Many different people use Google Translate for a multitude of purposes – individual users use the service to translate recipes, some organisations use it to translate public signs (like the one above) and some companies use it to translate their websites. This last practice, however, is frowned upon by developers of the tool as it affects the service’s software.
On website the Register, Jack Clark reveals a big flaw in the Google Translate service. Even though he says engineers from all over the world have tried to fix the problem, the translations the service produces are of poor quality when the source texts used by the software aren’t too great to start with.
According to Clark, Google is aware of this issue.
In fact, the company’s director of research, Peter Norvig, discussed the problem when it was mentioned by an audience member at the Nasa Innovative Advanced Concepts conference held on the 5th of February in Stanford, America.
Norvig stated Google knows about the problem that arises when websites use the ‘cut-rate’ method to translate their website, i.e. use the Google service for their localization efforts.
When Google indexes websites translated with the cut-rate method, Clark says, the software might decide to use the translation for its own translation memory. Most additions to the Google Translate translation memory improve the quality of future translations, but Clark believes this is not the case when texts translate by the tool itself are added.
According to Norvig, the consequences of this flaw in the software are not substantial, but might become worse in the future.
Google tries to reduce the effects to a minimum by looking at each site individually, Norvig says. If the translation seems to be off, the website will be deleted from the translation memory.
Clark states that Google has already taken measures to decrease the amount of badly translated texts by closing down their Google Translate API service.
In addition to his comments about the cut-rate method, Norvig also revealed that Google has tried to ‘fingerprint’ every translation by looking at word and syntax choice, but that this practice was discontinued because of disappointing results.
[From The Scary Truth about Google Translate by Mox]