We at the Translation Center have never been a supporter of machine translation and have always advocated the use of humans, but I have to admit that we have been following Google Translate closely.
Traditional machine translation programs use a decode/recode methodology, with programmers teaching the machines the syntactic and lexical rules of grammar in the two respective languages. Over the last few decades, despite the investment of millions of dollars, the results have been disappointing, averaging about 75% accuracy and useful for very little besides weather reports.
Google Translate, however, uses a corpus-based approach, searching its massive databases of human-generated translations for matches. In the past, computers were never large enough to handle the data needed, but Google, by networking is resources, is quickly approaching the critical mass.
In a March 8 New York Times article titled “Google’s Computing Power Refines Translation Tool,” Miguel Helft reports on Google’s efforts to feed their program millions of passages and then search the translations for the best fit. It seems as if Google’s infrastructure, collected data, including its book-scanning projects, and search engines are particularly well-suited for the new technology.
Google Translate now handles 52 languages, more than other systems, and draws on billions of words, again more than any competitor. The larger the corpus to draw upon, the better the translations. And people are using it, too. According to Helft, hundreds of millions times a week to translate Web pages as well as other texts.
New projects, such as combining text searches with image searches, allowing people to take a cell-phone picture of a text and use Google Translate for an instant translation.
While the accuracy of Google Translate is still less than perfect, and we at the Center still recommend using humans, we need to keep an eye on Google Translate as it continues to expand and improve.