|
Introduction
|
|
| This is the oldest application of language technology, dating back to first experiments in Russian-English translation in the 1950s. It has had a chequered history, often attracting considerable government funding, but just as often blighted by discouraging reports about its long-term prospects. Nevertheless systems which have resulted from development over this long period are widely and intensively used in international organizations (such as the European Commission, the United Nations and the Panamerican Health Organization), and are now often freely available to casual users.
Machine translation is generally agreed to be best applied to generate rough drafts of translated documents. As such, it can assist translators to speed production; and it can also aid users directly when all that is required is a rough indication of the content of documents in unknown languages. It is currently in vogue, therefore, to provide access to web pages.
With the advent of memory-based translation, a new aid to translators has become available: they can store, and conveniently re-use, translations of words, phrases and paragraphs from previous work. This effects a massive saving of time, and gain in consistency, over the long term.
|
|
Where the Progress is Being Made
|
|
There are a number of approaches to machine translation. They are often presented as competitors, but probably all of them will make contributions to different aspects of translation performance.
- Grammar-based machine translation. This is the traditional approach, involving dictionary look-up and syntactical analysis. In the 1980s, it was pursued Europe-wide in the EUROTRA project, and the Japanese national programmes MU and EDR . It is capable of continuing gradual improvement, which is going on at the same centres.
- Knowledge-based machine translation. This is an extension of the grammar-based approach, but emphasizes the building of a coherent conceptual structure, or ontology, which helps to interpret a text, and so give a basis for choosing among translation-equivalents. Noted centres for this are CMU and NMSU in the USA.
- Statistically-based machine translation. This approach is radically different, by-passing any explicit analysis of grammar, and using pre-existing bilingual texts to predict, on the basis of probability, what would be an appropriate translation for a given word in a given context. The approach has been pioneered in the USA, especially at IBM .
- Terminology. Technical texts are an important area where translation is often necessary. However, they are often written in specialized language which is not available in bilingual dictionaries. (Automatic processes have been developed to identify these terms. See the University of Surrey . ) It therefore becomes necessary, and cost-efficient, to compile special lists of terms and their equivalents in foreign languages. These lists are typically much longer than general-language dictionaries.
- Memory-based machine translation. This approach, like the statistically-based, relies on records of previous translation, but is deterministic, providing automatic access to translation equivalents which have already been selected in a translator's past work. It was first developed in Kyoto University , but has since been take up as an approach within Europe, notably at The Trados Company..
|
|
Sources for Products
|
