Translation Memory Takes Diverging Paths in Automotive Industry

Article Details

Global Watchtower

Common Sense Advisory Blogs

Posted by Donald A. DePalma on January 10, 2007 in the following blogs: Translation and Localization, Web Globalization, Business Globalization, Technology, Interpreting, Market Data, Global Marketing, Best Practices, Supplier Business Issues
Pages \| \|

Earlier this year the world's currently largest automaker announced that it would license translation memories through TM Marketplace (TMMP). Late last year a language service provider (LSP) for the world's soon-to-be-largest manufacturer of cars said that it would use automated translation technology to improve its ability to serve its automotive customers in Europe.

GM Service Parts and Operations made translation memories in six language combinations available for licensing through TMMP. As the first company using TMMP's honest broker service, GM hopes to increase the use of its TMs as it monetizes them -- and increase its "terminological lead in the automobile industry."As we noted when we wrote about TMMP 18 months ago, the best and most likely use of TMs is within a company's ecosystem of parts and component suppliers. We agree that anybody selling to GM will be interested, but we believe GM should readily and freely share this asset with all its suppliers in order to decrease friction in its supply chain.

And we don't expect Daimler, Hyundai, or Toyota to lay out cash for terms used in GM shop floor manuals, service bulletins, and owner manuals -- while there is a lot of potential TM leverage in this highly competitive business, each manufacturer does market distinctive features and functions using terms that it would prefer to own. However, up-and-coming carmakers in China may choose to jumpstart their own supply chains with another company's well-established terminology -- that is something GM may ultimately be unhappy about.

The manager of Japan-based language service provider Ishida Taiseisha's (ITP) European operations announced that the firm had evaluated Language Weaver's statistical machine translation (SMT) engine and saw a 30% boost for its French and Spanish translations. With Language Weaver's SMT trained on its existing memories, ITP found this productivity gain with "quality maintained" and enhanced production control. The company is sold on the SMT concept and will roll out additional European languages during 2007. ITP derives nearly 85% of its revenue from developing technical and marketing publications for Asian automakers.Is this machine-assisted human translation for everyone? No. ITP fits a profile that SMT vendors seek -- a vertical industry focus, large volumes of previously translated words, and previous investment in optimizing its TM servers and its workflow processes. ITP's clients will see faster and cheaper translation from this work, but unless they require the LSP to share its TMs, the productivity benefit stops there.

The fact is, short sentences find the most matches in any TM. "Turn the key" probably will be written the same way at GM and Toyota. It's also something that both companies already have in their TM databases. Innovative manufacturing techniques and feature introductions like onboard navigation create high volumes of "new words" -- and also a greater risk of mistranslation from using inappropriate memories.

Ultimately, it comes down to euros and cents. Is the benefit worth the added cost and risk? Cost is added by buying segments, extra time spent by translators/editors selecting and tweaking matches, and dealing with service and liability issues from any errors. Some buyers may find a net benefit, but most will not.

One fundamental shortcoming of the TMMP approach is its categorization method, which is based on self-certification. Contributors describe their translation process in a limited sort of way, but no formal criteria exist for assessing quality. Some professionals suggest that companies with large TM databases would benefit more by throwing the accumulated detritus overboard and starting fresh. Powerful tools from companies like Language Weaver, MultiCorpora, and SDL harvest terminology and memories from parallel multi-language texts, which may be faster and more effective than manually scrubbing dubious or aging TMs.

For shared TMs to work at all, we think that they must be free or at the very least charged only according to actual words used. This latter option is only feasible within multi-tenant applications, such as Lingotek and Lionbridge's Logoport. Why pay for 1.7 billion translation units when you might only use 30,000, or 3,000, or 30?

Categories