MT translation
began in 1953, seventeen years after Konrad Zuse, a construction engineer for
the Henschel Aircraft Company in Berlin, developed a series of automatic
calculators to help him with complex engineering calculations that earned him
the semiofficial title of “inventor of the modern computer.” The first MT
systems were Direct Systems. They simply translated the Source Language (SL)
text to the Target (TL) directly using a bilingual dictionary as reference.
Think of the hand-held travel dictionary that converts phrases that travelers
use most frequently. Direct MT systems do not analyze or disambiguate the SL
text. They translate only simple text, or text with a very low level of
ambiguity (some are capable of finding entries for past participles, gerunds,
noun plural forms, and adjective forms). If a SL word has more than one meaning,
this approach could produce the wrong results in the TL. For example the
translation of the “I am fine” response to “How are you?” could
conceivably become “I am a traffic ticket” in the TL. Mistakes such as
this will amuse your hosts in the visiting country, who will still appreciate
your effort to communicate with them in their language, but they are
unacceptable in a formal translation.Since the 1950’s computing technology has taken giant leaps, and
MT along with it. These advances enabled the development of Indirect MT systems.
These systems begin the translation process by analyzing the source text to
disambiguate the SL sentences and enhance the quality of the translation.
In 1992, Hutchins and Somers (An introduction to machine translation, Academic
Press, London) identified two types of MT systems: Transfer-based and
Interlingual.
In “The development and use of machine translation systems and computer-based
translation tools”, John Hutchins explains the differences between these to
MT systems:
“There have in fact been two basic ‘indirect’ approaches. In one the abstract
representation is designed to be a kind of language-independent ‘interlingua’,
which can potentially serve as an intermediary between a large number of natural
languages. Translation is therefore in two basic stages: from the source
language into the interlingua, and from the interlingua into the target
language. In the other indirect approach (in fact, more common approach) the
representation is converted first into an equivalent representation for the
target language. Thus there are three basic stages: analysis of the input text
into an abstract source representation, transfer to an abstract target
representation, and generation into the output language.”
Transfer Based Indirect MT System
A Transfer Based Indirect MT System includes analysis and
parsing of the source language text, independently of the TL text. The results
of this analysis are used during the transfer to TL process to find the
corresponding words in the TL.
Interlingual MT Systems
The Interlingual system adds two additional steps to the
process: Translation to an intermediate language (interlingua) before
transferring the SL text to the TL.
According to Hutchins & Somers (1992), an interlingua is the intermediate
representation of meaning that “includes all information necessary for the
generation of the target text without `looking back' to the original text. The
representation is thus a projection from the source text and at the same time
acts as the basis for the generation of the target text; it is an abstract
representation of the target text as well as a representation of the source
text.”
Some researchers used an artificial language, like Esperanto as the interlingua
because artificial languages are considered to be more regular and consistent in
their morphology and syntax.
Commercially available MT systems fit the three basic system types (‘transfer’,
‘direct’ and ‘interlingual’). Some best known of the systems in the industry,
such as Systran, Logos and the Fujitsu Atlas systems, are based on 'direct
translation', but they are vastly improved versions of their predecessors. They
are highly modular, easily modifiable and extendable. Systran powers the popular
Babel Fish Translation that is available in the Internet. Systran was originally
developed in the 1970 to translate Russian to English. Today, it offers a large
number of language pairs, including most European as well as some Asian
languages. The Commission of European Communities bought an English-French
version of this system even though the quality of the translation. After a
considerable investment in time and effort by the Commission's evaluators and
lexical coding specialists, the quality of the translation had improved enough
to do post level editing in multiple language pairs.
There have been also a number of ‘transfer' systems. One of these is METAL,
which was supported by Siemens, Germany during the 1980’s. The system became
available commercially in the 1990’s but sales were disappointing, and Siemens
transferred the rights to METAL to GMS and LANT. The most famous systems based
on the 'transfer' process were Ariane at GETA in Grenoble, France - a project
that began in the 1960's - and Eurotra - a project funded by the Commission of
the European Communities. Neither system met the desired expectations. In the
late 1980's Japanese Government agencies, in cooperation with researchers from
China, Thailand, Malaysia and Indonesia, sponsored an 'interlingua' system for
Asian languages. After more than a decade of effort, this system has not
produced the desired results. (For surveys of MT research and development in
1980s and early 1990s see Hutchins 1993, 1994.)
Our own experience with ‘untrained’ MT systems is that translators prefer to
start ‘from scratch’ rather post-edit an automated translation. However, even
‘untrained’ MT systems are helpful in many ways. For example, Alta Vista Babel
Fish makes it possible to get the idea of what a website in a language we do not
know is about - we call this 'gisting'.
Computer Assisted Translation (CAT)
Reverse the letters that make up the MT acronym and you get TM, (Translation
Memory), a CAT system that provides a middle ground between MT and unassisted
translation. TM is embraced by most translators and translation services. For
more on TM systems, please refer to http://www.intersolinc.com/terminology_management.htm
The basic difference between MT and TM systems is that TM systems are intended
as an aid to the translator, they are not automated translation systems.
Essentially, they allow translators to maintain a collection of terms that they
have previously translated so they can re-use them. Translators can then post-edit, or modify their own translations so they become a 100% match with a
modified or new version of the SL text. This process is called “fuzzy matching”.
The following figure illustrates this process:
In this English to Spanish translation, using the TRADOS TM system, the phrase
on the top row, (with the MS Word icon), is the new text to be translated. The
phrase in the bottom row indicated by the US flag (for English –US version) is
the SL phrase stored in the TM database. The SL (Spanish) text directly
underneath is the "fuzzy match" phrase that the translator will have to edit to
get to a 100% match with the new SL text.
TM is also been used as the ‘front-end' system for MT systems. MT systems can be
‘trained’ using phrases taken from TM databases.
In conclusion, TM and MT systems are only as good as the data that is fed into
them and in both cases these are words and phrases originally produced by human
translators. As computing technology continues to advance, the synergy between
the computer and the human translator will increase and the results will be
mutually beneficial.
References
http://inventors.about.com/library/blcoindex.htm
http://www.foreignword.com/Technology/art/Hutchins/hutchins99_2.htm#hist
Different Approaches to Machine Translation, by Jaap Van Der Meer published in
MultiLingual Computing Technology, #71, volume 16 Issue 3
Print this page