JJ's Journal: Stat Translation

All this writing about statistical translation is intruguing (sp) to me. It is in reality, right up my alley. My first master's thesis was on a specific topic in mechanical translation (MT). Specifically, interlinguas. I proposed using a dynamic one which was built up by reading text (aka processing Rosetta stones.) My second master's thesis was on a grammatical model of medieval Chinese, again based on stats, which could be utilized for MT. (Of course, who needs to translate Tang dynasty Chinese nowdays?) My third masters (yes, I am a glutton for punishment) was in stats.
      Now that I am a professional programmer, I can visualize how difficult it would be to put all of this together and actually get reasonable text out.
      I am befuddled why MT efforts have been so dismal. I mean, we've been working at it for nearly 60 years (my first advisor gave me the minutes from a 1951 meeting he attended where several people said 3 or 4 years of work already.) And our most succesful effort is a blind statistical one from USC (mentioned in Slashdot's Romancing the Rosetta Stone.) Stats will take you a long way, particularly with relatively straightforward (grammar-wise) English-Chinese translations. However, the best this effort can ever do is the average of all input human translators. And there are language pairs, English to Inuktatuk (Greenlandic Eskimo) or English to Basque, where this type of approach would just wilt on the fine.