Machine Translation using Moses

I’ve read critiques (link? gah. lost it) of the open source machine translation system Moses revolving around the need for a *nix (Unix based operating systems like Linux and the Mac OSX) savvy systems administrator – this is no out of the box solution for the home user.

TAUS have recently posted a list of the highlights from a recent conference discussion about businesses that use it. As noted in the first account (from Adobe, no less), Moses doesn’t take that many resources – I’ve got it installed here on my 4 year old Ubuntu desktop – but it does give better and faster results the more you have at hand. It’s not just CPU cycles, but a host of other, potentially costly, resources – programmers, corpus tamers, and the time intensive training and output analysis phases:

In the general discussion, most Moses users said they tend to test their raw output against Google or Bing Translate output, but the results naturally varied with the type of content. Nearly all are now using Moses in production, and in many cases they use TDA data to help train the engines. Moses makes most sense when used for very high volumes of throughput. ROI (the investment is mostly labor) can be obtained in about two weeks.