Machine Translation

Formal evaluations of MT systems are particularly difficult to conduct, since there is no agreed-upon "best output" for any single set of input data, and the multiple "good outputs" are difficult to compare in a quantifiable way in order to achieve a score. Since the first evaluation, there have been no new evaluations in almost a decade (Chinchor 2000).
On the other hand, some level of commercial use has been made of MT technology, as evidenced by the BabelFish translation system from Systran, available on AltaVista. However, the technology has definite limitations, as anyone with some knowledge of foreign languages will understand. For instance, if the user chooses to translate a page from French into English, it will often be immediately obvious that the grammatical structure of the original document has been retained in the translation.
Additionally, idioms and proper names that do not exist in the "dictionary" used by the system will either be translated literally, or simply left untranslated. Given this limitation, if the user is familiar with the source language, one can often make sense of the translation with the awareness that certain features of the original language structure and form will affect the translation. "Word-sense disambiguation" is an additional concern for MT systems, since many words can be translated in multiple ways. Automatic word sense disambiguation can help a system determine which usage of the word is meant - for example, the translation of the French word grille could be made as either "railings, gate, bar, grid, scale, [or] schedule" (Ide and Véronis 1998).
Elliott Macklovich, of the Computational Linguistics Laboratory at the University of Montreal states that "for high-quality, polished text, we just won't be able to dispense with [human translators] for some time" (Caragata 1999). Later in the same article, he clarifies: "All the machine has is access to the words, but you need more than words to translate correctly. You need background knowledge, common sense and culture. Machines have none of that."
When automatic systems were used to translate languages for the Eurpean Union (which includes countries speaking eleven different languages), the drafts produced had to be so heavily edited that any savings in cost was eliminated (Caragata 1999).
The difficulty can be easily illustrated. On a given day, I used the BabelFish translator to convert the following two sentences from English into French, then from French back into English:

"Translating two languages automatically is a piece of cake".
"Summarization of text is no joyride".

The French translations are given below:

"La traduction de deux langages est automatiquement un morceau de gâteau".
"Summarization de texte n'est aucun joyride".

As you see, the English word "joyride" was not translated into French, since the dictionary for the translation system did not apparently contain that word.

The English re-translations are given next, showing the stilted nature of the new sentence structures:

"The translation of two languages is automatically a piece of cake".
"Summarization of text is not any joyride".

This page has been accessed - - times since November 14, 2000