$Date: 2004/04/13 06:21:41 $
From Kevin Duh: Some pondering for today's interesting discussion:
  1. Example-based MT, in my little knowledge of it, basically performs translation by analogy. It uses the assumption: ``/If a previously translated sentence occurs again, the same translation is likely to be correct again./'' (Brown 99). So the basic algorithm is to match the source sentence to the most similar example in the bilingual corpus, then find the corresponding translation (using a bilingual dictionary, word alignment methods, etc), and finally piece together all the translated matches. The assumption has a big IF (if a previously translated sentence occurs), so we basically end up with something similar to the generalization problem in statistical learning. To get around this problem, EBMT systems "generalize" the bitext by allowing "classes" to exist in place of words. However, the "classes" must be restricted so that the rules of syntax, etc are observed. That's where I see this paper fits in. All the knowledge-based rules and "feature bundles" are basically ways to restrict the sentence matching. So I think it's addressing a very important issue. I'm curious to see how other researchers in EBMT deal with this problem.
  2. I was a bit surprised that the EBMT paper cited various papers I used to consider to belong to the Statistical MT (SMT) camp. It seems that some researchers don't draw such a sharp distinction between EBMT and SMT. After all, they are both based on example corpus. I'm sure more similarities can be drawn between the two approaches. What I realized from this observation is that in my own brain I saw EBMT and SMT as totally separate things and therefore did not survey any papers in EBMT. I think I fell in the trap of building a wall around what I consider my own area of interest and totally paid no attention to what's outside. bad!
-Kevin

ps. "The cat the dog the rat chased saw slept." Another reading from a friend in the lab (who will remain anonymous) is: "The cat, dog, and rat are all friends. They played tag, watch movies, and sleep together." :)