Is Translation Memory dead?

Following a presentation by Jaap van der Meer at the LISA standards summit a number of people have been expressing the view that translation memory as we currently know it will not be around in five years. I think they are wrong and I would like to explain why.

Those that claim that translation memory is on the way out have come to this conclusion because they believe that statistical machine translation is becoming more and more important and as the corpus for large MT engines such as the Google one grow, translation memory will become increasingly irrelevant. There are even claims that while translation memory may be here in five years it will not have an economic value.

There is a huge volume of commercial translation which goes through translation memory tools and a relatively small volume using machine translation tools. However, recently the amount of commercial organisations using machine translation has grown and there is a very large amount of non-commercial machine translation being done, particularly through Google. That said, if an organisation wants a quality translation for professional purposes they are likely to use a translation agency or translators who uses a translation memory tool.

One of the ideas behind the belief that the translation memory is on the way out is that the larger the corpus the better and we will soon get to a stage where an enormous corpus will produce almost perfect results from statistical machine translation engines. I believe there is an optimum level with statistical machine translation. The law of diminishing returns kicks in when you get beyond this level. There are more possibilities for errors in both the corpus and how systems manage it and even for creative sabotage. Google give translators and everyone else the opportunity to provide a better translation. While this will usually improve it, there is an opportunity to sabotage the corpus or at least make it funnier. Use Google Translate to translate ‘Quid Pro Quo’ from Latin to English for an example of this. However, I have no idea what the optimum level will prove to be.

The value behind translation memory is that a professional translation has worked on the matching segments or corpus. Translation memory is of greatest value when the content of the translation memory matches closely the content of the document to be translated. Translator and translation agencies are often specialists. They may work for the same customers for a long time and often concentrate on particular subject matter. Statistical machine translation such as Google’s may be able to offer a gist of a document but unless the corpus has matches which originated from a professional translator and matches their areas of expertise it cannot offer the same value as a well maintained translation memory. As long as translation memory continues to make translation work easier it will be used. I believe that in five years there may be changes but we are more likely to see improvements with translation memory than its demise.