The Problem with Machine Translation

The dream of machine translation is perhaps as old as computers themselves. Before command-line interfaces, email or word processing, machine translation (MT) was one of the first two computer applications designed to act upon words instead of numbers. But the infamous tale of the American military system which is said to have rendered the popular Russian saying "The spirit is willing but the flesh is weak" into "The vodka is good but the meat is rotten" is just a sign of the bitter frustrations that followed the unbridled optimism of the early times.

The Evolution of Machine Translation

Languages are perhaps the most elusive element of the human culture: no one knows the creator, but still they survived through the centuries in hybrid forms, dashing out the ambitions of those, like some structuralists, who wanted to confine them to a strict domain of logical rules and signs.

It isn't that languages do not have rules. But those rules (or at least some of them) are only learnt after years of human interaction. Such a dictionary that would allow the user to find every single occurrence of every word in any language would be no more than a chimera.        

But it was such a chimera that presided the first feasibility trial of a translation machine that was carried out in 1954, as a joint project between IBM and the University of Georgetown. The first versions of machine translation software were based on detailed bilingual dictionaries that offered a number of equivalent words in the target language for each word listed in the source language, as well as a series of rules on word order.

However, initial optimism soon disappeared. Researchers began to think that the semantic barriers were overwhelming and no longer saw a solution on the near horizon to the problem of machine translation.

Research continued in Canada, France and Germany and two machine translation systems came into being several years later: "Systran", used by the European Union Commission and "Taum-météo", created by the University of Montreal to translate weather forecasts from French to English.

Important advances occurred during the 1980s. The administrative and commercial needs of multilingual communities stimulated the demand for translation, leading to the development in countries such as France, Germany, Canada and Japan of new machine translation systems such as "Logos" (from German to French and vice-versa) and the internal system created by the Pan-American Health Organization (from Spanish to English and vice-versa), as well as a number of systems produced by Japanese computer companies. Research also revived in the 1980s because large-scale access to personal computers and word-processing software produced a market for less expensive machine translation systems.

The beginning of the 1990s saw vital developments in machine translation with a radical change in strategy from a translation based on grammatical rules to that based on bodies of texts and examples.

Language was no longer perceived as a static entity governed by fixed rules, but as a dynamic body that changes according to use and users, evolving through time and adapting to social and cultural realities. But the so-far results are still far from acceptable, despite some progresses on the technical field.

To this day, machine translation continues its progress. Large companies are beginning to use it and software sales to the general public are increasing as well. This situation has led to the creation of on-line machine translation services which offer quick (but rarely efficient) translations in the desired language, as well as multilingual dictionaries, encyclopaedias and free terminology databases.

Machine Translation vs. Computer Aided Translation (CAT)

Machine Translation is often wrongly mixed with Computer Aided Translation (CAT). These two technologies are the offspring of different approaches. They do not produce the same results, and are used in distinct contexts.

MT aims at assembling all the information necessary for translation in a software programme so that a text can be translated without human intervention whatsoever. It exploits the computer's capacity to analyse the structure of a statement or sentence in the source language, break it down into easily translatable elements and then create a statement with the same structure in the target language.

CAT uses a number of tools to help the translator work accurately and quickly, the most important of which are terminology databases and translation memories. In effect, the computer offers a new way of approaching text processing of both the source and target text.

The technology basically acts as a recycler, offering the professional possible translations for the text he's working on, which are based on previous material. CAT is not capable of producing an immediately useable text, as languages are highly dependant on context. Backed by a translation memory, CAT is considered mainly a save-time tool, rather than a replacement for human activity. It requires post-editing in order to yield a quality target text.

In its simplest form, a translation memory is a database in which a translator stores translations for future re-use, either in the same text or other texts. Basically the software records bilingual pairs: a source-language segment (usually a sentence) combined with a target-language segment.

If an identical or similar source-language segment comes up later, the translation memory will find the previously-translated segment and automatically suggest it for the new translation. The translator is free to accept it without change, edit it to fit the current context, or reject it altogether.

The Limits of Machine Translation

Pure-machine translation can deliver acceptable results when dealing with very predictable technical texts, which never go beyond the expected domain of discourse. But this is little help in the domains where people want translation the most: in spontaneous conversations, in person, on the telephone and on the internet.

Computers just do not have the ability to deal adequately with the various complexities of language than humans handle naturally: ambiguity, syntactic irregularity, multiple word meanings and the influence of context.

A classic example is illustrated in the following pair of sentences: "He drives too fast" and "Patients must come for the blood test on fast".  A computer can be programmed to 'understand' either of these examples, but not to distinguish between the two occurrences of "fast".  A pure-computer translation is similar to the one performed by a human without a deep knowledge of the target language.

Grammatical rules can be memorised, or programmed. But without real knowledge of a language, a human or a computer simply looks up words in a dictionary and has no way to choose between diverse meanings. Computers not only lack the knowledge of the world to deal with word choice, but they also lack the knowledge necessary for cultural sensitivity.

Acknowledgements