Machine Translation Revisited

Last month, Google Translate celebrated its 6th birthday. According to Google, more than 200 million people use this simple Internet translation service monthly. “What all the professional human translators in the world produce in a year, our system translates in roughly a single day,” said Google Translate engineer Franz Och, who originally invented this Google service.

In 2010, the US Defense Department spent nearly $ 700 million on a single translation contract for human interpreters – mostly in Afghanistan. A high demand for machine translation systems that would reduce this vast budget is obvious.

Statistical methods are state-of-the-art

Today, concepts of statistical machine translation (SMT) seem to be the most reliable and successful approaches compared to other models that did not perform in desired quality standards during the past decades. (A good overview on machine translation history can be found HERE). Rather than trying to explicitly encode rules for translating from one language to another, the idea of SMT is to get algorithms to infer those rules from existing databases of translated texts. The European Union, with its huge translation projects (up to 27 languages served), provides one of the world’s largest text databases. With MT@EC, the Commission is developing a new system based on SMT that will provide an improved machine translation service in terms both of quality of output and number of supported languages. Google Translate is based on statistical procedures as well. To acquire this huge amount of linguistic data, Google used United Nations documents in six different languages.

Machine translation systems have clear limitations

The key question remains, however: Does a machine translation get the meaning right? Metrics like BLEU (Bilingual Evaluation Understudy, NIST (National Institute of Standards and Technology) or METEOR (Metric for Evaluation of Translation with Explicit Ordering) try to evaluate the output of a machine translation. The measure of evaluation for metrics is correlation with human judgment. A very good overview on evaluation methods is given by Christian Federmann in his scientific paper “How can we measure machine translation quality”.

Machine results have clear advantages as long as a rough translation is sufficient. When it comes to poetry, however, the limitations of systems like Google Translate are obvious as the following example shows. The opening sentence in Marcel Proust’s “In search of Long Times” – “Longtemps, je me suis couché de bonne heure” – is translated by Google into: “Long time, I went to bed early.” The human translator, with a more precise feeling of Proust’s meaning, instead wrote: “For a long time I used to go to bed early.”

The same applies to highly complex issues like technical descriptions or patent translations. A single error in meaning of terminology in a patent translation can lead to severe and costly patent infringements.
The patent EP 0695351 on the “isolation, selection, and propagation of animal transgenic stem cells,” which originally belonged to Edinburgh University and was licensed to the Australian biotechnology firm Stem Cell Sciences, is a good example. In the year 2000, the European Patent Office issued a statement expressing regret about a critical mistake. It explained that an error arose because in the terminology used in English (“transgenic animal”) “animal” includes the notion of “human,” whereas a direct German or French translation of “animal” does not. As the qualifier “non-human” was missing, the patent also covered the genetic manipulation of human stem cells. Though this translation process was not delivered by a machine, the example illustrates the potential of damage that lies in a critical translation.

Machine translation can save lives

Nevertheless, in an article published in The New York Times in 2010, author David Bellos gives a very good example of successful work with a machine translation program: “When Haiti was devastated by an earthquake in January, aid teams poured in to the shattered island, speaking dozens of languages — but not Haitian Creole. How could a trapped survivor with a cellphone get usable information to rescuers? If he had to wait for a Chinese or Turkish or an English interpreter to turn up he might be dead before being understood. Carnegie Mellon University instantly released its Haitian Creole spoken and text data, and a network of volunteer developers produced a rough-and-ready machine translation system for Haitian Creole in little more than a long weekend. It didn’t produce prose of great beauty. But it worked.” (Read the full article HERE.)

This may as well be the current conclusion of the never-ending discussion on whether machine translation will ever replace human translation efforts: yes, as long as quality, meaning and stylistic matters are not really important.