[stevenbedrick]

My dissertation research involved using (and, most importantly, evaluating) GT for certain medical text problems (going from EN->ES), and I found that its performance is highly domain-dependent (i.e., the quality of its translations depends a lot on what kind of input text you're feeding it). Furthermore, I found that whether or not it's "good enough" depends heavily on the user, specifically that user's level of proficiency in English.

GT is better than a lot of machine translation systems, but it still falls victim to a number of common problems that face such systems. Specifically, I've run into a lot of major word-sense-disambiguation issues, especially when working with the sort of short snippets of text that the gem in question uses as examples. Basically, that sort of short bit of text is a worst-case scenario for statistical machine translation, since there's so little context. Google generally does much better with longer runs of text than it does with short phrases.