SYSTRAN just released the latest version of its MT software, SDL announced a KbTS success story, and Lionbridge said on its 2006 earnings call that it will join companies like ITP in using MT in its translation service. However, for most companies delivering multilingual content machine translation remains a controversial subject — quality concerns top the list of issues and thus limit its acceptance.

Just how much do consumers use this “flawed” technology? Lots. Last year we surveyed 2,430 consumers in 8 countries where English is not the official language. Here’s what we asked about their use of online machine translation (OLMT): “how often do you use automated machine translation such as BabelFish, Google, or Systran to better understand the English you read at a website?”

As a whole, 53.5% said they use MT sometimes (35.0%), frequently (15.5%), and always (3.0%).

When we cut the data according to their language proficiency (those with both no-or-low knowledge of English versus those who characterized themselves as proficient in English) we saw a bell curve: “Always” was a scant 5.0% and 1.6%, respectively; “frequently” scored 20.2 and 12.2%; “sometimes” was the biggest, coming in at 31.5 and 37.4%; “rarely” was 22 and 27.2%, and “never” was chosen by 21.4% of the no-or-low segment and 21.6 percent of those with English competency.

Knowing that so many consumers turn to machine translation anyway, how should companies use MT to facilitate communications?

• Offer an MT choice. You’ll never be able to translate everything, so when you have insufficient budget or time for human translation, provide an MT button. But remember to alert visitors to potential issues, and monitor the requests so you know what content receives sufficient interest to justify human translation. This automation, opt-in, and monitoring option gives foreign-language visitors a choice between the de facto zero translation and the imperfect machine translation they will use anyway.
• Trust some MT more than others. We found that this usage data maps intriguingly to the relative quality of available machine engines (with Spain being an exception). For French and German, low quality is less low than for Turkish, Chinese, and Brazilian Portuguese. Japanese users are more acclimatized to computer-translated output, having relied on it for many years for online browsing. Use these trends as guidelines in selecting languages for MT rendering.
• Never publish MT output as your own content. Do NOT publish output from MT without post-editing and review. If you position MT output as your own, visitors will naturally assume you have read and approved the possibly bizarre content they are reading.

The bottom line: If you don’t provide human-translated content for your website visitors who don’t speak your language, they’ll find some free MT on their own. The quality they get at an OLMT site will surely be less that what you could offer — just as the damage to your brand could be substantially greater from untuned free translation.