten years ago a $45 million investment 
in the Eurotra project as part of their overall information 
technology drive. 
Why this is a good time to get back into MT. 
There is a growing need for translation in intelligence, 
commerce, science, government, and international organi- 
zations. This is due to factors such as the following: 
• Increases in international cooperation and competition, 
which involve an ever-growing volume of text to be 
communicated. 
• World-wide electronic networks have made international 
communication much easier. 
• Reports, documentation, legal papers, and manuals are 
increasingly produced in one culture and exported to 
various other cultures, often in multiple languages. 
• More emphasis is being placed on the use of national 
languages in documents and systems. 
• The economic rise of South-East Asia and the opening 
of the European market in 1992 add significantly to 
these factors. 
371 
Strategic Reasons for an MT Effort: 
MT systems live and decay like natural organisms: they 
have natural life spans that cannot be indefinitely 
prolonged. The SYSTRAN system has lived long and 
done well but it is 30 years old and cannot be optimized 
above the 75% level. Later systems from the early 1970s 
(GETA, LOGOS, ALPS, WEIDNER, MU, etc) were better 
constructed but cannot rise above their current levels---the 
evidence for this being that the two research systems in 
that list (GETA & MU) have now effectively collapsed and 
their teams dispersed. The right thing is now to promote a 
new design using the enormous and transferable advances 
that have been made in interfaces, hardware, linguistics, 
AI, machine lexicons etc. 
The most recent new large-scale efforts have either been 
badly managed and proved impractical (like EUROTRA) 
or set very mow commercial goals (usually involving 
only Japanese and English or Asian languages) like the 
major Japanese systems. 
The need has never been greater, not only for MT itself 
but all the associated technologies that can be integrated 
into a well-designed MT system. The rapid integration of 
the world is not leading to a common language (English) 
nearly as fast as it is leading to the absolute need to read 
masses of documentation produced in foreign languages. It 
is also necessary to choose a system to which NEW 
languages can be rapidly added, i.e. a modular, interlingual 
one. 
Much of the MT-related research performed in the US is 
being applied elsewhere. No nationwide project utilizing 
the best research talents in NLP has been attempted in the 
U.S. in over two decades. Today, Darpa is probably the 
only institution with the resources and scope to mount a 
large-scale MT effort successfully. Such an effort would 
harness and coordinate NLP work of various kinds and 
would create a setting in which new innovations could be 
used within this country first. 
A second strategic reason pertains to interproject 
collaborations. Currently, there is relatively little 
collaboration and sharing of resources and expertise among 
NLP research groups in this country. A new national 
agenda with a set of clearly focused goals could serve as an 
integrating agent. The development of a standard 
interlingua representation, a set of standardized lexicons, 
one or more grammars, support tools and interfaces, and 
additional software, can shape much future NLP research 
in this country by enabling researchers to make use of 
existing work and tools with much less effort than is 
currently the case. 
Technical Reasons for an MT Effort: 
Steady developments in various aspects of NLP make 
available large portions of an MT system more or less off 
the shelf, which greatly facilitates the construction of new 
MT systems. These developments are the following: 
1. Clearer understanding of semantics: Recent 
refinements of taxonomical ontologies of 
representation provide an interlingua-like basis for a 
new, more powerful, MT. Making maximal use of 
the high-level linguistic and semantic generalizations 
shared among languages, one can minimize 
language-to-language lexical or structural transfer 
rules and so increase the portability of the system 
across domains. 
2. More complete grammars: Development of 
grammars is an ongoing process. There exist today 
grammars that cover English (and other languages 
such as German, Chinese, Japanese, and French) far 
more extensively than the most comprehensive 
grammars of 20 years ago did. 
3. Better existing generation and parsing technology: 
Single-sentence parsing and generation has been 
studied to the point where a number of well- 
established paradigms and algorithms exist, each 
with known strengths and weaknesses, a situation 
which greatly facilitates the construction of a new 
MT system (in fact, in the last 5 years a number of 
general-purpose generators have been distributed: 
Penman, mumble, frege, etc.). 
4. In addition, the number of existing MT systems and 
the amount of MT experience is also much larger 
than it was in the early days, especially in Europe 
and Japan. 
An Interlingual Approach Versus Transfer Or Massive 
Statistics. 
A fundamental technical notion in our proposal is 
interlinguality: it is one of the three basic structural 
methods for MT, contrasted with direct and transfer 
approaches. The direct method was used for early systems 
like SYSTRAN as well as large recent ones like SHALT 
from IBM Japan. If one is only every going to be interested 
in one language couple in one direction, as SYSTRAN 
originally was, there is no reason not to use it. We assume, 
however, that that is not our situation and muldlinguality is 
essential. It should also be noted that some form of 
intedinguality is now becoming the standard position in 
AI-knowledge representation and our approach meshes best 
with that. The interlingua approach overcomes the 
problem of building thousands of transfer rules by using a 
central representation into which and from which all the 
languages are parsed and generated. 
Of major concern is to design an intedingua which is 
both specific enough to allow simple and unambiguous 
processing and general enough to enable different 
approaches with different theoretical strengths to represent 
the information they can extract from the text. Fortunately, 
none of the parties involved have ever been committed to 
the highly formalized representation languages and systems 
which have been (and still are) popular in various areas of 
NLP, formalisms whose logical properties have been 
studied extensively but whose practical utility is low. 
372 
Consider now the following example: 
"Mary was in a severe accident. She lost a foot." 
vs. "Mary was buying cloth, but measured it incorrectly by 
accident. She lost a foot." 
There is no statistical measure (e.g., no low-order n- 
grams) that will disambiguate reliably. Yet, if a sentence 
similar to the above concerned the Lybian Colonel or Abu 
Nidal it might be useful to have accurate intelligence. 
Language other than English have different ambiguities 
that must be resolved to translate to English or to fill a 
database for an analyst. 
The interlingua approach is far better able to exploit 
domain knowledge in order to produce reliable translations 
than the other two approaches. The massive statistical 
approach is inimical to any infusion of domain knowledge 
or any comprehension of the language. Pure statistical 
translation had been rejected in the early years, but has 
been brought back to life in the recent IBM re~arch effort. 
Experience has consistently shown that unaided statistical 
methods perform only at a low level which cannot be 
raised much, and only on a carefully selected materials (in 
the IBM project based on the copious high-quality parallel 
French-English Hansard texts from Canada -- data not 
found for other language pair. Even the 50 success 
claimed may depend crucially on order similarities between 
English and French. The paper claims that for 63 of 
tested sentences under 10 words, the most probable word 
order, based on trigram probabilities, was the correct one 
80 of the time, which together produce the figure above. 
The transfer approach is indeed capable of using domain 
knowledge, but the software engineering is much worse 
than an interlingual approach. If one must translate among 
N languages, there are N(N-1)/2 language pairs. A transfer 
approach would require on the order of N**2 transfer 
grammars (2 per language pair, one for each direction), if 
these must be augmented with domain semantics, a task 
that was gargantuan to start becomes totally intractable to 
hardiest of souls. In contrast, the interlingua approach 
requires 2N grammars (1 for analysis and 1 for generation 
for each language, into and out of the standardized 
common interlingual knowledge representation). Domain 
knowledge, though potentially complex, need be added 
only once per domain and retained in modular, reusable 
declarative data structures that serve as input to a unifying 
compiler. This compiler combines modular domain 
knowledge and grammar files dynamically to produce a run 
time translator among two languages for a given domain 
(or set of domains). 
Statistics, although not the preferred translation 
paradigm, plays several important roles in MT, including: 
Once the meaning of a text is analyzed, selecting the most 
normative (frequent) rendition into words in each target 
language. Statistics can select collocations from large text 
corpora (such as the preferred use of "pitch black" rather 
than "asphalt black"). Given a large potential lexicon, 
simple frequency analysis can direct the dictionary-building 
work towards the most frequent words first, so as to obtain 
maximal utility of a system during development phases. 
All evaluation metrics of fluency, accuracy and cost of 
translation are statsfically based. 
Machine translation systems must be concerned with the 
knowledge encoding, with modular software architectures, 
with good engineering, with scalable and evaluable systems 
development, much more so than with specific finguisfic 
theories prevalent in modern transfer approaches. In 
practice, MT approaches motivated by theoretical-linguistic 
concerns, like EUROTRA, tend to be too driven by 
finguistic fashion (since their chief motivation is to be 
theoretically interesting rather than effective). This opinion 
is shared by the Japanese researchers. Thus, the 1989 
JEIDA report concluded that linguistic theory had made no 
discernible contribution to the advance of MT. Key 
features of the cooperative approach we advocate are: 
1. The use of an interfingua instead of transfer rules or 
statistical cooccurrences; 
2. Modularity : both programs and data will be 
produced in a modular fashion allowing them to be 
assembled into a number of prototype MT systems; 
3. Commitment to gradual increase in the levels of 
automation of the systems we create; 
4. The central role of world knowledge in addition to 
knowledge about language; 
5. The use of a representation based on commonsense 
semantics and pragmatics ; 
6. Emphasis on the large scale of the systems under 
construction; 
7. Ensuring portability across domains by building 
reusable tools and information repositories such as 
lexicons; 
8.Developing a translator's workstation environment 
a) to facilitate the integration of the above modules 
and b) to support the creation of useful machine- 
aided translation systems at the earlier stages of the 
project, while the various automatic processing 
modules are being developed. Included here will be a 
separate, but compatible, lexicology workstation, to 
assist the incorporation of large-scale semantic, 
syntactic and collocational information from machine 
-readable dictionaries and text corpora. 
The Modularity Assumption 
Modularity is independent of interlinguality though 
opting for the latter requires the former. Strong modularity 
of language components would now be supported by most 
researchers and developers in MT, largely because it allows 
the addition of new languages with minimum dislocation. 
It is also essential if it is to be possible to treat different 
languages by different methods and to combine work at a 
range of sites. Agreeing on suitable interfaces is a practical 
not a theoretical matter, and the experience of EUROTRA 
has shown it is perfectly feasible (this is the main scientific 
contribution of EUROTRA). 
373 
In order to harness the NLP research potential in this 
country, a modular approach to the construction of 
prototype MT systsems is proposed. Under this approach, 
various sites will build various modules which can be 
assembled in various ways to construct various prototype 
systems. 
Two advantages of the modular approach are: new 
languages and additional functionalities such as gisting can 
be added with minimal disruption to the existing system, 
and the system can support various theoretical approaches 
(which may be required by various languages, or which 
may be the best way to foster collaborations among groups 
with different research methodologies). 
MT system modules are either theory-neutral or theory- 
based. Theory-neutral modules are typically receptacles of 
basic information, such as core lexicons, morphological 
information, models of the application domain, domain 
lexicons, etc. Theory-based modules are modules whose 
construction and performance depends on a particular 
theoretical approach (of Linguistics, semantics, etc.); 
typical instances are parsers, generators, and theory-based 
grammars. 
In order to limit redundancy, this proposal calls for the 
straightforward incorporation of existing theory-neutral 
modules from any available source. Various lexicons and 
some theory-neutral grammars of several languages exist in 
the public domain. Only when such information is 
unavailable, or when the available information is not 
structured in a useful way, should a new module be 
constructed. In such cases, the proposal calls for the 
construction of a single module, to be shared by all 
participants in the MT program. 
Some modules, however, must be structured to conform 
to the requirements of a particular theoretical approach. In 
order to allow various approaches to participate (and be 
tested) in the MT program, the proposal calls for the 
parallel construction of various theory-based modules that 
perform the same function. Different modules will be 
constructed at different sites, but the enforcement of an 
intermodule communication protocol will ensure that the 
modules are mutually replaceable. 
To summarize the modularity issue, we propose to 
enforce standard interfaces, modular development and 
maintenance in the following dimensions: 
• Modular knowledge bases 
• Common "top" of ontology across all domains 
Combinable "subworld" ontologies for specific domains 
No language-specific info in knowledge bases, only 
domain info. 
• Modular unification-grammar files 
• High-level well-structured grammars for each language 
No domain-specific info in any language specification 
• Unifying grammar compiler (or interpreter) 
374 
• Takes language, domain and dictionary to produce run- 
time working MT system No human needs to cope with 
the output object code of unifying compiler, in the same 
way that no one needs to look at output of ADA 
compiler once verified. 
The advantages of this modular approach include the 
following: 
1.Various projects and various theoretical approaches 
will be able to participate. 
2.Projects need not have experience in all aspects of 
MT to participate. 
3.Redundant development of modules will be 
eliminated. 
4.Interproject collaboration will be stimulated 
throughout the U.S. 
5.The common goal of translation will provide a more 
coherent focus for the various research endeavors and 
will facilitate the comparison of various approaches 
to tease out their slrengths and weaknesses. 
6.As new and promising projects are found, they can 
be included into the program. 
7.The theory-neutral modules, all in a standard form, 
will be made available to the whole NLP community 
as a basic resource. 
8.Large-scale lexicons, automatically constructed from 
text, can be used in parsing and generation and in 
interactive help. 
9.Existing work on collocation, cooccurrence, and 
clustering of words and phrases can be put to use 
(for example, to guide lexicon construction). 
The attached "mountain" figure shows a possible 
cooperation being established between the Center for 
Machine Translation (Carnegie Mellon University), the 
Computing Research Laboratory (New Mexico State 
University) and the Information Sciences Institute 
(University of Southern California), whose groups share 
both experience of MT and the above assumptions. 
Figure- mountain shows the anticipated modules, starting 
from the bottom left-hand comer upward (parsing) going 
up, and then down again to the bottom right (generation). 
The approximate number of modules and the sites with 
expertise in them are indicated, with the sites responsible 
for the module in larger font. Boxes in the middle represent 
the tasks of knowledge acquisition and system integration, 
also annotated with ressponsible sites. 
Gradual Improvement 
It is important to note that not all modules will be 
required for the MAT system to run. A number of the 
more experimental aspects can be "short-circuited", 
resulting in a leaner representation of the input text (and 
weaker output, or correspondingly more work for the 
augmentor or post-editor). We plan to adopt the policy of 
first creating a machine-aided translation system and then 
gradually enhance the levels of automation in the 
subsequent versions by implementing new descriptive 
In order to harness the NLP research potential in this 
country, a modular approach to the construction of 
prototype MT systems is proposed. Under this approach, 
various sites will build various modules which can be 
assembled in various ways to construct various prototype 
systems. 
Two advantages of the modular approach are: new 
languages and additional functionalities such as gisting can 
be added with minimal disruption to the existing system, 
and the system can support various theoretical approaches 
(which may be required by various languages, or which 
may be the best way to foster collaborations among groups 
with different research methodologies). 
MT system modules are either theory-neutral or theory- 
based. Theory-neutral modules are typically receptacles of 
basic information, such as core lexicons, morphological 
information, models of the application domain, domain 
lexicons, etc. Theory-based modules are modules whose 
construction and performance depends on a particular 
theoretical approach (of Linguistics, semantics, etc.); 
typical instances are parsers, generators, and theory-based 
grammars. 
In order to limit redundancy, this proposal calls for the 
straightforward incorporation of existing theory-neutral 
modules from any available source. Various lexicons and 
some theory-neutral grammars of several languages exist in 
the public domain. Only when such information is 
unavailable, or when the available information is not 
structured in a useful way, should a new module be 
constructed. In such cases, the proposal calls for the 
construction of a single module, to be shared by all 
participants in the MT program. 
Some modules, however, must be structured to conform 
to the requirements of a particular theoretical approach. In 
order to allow various approaches to participate (and be 
tested) in the MT program, the proposal calls for the 
parallel construction of various theory-based modules that 
perform the same function. Different modules will be 
constructed at different sites, but the enforcement of an 
intermodule communication protocol will ensure that the 
modules are mutually replaceable. 
To summarize the modularity issue, we propose to 
enforce standard interfaces, modular development and 
maintenance in the following dimensions: 
*Modular knowledge bases 
*Common "top" of ontology across all domains 
Combinable "subworld" ontologies for specific domains 
No language-specific info in knowledge bases, only 
domain info. 
*Modular unification-grammar files 
*High-level well-structured grammars for each language 
No domain-specific info in any language specification 
*Unifying grammar compiler (or interpreter) 
375 
*Takes language, domain and dictionary to produce run- 
time working MT system No human needs to cope with 
the output object code of unifying compiler, in the same 
way that no one needs to look at output of ADA 
compiler once verified. 
The advantages of this modular approach include the 
following: 
1.Various projects and various theoretical approaches 
will be able to participate. 
2.Projects need not have experience in all aspects of 
MT to participate. 
3.Redundant development of modules will be 
eliminated. 
4.Interproject collaboration will be stimulated 
throughout the U.S. 
5.The common goal of translation will provide a more 
coherent focus for the various research endeavors and 
will facilitate the comparison of various approaches 
to tease out their strengths and weaknesses. 
6.As new and promising projects are found, they can 
be included into the program. 
7.The theory-neutral modules, all in a standard form, 
will be made available to the whole NLP community 
as a basic resource. 
8.I.arge-scale lexicons, automatically constructed from 
text, can be used in parsing and generation and in 
interactive help. 
9.Existing work on collocation, cooccurrence, and 
clustering of words and phrases can be put to use 
(for example, to guide lexicon construction). 
The attached "mountain" figure shows a possible 
cooperation being established between the Center for 
Machine Translation (Carnegie Mellon University), the 
Computing Research Laboratory (New Mexico State 
University) and the Information Sciences Institute 
(University of Southern California), whose groups share 
both experience of MT and the above assumptions. 
Figure-. mountain shows the anticipated modules, starting 
from the bottom left-hand comer upward (parsing) going 
up, and then down again to the bottom right (generation). 
The approximate number of modules and the sites with 
expertise in them are indicated, with the sites responsible 
for the module in larger font. Boxes in the middle represent 
the tasks of knowledge acquisition and system integration, 
also annotated with responsible sites. 
Gradual Improvement 
It is important to note that not all modules will be 
required for the MAT system to run. A number of the 
more experimental aspects can be "short-circuited", 
resulting in a leaner representation of the input text (and 
weaker output, or correspondingly more work for the 
augmentor or post.editor). We plan to adopt the policy of 
first creating a machine-aided translation system and then 
gradually enhance the levels of automation in the 
subsequent versions by implementing new descriptive 
World Knowledge 
Ours is an AI approach in that we shall, in processing 
expressions so as to select a particular interpretation, apply 
computationally expressed knowledge of the world, as well 
as our knowledge of language. We thus select the most 
sensible interpretation of ambiguous expressions, 
recovering the most sensible referents for pronouns and 
inferring information which is implicit. This knowledge of 
the world is general in the sense that we know a great deal 
about objects, actions, states, events and situations, such as 
the classes to which they belong and the attributes they 
possess. Through the application of such knowledge, we 
weed out incoherent interpretations as they develop and 
select the most appropriate interpretation from those that 
survive. 
Commonsense Semantics and Pragmatics 
A crucial component is a realistic pragmatics, bringing 
in the best of AI work on speech act, belief etc. 
phenomena. These are now tractable and usable notions in 
MT systems. We shall commit ourselves to commonsense 
semantic approaches rather than formal ones since these 
have not proved fruitful in MT in any language. This will 
also involve a commitment to algorithmic elements of AI- 
based semantics (such as Preference Semantics) that have 
already proved useful in message-understanding work, and 
have an intimate connection with understanding of ill- 
formed, metaphor-laden text that is the normal form of 
actual documents. 
In order to build working, portable prototype systems, 
the most practical and useful notations must be. used. As 
mentioned above, the selection of notations whose 
properties are desirable on formal grounds but whose 
practical utility is low will be avoided. 
In order to produce MT of superior quality that existing 
systems, one of the most powerful key ideas is the use of 
discourse-related and pragmatic terms. Most MT systems 
operate on a sentence-by-sentence basis only; they take no 
account of the discourse structure. Given recent work on 
discourse structure at various centers in the U.S., structural 
information should be taken into account and can be used 
to improve the quality of the translation. Similarly, 
pragmatic information, such as Speech Acts, reference 
treatment, and perhaps even some stylistic notions (to the 
extent that notations have been developed to represent 
them) will be used to improve the quality of the translation. 
Scale 
We emphasize scale phenomena, both in the sense of 
bringing large-scale lexical material automatically via 
existing work on machine readable dictionaries, but also 
making use where possible of statistically-based work on 
corpora to guide lexical entry selection, corrigibility of 
sentences to particular syntax rules etc. 
Portability 
• Construct reusable tools, general information 
repositories (e.g., lexicons, grammars) 
* Establish nationwide resources in standard form 
• Ensure future reusability 
• Construct reusable tools, general information 
repositories (e.g., lexicons, grammars) 
• Establish nationwide resources in standard form 
• Ensure future reusability 
One of the well-known weaknesses of current MT 
systems is their limited applicability. In order to achieve an 
acceptable level of translation quality, the current brute- 
force approaches require large collections of translation 
rules which invariably contain increasingly domain-specific 
information. Porting these systems to a new domain 
becomes a major undertaking. 
By using the newest NLP technology while focusing on 
the development and use of a number of very general 
information resources (such as a high-level concept 
ontology under which domain-specific ontologies are 
subordinated, and a general lexicon for closed-class words), 
this proposal is aimed at overcoming the problem of 
domain-dependence without compromising on translation 
quality. 
A major factor supporting the domain-independence is 
the ability to acquire information --- conceptual, lexical, 
phrasal, translational -- interactively during the translation 
process. When the system encounters input it cannot 
handle, it queries the human assistant, who decides what 
type of information the input is and then inserts appropriate 
definitions into the system's information banks for future 
use, using the interfaces and acquisition tools provided. 
The proposed MT program devotes a ~arge amount of 
effort on the development of interactive acquisition 
software and interfaces, via the notions of the Translator's 
and Lexicologist's workstations. 
The strengths and weaknesses of this interlinguai approach 
The strengths of the interlingua approach have been 
briefly discussed above. 
The central weakness is the necessity to build a 
knowledge base, and therefore the initial development cost, 
though it can be amortized over other languages, and many 
translations in the context of a well-engineered modular 
system. 
We would like now to defend the interlingua approach 
against three most commonly held negative opinions. 
OPINION 1: An interlingual approach forces unneeded 
processing 
If a source language has, say, an expression which is three ways 
ambiguous and some target language has an expression which has 
precisely the same three-way ambiguity, unanalyzed why not simply 
carry the ambiguity from the source to the target and let the reader 
376 
figure it out? Why disarnbiguate needlessly? 
The response is, on the one hand, that a third language 
probably has different expression for each of the possible 
interpretations, so that if the same representational 
apparatus is to be applied to translations between the 
source language and a third language or from the target 
language and a third language, such processing is necessary 
in any case. On the other hand, a quick inspection of 
bilingual dictionaries shows that cases of complete 
correspondence of ambiguity across languages is extremely 
rare, even in closely related languages such as German and 
Dutch. 
The issue of "since we sometimes can get away with 
less processing, why risk doing unnecessary work?" can be 
compared with intelligence-gathering work, where much of 
the effort is routine; information often confirms 
expectations; and therefore much of the work is 
"unnecessary." With such an attitude, all unexpected, 
important intelligence would often be ignored, much to the 
detriment of the analysts and policymakers. Ignoring 
meaning in translation because it need not always be 
interpreted, is an equally flawed philosophy. The times 
when deeper analysis is required can be absolutely crucial 
to produce meaningful, rather than misleading, translations. 
language particular bias would simply be defect of the 
approach, rather than wholly invalidating it. 
A standard example here would be the case of the verb 
"wear" in English and the problem of expressing the 
notion in Japanese or Chinese. It so happens that in 
Japanese the corresponding verb depends entirely on what 
is worn e.g. shoes (verb= hateiru ), coat (verb= kiteiru ), 
spectacles (verb= kaketeiru ) and so on (and similarly for 
Chinese). It is thus reasonable to say that Japanese does not 
have a concept of "wear" in the way English does. 
However, that observation is no kind of argument at all 
against an interlingual approach, merely one for intelligent 
generation. In an interlingual environment there will be at 
least one interlingual node (which may or may not 
correspond to "wear") that links the relevant sense 
representations. The crucial point is that it would be the 
intelligent Japanese generator (since no problem arises in 
the Japanese to English direction) that makes the choice of 
output verb based simply on selection semantics (e.g. if the 
worn object is "koutoo" the verb is "kiteiru" and so on). 
Conclusion 
There are several aspects of the knowledge-based 
interlingua MT project that incur a measure of risk. 
Foremost among these is the distributed management risk 
OPINION 2: Interlingual approaches are heavily knowledgeamong the three centers. 
dependent and the task of working out appropriate 
representations is too demanding to be practical. 
It has been our experience that some, even if 
incomplete, level of knowledge representation is crucial to 
machine translation. The need for such knowledge is 
especially obvious in the translation of technical text, 
where translations based on a general knowledge of the 
world are markedly inferior to translations based on a 
specific knowledge of the subject domain of the translation. 
Large-scale know.ledge bases are being actively developed 
in the field, and domain models have come to be 
considered a standard component of many AI-related 
application systems. In our system, acquisition of world 
knowledge will be an ongoing task, and we fully intend to 
prove the feasibility of working with large knowledge 
bases in practical terms. To wit, EDR laboratories, and 
Fujitsu in Japan have come to the same conclusion and are 
actively building large knowledge-bases for interlingua- 
based machine translation, with initial success. 
Although it is clearly in the 
national interest to establish several sites developing MT 
technology in mutual cooperation, special effort must be 
made to address communication, establishment of 
standards, mutual responsibility relationships, fall-back 
positions and so on. We think this is an eminently 
manageable risk, but nonetheless a omnipresent one. We 
are fully cognizant of this risk, and are prepared to 
minimize it by establishing common procedures, open lines 
of communication, and accommodation where necessary. 
OPINION 3: interlingual approaches are based on a 
particular language, thus creating unnatural analyses 
for other languages. 
This is the "cultural imperialism" argument. If, 
however, there exists such a thing as a universal descriptive 
linguistic framework, then there is no reason to assume that 
language imperialism must be a side-effect of the 
interlingual approach. Our experience in the development 
of an interlingual representation based on a cross-linguistic 
comparison of parallel texts has indicated, at least, that 
such a language independent framework is possible. But 
even if no such framework exists, then at worst, such 
377 
\ 
r-'- 
'9..- -~ 
.o ,n 
F~gm'e 1: MAT Sys~m ModuMs. 
378 
